From 17f37225f68ef2a1e640541bba62c5abf066f6da Mon Sep 17 00:00:00 2001
From: Wojciech Indyk <wojciechindyk@gmail.com>
Date: Fri, 28 Aug 2020 19:22:23 +0200
Subject: [PATCH 1/6] Automatic conversion via tf_upgrade_v2.

---
 btgym/algorithms/aac.py                   | 108 ++---
 btgym/algorithms/math_utils.py            |  16 +-
 btgym/algorithms/nn/ae.py                 |  24 +-
 btgym/algorithms/nn/layers.py             |  62 +--
 btgym/algorithms/nn/losses.py             |  92 ++--
 btgym/algorithms/nn/networks.py           |  36 +-
 btgym/algorithms/policy/base.py           |  40 +-
 btgym/algorithms/policy/meta.py           |  42 +-
 btgym/algorithms/policy/stacked_lstm.py   |  52 +--
 btgym/algorithms/utils.py                 |  12 +-
 btgym/algorithms/worker.py                |  32 +-
 btgym/monitor/tensorboard2.py             |  20 +-
 btgym/research/b_vae_a3c.py               |  64 +--
 btgym/research/casual/aac.py              |  16 +-
 btgym/research/casual_conv/layers.py      |  20 +-
 btgym/research/casual_conv/networks.py    |  14 +-
 btgym/research/encoder_test/aac.py        |  70 +--
 btgym/research/encoder_test/networks.py   |  10 +-
 btgym/research/encoder_test/policy.py     |  38 +-
 btgym/research/gps/aac.py                 |   2 +-
 btgym/research/gps/loss.py                |  24 +-
 btgym/research/metalearn_2/_fwrnn_aac.py  | 188 --------
 btgym/research/metalearn_2/_mldg_batch.py | 530 ----------------------
 btgym/research/metalearn_2/loss.py        |  28 +-
 btgym/research/mldg/aac.py                |  30 +-
 btgym/research/mldg/aac_1.py              |  18 +-
 btgym/research/mldg/aac_1d.py             |  10 +-
 btgym/research/mldg/aac_1s.py             |  38 +-
 btgym/research/mldg/policy.py             |   2 +-
 btgym/research/model_based/aac.py         |  68 +--
 btgym/research/policy_rl2.py              |   2 +-
 setup.py                                  |   2 +-
 32 files changed, 496 insertions(+), 1214 deletions(-)

diff --git a/btgym/algorithms/aac.py b/btgym/algorithms/aac.py
index 1f00b47c..43a0539a 100644
--- a/btgym/algorithms/aac.py
+++ b/btgym/algorithms/aac.py
@@ -198,7 +198,7 @@ def __init__(self,
             self.random_seed = random_seed
             if self.random_seed is not None:
                 np.random.seed(self.random_seed)
-                tf.set_random_seed(self.random_seed)
+                tf.compat.v1.set_random_seed(self.random_seed)
             self.log.debug('rnd_seed:{}, log_u_sample_(0,1]x5: {}'.
                            format(random_seed, log_uniform([1e-10,1], 5)))
 
@@ -405,7 +405,7 @@ def __init__(self,
             self.log.debug('started building graphs...')
             if self.use_global_network:
                 # PS:
-                with tf.device(tf.train.replica_device_setter(1, worker_device=self.worker_device)):
+                with tf.device(tf.compat.v1.train.replica_device_setter(1, worker_device=self.worker_device)):
                     self.network = pi_global = self._make_policy('global')
                     if self.use_target_policy:
                         self.network_prime = self._make_policy('global_prime')
@@ -417,7 +417,7 @@ def __init__(self,
 
             # Worker:
             with tf.device(self.worker_device):
-                with tf.variable_scope(self.name):
+                with tf.compat.v1.variable_scope(self.name):
                     self.local_network = pi = self._make_policy('local')
 
                     if self.use_target_policy:
@@ -429,7 +429,7 @@ def __init__(self,
                     self.worker_device_callback_0()  # if need more networks etc.
 
                     # Meant for Batch-norm layers:
-                    pi.update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='.*local.*')
+                    pi.update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS, scope='.*local.*')
 
                     # Just in case:
                     self.dummy_pi = self._make_dummy_policy()
@@ -440,7 +440,7 @@ def __init__(self,
                         self.log.debug('{}: {}'.format(v.name, v.get_shape()))
 
                     #  Learning rate annealing:
-                    self.learn_rate_decayed = tf.train.polynomial_decay(
+                    self.learn_rate_decayed = tf.compat.v1.train.polynomial_decay(
                         self.opt_learn_rate,
                         self.global_step + 1,
                         self.opt_decay_steps,
@@ -507,13 +507,13 @@ def _make_base_loss(self, pi, pi_prime, name='base', verbose=True):
             tensor holding estimated loss graph
             list of related summaries
         """
-        with tf.name_scope(name):
+        with tf.compat.v1.name_scope(name):
             # On-policy AAC loss definition:
-            pi.on_pi_act_target = tf.placeholder(
+            pi.on_pi_act_target = tf.compat.v1.placeholder(
                 tf.float32, [None, self.ref_env.action_space.one_hot_depth], name="on_policy_action_pl"
             )
-            pi.on_pi_adv_target = tf.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
-            pi.on_pi_r_target = tf.placeholder(tf.float32, [None], name="on_policy_return_pl")
+            pi.on_pi_adv_target = tf.compat.v1.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
+            pi.on_pi_r_target = tf.compat.v1.placeholder(tf.float32, [None], name="on_policy_return_pl")
 
             clip_epsilon = tf.cast(self.clip_epsilon * self.learn_rate_decayed / self.opt_learn_rate, tf.float32)
 
@@ -534,10 +534,10 @@ def _make_base_loss(self, pi, pi_prime, name='base', verbose=True):
             model_summaries = on_pi_summaries
 
             # Off-policy losses:
-            pi.off_pi_act_target = tf.placeholder(
+            pi.off_pi_act_target = tf.compat.v1.placeholder(
                 tf.float32, [None, self.ref_env.action_space.one_hot_depth], name="off_policy_action_pl")
-            pi.off_pi_adv_target = tf.placeholder(tf.float32, [None], name="off_policy_advantage_pl")
-            pi.off_pi_r_target = tf.placeholder(tf.float32, [None], name="off_policy_return_pl")
+            pi.off_pi_adv_target = tf.compat.v1.placeholder(tf.float32, [None], name="off_policy_advantage_pl")
+            pi.off_pi_r_target = tf.compat.v1.placeholder(tf.float32, [None], name="off_policy_return_pl")
 
             if self.use_off_policy_aac:
                 # Off-policy AAC loss graph mirrors on-policy:
@@ -558,8 +558,8 @@ def _make_base_loss(self, pi, pi_prime, name='base', verbose=True):
 
             if self.use_pixel_control:
                 # Pixel control loss:
-                pi.pc_action = tf.placeholder(tf.float32, [None, self.ref_env.action_space.tensor_shape[0]], name="pc_action")
-                pi.pc_target = tf.placeholder(tf.float32, [None, None, None], name="pc_target")
+                pi.pc_action = tf.compat.v1.placeholder(tf.float32, [None, self.ref_env.action_space.tensor_shape[0]], name="pc_action")
+                pi.pc_target = tf.compat.v1.placeholder(tf.float32, [None, None, None], name="pc_target")
 
                 pc_loss, pc_summaries = self.pc_loss(
                     actions=pi.pc_action,
@@ -574,7 +574,7 @@ def _make_base_loss(self, pi, pi_prime, name='base', verbose=True):
 
             if self.use_value_replay:
                 # Value function replay loss:
-                pi.vr_target = tf.placeholder(tf.float32, [None], name="vr_target")
+                pi.vr_target = tf.compat.v1.placeholder(tf.float32, [None], name="vr_target")
                 vr_loss, vr_summaries = self.vr_loss(
                     r_target=pi.vr_target,
                     pi_vf=pi.vr_value,
@@ -586,7 +586,7 @@ def _make_base_loss(self, pi, pi_prime, name='base', verbose=True):
 
             if self.use_reward_prediction:
                 # Reward prediction loss:
-                pi.rp_target = tf.placeholder(tf.float32, [None, 3], name="rp_target")
+                pi.rp_target = tf.compat.v1.placeholder(tf.float32, [None, 3], name="rp_target")
 
                 rp_loss, rp_summaries = self.rp_loss(
                     rp_targets=pi.rp_target,
@@ -613,7 +613,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
         """
 
         # Each worker gets a different set of adam optimizer parameters:
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
 
         # self.optimizer = tf.train.RMSPropOptimizer(
         #    learning_rate=train_learn_rate,
@@ -624,10 +624,10 @@ def _make_train_op(self, pi, pi_prime, pi_global):
 
         # Clipped gradients:
         self.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.loss, pi.var_list),
+            tf.gradients(ys=self.loss, xs=pi.var_list),
             40.0
         )
-        self.grads_global_norm = tf.global_norm(self.grads)
+        self.grads_global_norm = tf.linalg.global_norm(self.grads)
         # Copy weights from the parameter server to the local model
         self.sync = self.sync_pi = tf.group(
             *[v1.assign(v2) for v1, v2 in zip(pi.var_list, pi_global.var_list)]
@@ -649,7 +649,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
             stream = pi.on_state_in['external'][list(pi.on_state_in['external'].keys())[0]]
         else:
             stream = pi.on_state_in['external']
-        self.inc_step = self.global_step.assign_add(tf.shape(stream)[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=stream)[0])
 
         train_op = self.optimizer.apply_gradients(grads_and_vars)
         self.log.debug('train_op defined')
@@ -666,71 +666,71 @@ def _combine_summaries(self, policy=None, model_summaries=None):
         if model_summaries is not None:
             if self.use_global_network:
                 # Model-wide statistics:
-                with tf.name_scope('model'):
+                with tf.compat.v1.name_scope('model'):
                     model_summaries += [
-                        tf.summary.scalar("grad_global_norm", self.grads_global_norm),
+                        tf.compat.v1.summary.scalar("grad_global_norm", self.grads_global_norm),
                         # TODO: add gradient variance summary
                         #tf.summary.scalar("learn_rate", self.train_learn_rate),
-                        tf.summary.scalar("learn_rate", self.learn_rate_decayed),  # cause actual rate is a jaggy due to test freezes
-                        tf.summary.scalar("total_loss", self.loss),
+                        tf.compat.v1.summary.scalar("learn_rate", self.learn_rate_decayed),  # cause actual rate is a jaggy due to test freezes
+                        tf.compat.v1.summary.scalar("total_loss", self.loss),
                         # tf.summary.scalar('roll_reward', tf.reduce_mean(self.local_network.on_last_reward_in)),
                         # tf.summary.scalar('roll_advantage', tf.reduce_mean(self.local_network.on_pi_adv_target)),
                     ]
                     if policy is not None:
-                        model_summaries += [tf.summary.scalar("var_global_norm", tf.global_norm(policy.var_list))]
+                        model_summaries += [tf.compat.v1.summary.scalar("var_global_norm", tf.linalg.global_norm(policy.var_list))]
         else:
             model_summaries = []
         # Model stat. summary:
-        model_summary = tf.summary.merge(model_summaries, name='model_summary')
+        model_summary = tf.compat.v1.summary.merge(model_summaries, name='model_summary')
 
         # Episode-related summaries:
         ep_summary = dict(
             # Summary placeholders
-            render_atari=tf.placeholder(tf.uint8, [None, None, None, 1]),
-            total_r=tf.placeholder(tf.float32, ),
-            cpu_time=tf.placeholder(tf.float32, ),
-            final_value=tf.placeholder(tf.float32, ),
-            steps=tf.placeholder(tf.int32, ),
+            render_atari=tf.compat.v1.placeholder(tf.uint8, [None, None, None, 1]),
+            total_r=tf.compat.v1.placeholder(tf.float32, ),
+            cpu_time=tf.compat.v1.placeholder(tf.float32, ),
+            final_value=tf.compat.v1.placeholder(tf.float32, ),
+            steps=tf.compat.v1.placeholder(tf.int32, ),
         )
         if self.test_mode:
             # For Atari:
-            ep_summary['render_op'] = tf.summary.image("model/state", ep_summary['render_atari'])
+            ep_summary['render_op'] = tf.compat.v1.summary.image("model/state", ep_summary['render_atari'])
 
         else:
             # BTGym rendering:
             ep_summary.update(
                 {
-                    mode: tf.placeholder(tf.uint8, [None, None, None, None], name=mode + '_pl')
+                    mode: tf.compat.v1.placeholder(tf.uint8, [None, None, None, None], name=mode + '_pl')
                     for mode in self.env_list[0].render_modes + self.aux_render_modes
                 }
             )
-            ep_summary['render_op'] = tf.summary.merge(
-                [tf.summary.image(mode, ep_summary[mode])
+            ep_summary['render_op'] = tf.compat.v1.summary.merge(
+                [tf.compat.v1.summary.image(mode, ep_summary[mode])
                  for mode in self.env_list[0].render_modes + self.aux_render_modes]
             )
         # Episode stat. summary:
-        ep_summary['btgym_stat_op'] = tf.summary.merge(
+        ep_summary['btgym_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode_train/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode_train/cpu_time_sec', ep_summary['cpu_time']),
-                tf.summary.scalar('episode_train/final_value', ep_summary['final_value']),
-                tf.summary.scalar('episode_train/env_steps', ep_summary['steps'])
+                tf.compat.v1.summary.scalar('episode_train/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode_train/cpu_time_sec', ep_summary['cpu_time']),
+                tf.compat.v1.summary.scalar('episode_train/final_value', ep_summary['final_value']),
+                tf.compat.v1.summary.scalar('episode_train/env_steps', ep_summary['steps'])
             ],
             name='episode_train_btgym'
         )
         # Test episode stat. summary:
-        ep_summary['test_btgym_stat_op'] = tf.summary.merge(
+        ep_summary['test_btgym_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode_test/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode_test/final_value', ep_summary['final_value']),
-                tf.summary.scalar('episode_test/env_steps', ep_summary['steps'])
+                tf.compat.v1.summary.scalar('episode_test/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode_test/final_value', ep_summary['final_value']),
+                tf.compat.v1.summary.scalar('episode_test/env_steps', ep_summary['steps'])
             ],
             name='episode_test_btgym'
         )
-        ep_summary['atari_stat_op'] = tf.summary.merge(
+        ep_summary['atari_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode/steps', ep_summary['steps'])
+                tf.compat.v1.summary.scalar('episode/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode/steps', ep_summary['steps'])
             ],
             name='episode_atari'
         )
@@ -798,24 +798,24 @@ def _make_step_counters(self):
         Returns:
             None, sets attrs.
         """
-        self.global_step = tf.get_variable(
+        self.global_step = tf.compat.v1.get_variable(
             "global_step",
             [],
             tf.int32,
-            initializer=tf.constant_initializer(
+            initializer=tf.compat.v1.constant_initializer(
                 0,
                 dtype=tf.int32
             ),
             trainable=False
         )
-        tf.add_to_collection(tf.GraphKeys.GLOBAL_STEP, self.global_step)
+        tf.compat.v1.add_to_collection(tf.compat.v1.GraphKeys.GLOBAL_STEP, self.global_step)
         self.reset_global_step = self.global_step.assign(0)
 
-        self.global_episode = tf.get_variable(
+        self.global_episode = tf.compat.v1.get_variable(
             "global_episode",
             [],
             tf.int32,
-            initializer=tf.constant_initializer(
+            initializer=tf.compat.v1.constant_initializer(
                 0,
                 dtype=tf.int32
             ),
@@ -837,7 +837,7 @@ def _make_policy(self, scope):
         Returns:
             policy instance
         """
-        with tf.variable_scope(scope):
+        with tf.compat.v1.variable_scope(scope):
             # Make policy instance:
             network = self.policy_class(**self.policy_kwargs)
             if 'global' not in scope:
@@ -1313,7 +1313,7 @@ def process_summary(self, sess, data, model_data=None, step=None, episode=None):
 
         # Every worker writes train episode summaries:
         if model_data is not None:
-            self.summary_writer.add_summary(tf.Summary.FromString(model_data), step)
+            self.summary_writer.add_summary(tf.compat.v1.Summary.FromString(model_data), step)
         self.summary_writer.flush()
 
     def process(self, sess, **kwargs):
diff --git a/btgym/algorithms/math_utils.py b/btgym/algorithms/math_utils.py
index 0bdb0c1b..94829a64 100644
--- a/btgym/algorithms/math_utils.py
+++ b/btgym/algorithms/math_utils.py
@@ -36,22 +36,22 @@ def log_uniform(lo_hi, size):
 
 
 def cat_entropy(logits):
-    a0 = logits - tf.reduce_max(logits, 1, keepdims=True)
+    a0 = logits - tf.reduce_max(input_tensor=logits, axis=1, keepdims=True)
     ea0 = tf.exp(a0)
-    z0 = tf.reduce_sum(ea0, 1, keepdims=True)
+    z0 = tf.reduce_sum(input_tensor=ea0, axis=1, keepdims=True)
     p0 = ea0 / z0
-    return tf.reduce_sum(p0 * (tf.log(z0) - a0), 1)
+    return tf.reduce_sum(input_tensor=p0 * (tf.math.log(z0) - a0), axis=1)
 
 
 def kl_divergence(logits_1, logits_2):
-    a0 = logits_1 - tf.reduce_max(logits_1, axis=-1, keepdims=True)
-    a1 = logits_2 - tf.reduce_max(logits_2, axis=-1, keepdims=True)
+    a0 = logits_1 - tf.reduce_max(input_tensor=logits_1, axis=-1, keepdims=True)
+    a1 = logits_2 - tf.reduce_max(input_tensor=logits_2, axis=-1, keepdims=True)
     ea0 = tf.exp(a0)
     ea1 = tf.exp(a1)
-    z0 = tf.reduce_sum(ea0, axis=-1, keepdims=True)
-    z1 = tf.reduce_sum(ea1, axis=-1, keepdims=True)
+    z0 = tf.reduce_sum(input_tensor=ea0, axis=-1, keepdims=True)
+    z1 = tf.reduce_sum(input_tensor=ea1, axis=-1, keepdims=True)
     p0 = ea0 / z0
-    return tf.reduce_sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=-1)
+    return tf.reduce_sum(input_tensor=p0 * (a0 - tf.math.log(z0) - a1 + tf.math.log(z1)), axis=-1)
 
 
 # def softmax(x):
diff --git a/btgym/algorithms/nn/ae.py b/btgym/algorithms/nn/ae.py
index 3cd25409..f88b60b9 100644
--- a/btgym/algorithms/nn/ae.py
+++ b/btgym/algorithms/nn/ae.py
@@ -31,7 +31,7 @@ def conv2d_encoder(x,
         level-wise list of encoding layers shapes, first ro last.
 
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         layer_shapes = [x.get_shape()]
         layer_outputs = []
         for i, layer_spec in enumerate(layer_config, 1):
@@ -83,7 +83,7 @@ def conv2d_decoder(z,
         list of tensors holding decoded features for every layer inner to outer
 
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         x = z
         layer_shapes = list(layer_shapes)
         layer_shapes.reverse()
@@ -91,7 +91,7 @@ def conv2d_decoder(z,
         layer_config.reverse()
         layer_output = []
         for i, (layer_spec, layer_shape) in enumerate(zip(layer_config,layer_shapes[1:]), 1):
-            x = tf.image.resize_images(
+            x = tf.image.resize(
                 images=x,
                 size=[int(layer_shape[1]), int(layer_shape[2])],
                 method=resize_method,
@@ -156,7 +156,7 @@ def conv2d_autoencoder(
         None value
 
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         # Encode:
         encoder_layers, shapes = conv2d_encoder(
             x=inputs,
@@ -223,7 +223,7 @@ def cw_conv2d_autoencoder(
         None value
 
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         ae_bank = []
         for i in range(inputs.get_shape().as_list()[-1]):
             # Making list of list of AE's:
@@ -313,7 +313,7 @@ def beta_var_conv2d_autoencoder(
         tensor holding estimated KL divergence
 
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
 
         # Encode:
         encoder_layers, shapes = conv2d_encoder(
@@ -358,8 +358,8 @@ def beta_var_conv2d_autoencoder(
 
         # Oversized noise generator:
         #eps = tf.random_normal(shape=[max_batch_size, half_size_z], mean=0., stddev=1.)
-        eps = tf.random_normal(shape=[max_batch_size, size_z], mean=0., stddev=1.)
-        eps = eps[:tf.shape(z)[0],:]
+        eps = tf.random.normal(shape=[max_batch_size, size_z], mean=0., stddev=1.)
+        eps = eps[:tf.shape(input=z)[0],:]
 
         # Get sample z ~ Q(z|X):
         z_sampled = mu + tf.exp(log_sigma / 2) * eps
@@ -405,18 +405,18 @@ def __init__(self, conv_input, layer_output):
             conv_input:         convolution stack input tensor
             layer_output:       tensor holding output of layer of interest from stack
         """
-        self.idx = tf.placeholder(tf.int32, name='kernel_index')
+        self.idx = tf.compat.v1.placeholder(tf.int32, name='kernel_index')
         self.conv_input = conv_input
         self.layer_output = layer_output
         # Build a loss function that maximizes the activation
         # of the n-th filter of the layer considered:
-        self.vis_loss = tf.reduce_mean(self.layer_output[:, :, :, self.idx])
+        self.vis_loss = tf.reduce_mean(input_tensor=self.layer_output[:, :, :, self.idx])
 
         # Gradient of the input picture wrt this loss:
-        self.vis_grads = tf.gradients(self.vis_loss, self.conv_input)[0]
+        self.vis_grads = tf.gradients(ys=self.vis_loss, xs=self.conv_input)[0]
 
         # Normalization trick:
-        self.vis_grads /= (tf.sqrt(tf.reduce_mean(tf.square(self.vis_grads))) + 1e-5)
+        self.vis_grads /= (tf.sqrt(tf.reduce_mean(input_tensor=tf.square(self.vis_grads))) + 1e-5)
 
     def _iterate(self, sess, signal, kernel_index):
         """
diff --git a/btgym/algorithms/nn/layers.py b/btgym/algorithms/nn/layers.py
index 6dcf01cf..19dd966b 100644
--- a/btgym/algorithms/nn/layers.py
+++ b/btgym/algorithms/nn/layers.py
@@ -31,7 +31,7 @@ def categorical_sample(logits, depth):
             tensor of shape [batch_dim, logits_depth]
     """
     # print('categorical_sample_logits: ', logits)
-    value = tf.squeeze(tf.multinomial(logits, 1), [1])
+    value = tf.squeeze(tf.random.categorical(logits=logits, num_samples=1), [1])
     one_hot = tf.one_hot(value, depth, name='sample_one_hot')
     return one_hot
 
@@ -40,9 +40,9 @@ def linear(x, size, name, initializer=None, bias_init=0, reuse=False):
     """
     Linear network layer.
     """
-    with tf.variable_scope(name, reuse=reuse):
-        w = tf.get_variable("/w", [x.get_shape()[1], size], initializer=initializer)
-        b = tf.get_variable("/b", [size], initializer=tf.constant_initializer(bias_init))
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
+        w = tf.compat.v1.get_variable("/w", [x.get_shape()[1], size], initializer=initializer)
+        b = tf.compat.v1.get_variable("/b", [size], initializer=tf.compat.v1.constant_initializer(bias_init))
         return tf.matmul(x, w) + b
 
 
@@ -56,29 +56,29 @@ def noisy_linear(x, size, name, bias=True, activation_fn=tf.identity, reuse=Fals
         https://arxiv.org/abs/1706.01905
 
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         # the function used in eq.7,8
         def f(x):
             return tf.multiply(tf.sign(x), tf.pow(tf.abs(x), 0.5))
         # Initializer of \mu and \sigma
-        mu_init = tf.random_uniform_initializer(minval=-1*1/np.power(x.get_shape().as_list()[1], 0.5),
+        mu_init = tf.compat.v1.random_uniform_initializer(minval=-1*1/np.power(x.get_shape().as_list()[1], 0.5),
                                                     maxval=1*1/np.power(x.get_shape().as_list()[1], 0.5))
-        sigma_init = tf.constant_initializer(0.4/np.power(x.get_shape().as_list()[1], 0.5))
+        sigma_init = tf.compat.v1.constant_initializer(0.4/np.power(x.get_shape().as_list()[1], 0.5))
         # Sample noise from gaussian
-        p = tf.random_normal([x.get_shape().as_list()[1], 1])
-        q = tf.random_normal([1, size])
+        p = tf.random.normal([x.get_shape().as_list()[1], 1])
+        q = tf.random.normal([1, size])
         f_p = f(p); f_q = f(q)
         w_epsilon = f_p*f_q; b_epsilon = tf.squeeze(f_q)
 
         # w = w_mu + w_sigma*w_epsilon
-        w_mu = tf.get_variable("/w_mu", [x.get_shape()[1], size], initializer=mu_init)
-        w_sigma = tf.get_variable("/w_sigma", [x.get_shape()[1], size], initializer=sigma_init)
+        w_mu = tf.compat.v1.get_variable("/w_mu", [x.get_shape()[1], size], initializer=mu_init)
+        w_sigma = tf.compat.v1.get_variable("/w_sigma", [x.get_shape()[1], size], initializer=sigma_init)
         w = w_mu + tf.multiply(w_sigma, w_epsilon)
         ret = tf.matmul(x, w)
         if bias:
             # b = b_mu + b_sigma*b_epsilon
-            b_mu = tf.get_variable("/b_mu", [size], initializer=mu_init)
-            b_sigma = tf.get_variable("/b_sigma", [size], initializer=sigma_init)
+            b_mu = tf.compat.v1.get_variable("/b_mu", [size], initializer=mu_init)
+            b_sigma = tf.compat.v1.get_variable("/b_sigma", [size], initializer=sigma_init)
             b = b_mu + tf.multiply(b_sigma, b_epsilon)
             return activation_fn(ret + b)
         else:
@@ -90,15 +90,15 @@ def conv2d(x, num_filters, name, filter_size=(3, 3), stride=(1, 1), pad="SAME",
     """
     2D convolution layer.
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         stride_shape = [1, stride[0], stride[1], 1]
         filter_shape = [filter_size[0], filter_size[1], int(x.get_shape()[3]), num_filters]
 
-        w = tf.get_variable("W", filter_shape, dtype, initializer=tf.contrib.layers.xavier_initializer(),
+        w = tf.compat.v1.get_variable("W", filter_shape, dtype, initializer=tf.compat.v1.keras.initializers.VarianceScaling(scale=1.0, mode="fan_avg", distribution="uniform"),
                             collections=collections)
-        b = tf.get_variable("b", [1, 1, 1, num_filters], initializer=tf.constant_initializer(0.0),
+        b = tf.compat.v1.get_variable("b", [1, 1, 1, num_filters], initializer=tf.compat.v1.constant_initializer(0.0),
                             collections=collections)
-        return tf.nn.conv2d(x, w, stride_shape, pad) + b
+        return tf.nn.conv2d(input=x, filters=w, strides=stride_shape, padding=pad) + b
 
 
 def deconv2d(x, output_channels, name, filter_size=(4, 4), stride=(2, 2),
@@ -107,10 +107,10 @@ def deconv2d(x, output_channels, name, filter_size=(4, 4), stride=(2, 2),
     Deconvolution layer, paper:
     http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         stride_shape = [1, stride[0], stride[1], 1]
 
-        batch_size = tf.shape(x)[0]
+        batch_size = tf.shape(input=x)[0]
         input_height = int(x.get_shape()[1])
         input_width = int(x.get_shape()[2])
         input_channels = int(x.get_shape()[3])
@@ -126,9 +126,9 @@ def deconv2d(x, output_channels, name, filter_size=(4, 4), stride=(2, 2),
         # initialize weights with random weights
         w_bound = np.sqrt(6. / (fan_in + fan_out))
 
-        w = tf.get_variable("d_W", filter_shape, dtype, initializer=tf.contrib.layers.xavier_initializer(),
+        w = tf.compat.v1.get_variable("d_W", filter_shape, dtype, initializer=tf.compat.v1.keras.initializers.VarianceScaling(scale=1.0, mode="fan_avg", distribution="uniform"),
                             collections=collections)
-        b = tf.get_variable("d_b", [1, 1, 1, output_channels], initializer=tf.constant_initializer(0.0),
+        b = tf.compat.v1.get_variable("d_b", [1, 1, 1, output_channels], initializer=tf.compat.v1.constant_initializer(0.0),
                             collections=collections)
 
         return tf.nn.conv2d_transpose(x, w, output_shape,
@@ -141,7 +141,7 @@ def conv1d(x, num_filters, name, filter_size=3, stride=2, pad="SAME", dtype=tf.f
     """
     1D convolution layer.
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         stride_shape = stride
 
         # print('stride_shape:',stride_shape)
@@ -150,11 +150,11 @@ def conv1d(x, num_filters, name, filter_size=3, stride=2, pad="SAME", dtype=tf.f
 
         # print('filter_shape:', filter_shape)
 
-        w = tf.get_variable("W", filter_shape, dtype, initializer=tf.contrib.layers.xavier_initializer(),
+        w = tf.compat.v1.get_variable("W", filter_shape, dtype, initializer=tf.compat.v1.keras.initializers.VarianceScaling(scale=1.0, mode="fan_avg", distribution="uniform"),
                             collections=collections)
-        b = tf.get_variable("b", [1, 1, num_filters], initializer=tf.constant_initializer(0.0),
+        b = tf.compat.v1.get_variable("b", [1, 1, num_filters], initializer=tf.compat.v1.constant_initializer(0.0),
                             collections=collections)
-        return tf.nn.conv1d(x, w, stride_shape, pad) + b
+        return tf.nn.conv1d(input=x, filters=w, stride=stride_shape, padding=pad) + b
 
 
 def conv2d_dw(x, num_filters, name='conv2d_dw', filter_size=(3, 3), stride=(1, 1), pad="SAME", dtype=tf.float32,
@@ -162,7 +162,7 @@ def conv2d_dw(x, num_filters, name='conv2d_dw', filter_size=(3, 3), stride=(1, 1
     """
     Depthwise 2D convolution layer. Slow, do not use.
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         stride_shape = [1, stride[0], stride[1], 1]
         filter_shape = [filter_size[0], filter_size[1], int(x.get_shape()[-1]), num_filters]
         fan_in = np.prod(filter_shape[:3])
@@ -170,8 +170,8 @@ def conv2d_dw(x, num_filters, name='conv2d_dw', filter_size=(3, 3), stride=(1, 1
         # initialize weights with random weights
         w_bound = np.sqrt(6. / (fan_in + fan_out))
 
-        w = tf.get_variable("W", filter_shape, dtype,
-                            tf.contrib.layers.xavier_initializer(), collections=collections)
-        b = tf.get_variable("b", [1, 1, 1, num_filters * int(x.get_shape()[-1])],
-                            initializer=tf.constant_initializer(0.0), collections=collections)
-        return tf.nn.depthwise_conv2d(x, w, stride_shape, pad, [1, 1]) + b
+        w = tf.compat.v1.get_variable("W", filter_shape, dtype,
+                            tf.compat.v1.keras.initializers.VarianceScaling(scale=1.0, mode="fan_avg", distribution="uniform"), collections=collections)
+        b = tf.compat.v1.get_variable("b", [1, 1, 1, num_filters * int(x.get_shape()[-1])],
+                            initializer=tf.compat.v1.constant_initializer(0.0), collections=collections)
+        return tf.nn.depthwise_conv2d(input=x, filter=w, strides=stride_shape, padding=pad, dilations=[1, 1]) + b
diff --git a/btgym/algorithms/nn/losses.py b/btgym/algorithms/nn/losses.py
index 1592b152..dc421067 100644
--- a/btgym/algorithms/nn/losses.py
+++ b/btgym/algorithms/nn/losses.py
@@ -25,28 +25,28 @@ def aac_loss_def(act_target, adv_target, r_target, pi_logits, pi_vf, pi_prime_lo
         tensor holding estimated AAC loss;
         list of related tensorboard summaries.
     """
-    with tf.name_scope(name + '/aac'):
-        neg_pi_log_prob = tf.nn.softmax_cross_entropy_with_logits_v2(
+    with tf.compat.v1.name_scope(name + '/aac'):
+        neg_pi_log_prob = tf.nn.softmax_cross_entropy_with_logits(
             logits=pi_logits,
             labels=act_target
         )
-        pi_loss = tf.reduce_mean(neg_pi_log_prob * adv_target)
-        vf_loss = 0.5 * tf.losses.mean_squared_error(r_target, pi_vf)
-        entropy = tf.reduce_mean(cat_entropy(pi_logits))
+        pi_loss = tf.reduce_mean(input_tensor=neg_pi_log_prob * adv_target)
+        vf_loss = 0.5 * tf.compat.v1.losses.mean_squared_error(r_target, pi_vf)
+        entropy = tf.reduce_mean(input_tensor=cat_entropy(pi_logits))
 
         loss = pi_loss + vf_loss - entropy * entropy_beta
 
-        mean_vf = tf.reduce_mean(pi_vf)
-        mean_t_target = tf.reduce_mean(r_target)
+        mean_vf = tf.reduce_mean(input_tensor=pi_vf)
+        mean_t_target = tf.reduce_mean(input_tensor=r_target)
 
         summaries = [
-            tf.summary.scalar('policy_loss', pi_loss),
-            tf.summary.scalar('value_loss', vf_loss),
+            tf.compat.v1.summary.scalar('policy_loss', pi_loss),
+            tf.compat.v1.summary.scalar('value_loss', vf_loss),
         ]
         if verbose:
             summaries += [
-                tf.summary.scalar('entropy', entropy),
-                tf.summary.scalar('value_fn', mean_vf),
+                tf.compat.v1.summary.scalar('entropy', entropy),
+                tf.compat.v1.summary.scalar('value_fn', mean_vf),
                 # tf.summary.scalar('empirical_return',mean_t_target),
                 # tf.summary.histogram('value_fn', pi_vf),
                 # tf.summary.histogram('empirical_return', r_target),
@@ -79,13 +79,13 @@ def ppo_loss_def(act_target, adv_target, r_target, pi_logits, pi_vf, pi_prime_lo
     #act_target = tf.placeholder(tf.float32, [None, env.action_space.n], name="on_policy_action_pl")
     #adv_target = tf.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
     #r_target = tf.placeholder(tf.float32, [None], name="on_policy_return_pl")
-    with tf.name_scope(name + '/ppo'):
-        pi_log_prob = - tf.nn.softmax_cross_entropy_with_logits_v2(
+    with tf.compat.v1.name_scope(name + '/ppo'):
+        pi_log_prob = - tf.nn.softmax_cross_entropy_with_logits(
             logits=pi_logits,
             labels=act_target
         )
         pi_old_log_prob = tf.stop_gradient(
-            - tf.nn.softmax_cross_entropy_with_logits_v2(
+            - tf.nn.softmax_cross_entropy_with_logits(
                 logits=pi_prime_logits,
                 labels=act_target
             )
@@ -95,27 +95,27 @@ def ppo_loss_def(act_target, adv_target, r_target, pi_logits, pi_vf, pi_prime_lo
         surr1 = pi_ratio * adv_target  # surrogate from conservative policy iteration
         surr2 = tf.clip_by_value(pi_ratio, 1.0 - epsilon, 1.0 + epsilon) * adv_target
 
-        pi_surr_loss = - tf.reduce_mean(tf.minimum(surr1, surr2))  # PPO's pessimistic surrogate (L^CLIP)
-        vf_loss = tf.losses.mean_squared_error(r_target, pi_vf)  # V.fn. loss
-        entropy = tf.reduce_mean(cat_entropy(pi_logits))
+        pi_surr_loss = - tf.reduce_mean(input_tensor=tf.minimum(surr1, surr2))  # PPO's pessimistic surrogate (L^CLIP)
+        vf_loss = tf.compat.v1.losses.mean_squared_error(r_target, pi_vf)  # V.fn. loss
+        entropy = tf.reduce_mean(input_tensor=cat_entropy(pi_logits))
 
         loss = pi_surr_loss + vf_loss - entropy * entropy_beta
 
         # Info:
-        mean_pi_ratio = tf.reduce_mean(pi_ratio)
-        mean_vf = tf.reduce_mean(pi_vf)
-        mean_kl_old_new = tf.reduce_mean(kl_divergence(pi_prime_logits, pi_logits))
+        mean_pi_ratio = tf.reduce_mean(input_tensor=pi_ratio)
+        mean_vf = tf.reduce_mean(input_tensor=pi_vf)
+        mean_kl_old_new = tf.reduce_mean(input_tensor=kl_divergence(pi_prime_logits, pi_logits))
 
         summaries = [
-            tf.summary.scalar('l_clip_loss', pi_surr_loss),
-            tf.summary.scalar('value_loss', vf_loss),
+            tf.compat.v1.summary.scalar('l_clip_loss', pi_surr_loss),
+            tf.compat.v1.summary.scalar('value_loss', vf_loss),
         ]
         if verbose:
             summaries += [
-                tf.summary.scalar('entropy', entropy),
-                tf.summary.scalar('Dkl_old_new', mean_kl_old_new),
-                tf.summary.scalar('pi_ratio', mean_pi_ratio),
-                tf.summary.scalar('value_fn', mean_vf),
+                tf.compat.v1.summary.scalar('entropy', entropy),
+                tf.compat.v1.summary.scalar('Dkl_old_new', mean_kl_old_new),
+                tf.compat.v1.summary.scalar('pi_ratio', mean_pi_ratio),
+                tf.compat.v1.summary.scalar('value_fn', mean_vf),
             ]
 
     return loss, summaries
@@ -136,11 +136,11 @@ def value_fn_loss_def(r_target, pi_vf, name='_vr_', verbose=False):
         list of related tensorboard summaries.
     """
     # r_target = tf.placeholder(tf.float32, [None], name="vr_target")
-    with tf.name_scope(name + '/value_replay'):
-        loss = tf.losses.mean_squared_error(r_target, pi_vf)
+    with tf.compat.v1.name_scope(name + '/value_replay'):
+        loss = tf.compat.v1.losses.mean_squared_error(r_target, pi_vf)
 
         if verbose:
-            summaries = [tf.summary.scalar('v_loss', loss)]
+            summaries = [tf.compat.v1.summary.scalar('v_loss', loss)]
         else:
             summaries = []
 
@@ -172,17 +172,17 @@ def pc_loss_def(actions, targets, pi_pc_q, name='_pc_', verbose=False):
     """
     #actions = tf.placeholder(tf.float32, [None, env.action_space.n], name="pc_action")
     #targets = tf.placeholder(tf.float32, [None, None, None], name="pc_target")
-    with tf.name_scope(name + '/pixel_control'):
+    with tf.compat.v1.name_scope(name + '/pixel_control'):
         # Get Q-value features for actions been taken and define loss:
-        pc_action_reshaped = tf.reshape(actions, [-1, 1, 1, tf.shape(actions)[-1]])
+        pc_action_reshaped = tf.reshape(actions, [-1, 1, 1, tf.shape(input=actions)[-1]])
         pc_q_action = tf.multiply(pi_pc_q, pc_action_reshaped)
-        pc_q_action = tf.reduce_sum(pc_q_action, axis=-1, keepdims=False)
+        pc_q_action = tf.reduce_sum(input_tensor=pc_q_action, axis=-1, keepdims=False)
 
-        batch_size = tf.shape(targets)[0]
-        loss = tf.reduce_sum(tf.square(targets - pc_q_action)) / tf.cast(batch_size, tf.float32)
+        batch_size = tf.shape(input=targets)[0]
+        loss = tf.reduce_sum(input_tensor=tf.square(targets - pc_q_action)) / tf.cast(batch_size, tf.float32)
         #loss = tf.losses.absolute_difference(targets, pc_q_action)
         if verbose:
-            summaries = [tf.summary.scalar('q_loss', loss)]
+            summaries = [tf.compat.v1.summary.scalar('q_loss', loss)]
         else:
             summaries = []
 
@@ -213,13 +213,13 @@ def rp_loss_def(rp_targets, pi_rp_logits, name='_rp_', verbose=False):
         list of related tensorboard summaries.
     """
     #rp_targets = tf.placeholder(tf.float32, [1, 3], name="rp_target")
-    with tf.name_scope(name + '/reward_prediction'):
-        loss = tf.nn.softmax_cross_entropy_with_logits_v2(
+    with tf.compat.v1.name_scope(name + '/reward_prediction'):
+        loss = tf.nn.softmax_cross_entropy_with_logits(
             labels=rp_targets,
             logits=pi_rp_logits
         )[0]
         if verbose:
-            summaries = [tf.summary.scalar('class_loss', loss), ]
+            summaries = [tf.compat.v1.summary.scalar('class_loss', loss), ]
         else:
             summaries = []
 
@@ -241,11 +241,11 @@ def ae_loss_def(targets, logits, alpha=1.0, name='ae_loss', verbose=False, **kwa
         tensor holding estimated reconstruction loss
         list of summarues
     """
-    with tf.name_scope(name + '/ae'):
-        loss = tf.losses.mean_squared_error(targets, logits)
+    with tf.compat.v1.name_scope(name + '/ae'):
+        loss = tf.compat.v1.losses.mean_squared_error(targets, logits)
 
         if verbose:
-            summaries = [tf.summary.scalar('reconstruct_loss', loss)]
+            summaries = [tf.compat.v1.summary.scalar('reconstruct_loss', loss)]
         else:
             summaries = []
 
@@ -274,14 +274,14 @@ def beta_vae_loss_def(targets, logits, d_kl, alpha=1.0, beta=1.0, name='beta_vae
         list of summarues
 
     """
-    with tf.name_scope(name + '/b_vae'):
-        r_loss = tf.losses.mean_squared_error(targets, logits)
-        vae_loss = tf.reduce_mean(d_kl)
+    with tf.compat.v1.name_scope(name + '/b_vae'):
+        r_loss = tf.compat.v1.losses.mean_squared_error(targets, logits)
+        vae_loss = tf.reduce_mean(input_tensor=d_kl)
         loss = alpha * r_loss + beta * vae_loss
         if verbose:
             summaries = [
-                tf.summary.scalar('reconstruct_loss', r_loss),
-                tf.summary.scalar('d_kl_loss', vae_loss),
+                tf.compat.v1.summary.scalar('reconstruct_loss', r_loss),
+                tf.compat.v1.summary.scalar('d_kl_loss', vae_loss),
             ]
         else:
             summaries = []
diff --git a/btgym/algorithms/nn/networks.py b/btgym/algorithms/nn/networks.py
index a077db07..2d998db9 100644
--- a/btgym/algorithms/nn/networks.py
+++ b/btgym/algorithms/nn/networks.py
@@ -36,7 +36,7 @@ def conv_2d_network(x,
     Returns:
         tensor holding state features;
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         for i, num_filters in enumerate(conv_2d_num_filters):
             x = tf.nn.elu(
                 norm_layer(
@@ -55,7 +55,7 @@ def conv_2d_network(x,
                 )
             )
             if keep_prob is not None:
-                x = tf.nn.dropout(x, keep_prob=keep_prob, name="_layer_{}_with_dropout".format(i + 1))
+                x = tf.nn.dropout(x, rate=1 - (keep_prob), name="_layer_{}_with_dropout".format(i + 1))
 
         # A3c/BaseAAC original paper design:
         # x = tf.nn.elu(conv2d(x, 16, 'conv2d_1', [8, 8], [4, 4], pad, dtype, collections, reuse))
@@ -123,27 +123,27 @@ def lstm_network(
          lstm state output tensor;
          lstm flattened feed placeholders as tuple.
     """
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         # Prepare rnn type:
         if static:
-            rnn_net = tf.nn.static_rnn
+            rnn_net = tf.compat.v1.nn.static_rnn
             # Remove time dimension (suppose always get one) and wrap to list:
             x = [x[:, 0, :]]
 
         else:
-            rnn_net = tf.nn.dynamic_rnn
+            rnn_net = tf.compat.v1.nn.dynamic_rnn
         # Define LSTM layers:
         lstm = []
         for size in lstm_layers:
             layer = lstm_class(size)
             if keep_prob is not None:
-                layer = tf.nn.rnn_cell.DropoutWrapper(layer, output_keep_prob=keep_prob)
+                layer = tf.compat.v1.nn.rnn_cell.DropoutWrapper(layer, output_keep_prob=keep_prob)
 
             lstm.append(layer)
 
         lstm = rnn.MultiRNNCell(lstm, state_is_tuple=True)
         # Get time_dimension as [1]-shaped tensor:
-        step_size = tf.expand_dims(tf.shape(x)[1], [0])
+        step_size = tf.expand_dims(tf.shape(input=x)[1], [0])
 
         lstm_init_state = lstm.zero_state(1, dtype=tf.float32)
 
@@ -183,7 +183,7 @@ def dense_aac_network(x, ac_space_depth, name='dense_aac', linear_layer_ref=nois
         for every space in ac_space_shape dictionary
     """
 
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         # Center-logits:
         logits = norm_layer(
             linear_layer_ref(
@@ -247,8 +247,8 @@ def pixel_change_2d_estimator(ob_space, pc_estimator_stride=(2, 2), **kwargs):
     Note:
         crops input array by one pix from either side; --> 1D signal to be shaped as [signal_length, 3]
     """
-    input_state = tf.placeholder(tf.float32, list(ob_space), name='pc_change_est_state_in')
-    input_last_state = tf.placeholder(tf.float32, list(ob_space), name='pc_change_est_last_state_in')
+    input_state = tf.compat.v1.placeholder(tf.float32, list(ob_space), name='pc_change_est_state_in')
+    input_last_state = tf.compat.v1.placeholder(tf.float32, list(ob_space), name='pc_change_est_last_state_in')
 
     x = tf.abs(tf.subtract(input_state, input_last_state)) # TODO: tf.square?
 
@@ -258,13 +258,13 @@ def pixel_change_2d_estimator(ob_space, pc_estimator_stride=(2, 2), **kwargs):
     else:
         x = tf.expand_dims(x, 0)[:, 1:-1, 1:-1, :]  # True 2D,  fake batch dim and crop H, W dims
 
-    x = tf.reduce_mean(x, axis=-1, keepdims=True)
+    x = tf.reduce_mean(input_tensor=x, axis=-1, keepdims=True)
 
-    x_out = tf.nn.max_pool(
-        x,
-        [1, pc_estimator_stride[0], pc_estimator_stride[1], 1],
-        [1, pc_estimator_stride[0], pc_estimator_stride[1], 1],
-        'SAME'
+    x_out = tf.nn.max_pool2d(
+        input=x,
+        ksize=[1, pc_estimator_stride[0], pc_estimator_stride[1], 1],
+        strides=[1, pc_estimator_stride[0], pc_estimator_stride[1], 1],
+        padding='SAME'
     )
     return input_state, input_last_state, x_out
 
@@ -285,7 +285,7 @@ def duelling_pc_network(x,
             x=x,
             size=np.prod(duell_pc_x_inner_shape),
             name='pc_dense',
-            initializer=tf.contrib.layers.xavier_initializer(),
+            initializer=tf.compat.v1.keras.initializers.VarianceScaling(scale=1.0, mode="fan_avg", distribution="uniform"),
             reuse=reuse
         )
     )
@@ -296,7 +296,7 @@ def duelling_pc_network(x,
     # Q-value estimate using advantage mean,
     # as (9) in "Dueling Network Architectures..." paper:
     # https://arxiv.org/pdf/1511.06581.pdf
-    pc_a_mean = tf.reduce_mean(pc_a, axis=-1, keepdims=True)
+    pc_a_mean = tf.reduce_mean(input_tensor=pc_a, axis=-1, keepdims=True)
     pc_q = pc_v + pc_a - pc_a_mean  # [None, 20, 20, ac_size]
 
     return pc_q
diff --git a/btgym/algorithms/policy/base.py b/btgym/algorithms/policy/base.py
index 91c07243..a7dd4b93 100644
--- a/btgym/algorithms/policy/base.py
+++ b/btgym/algorithms/policy/base.py
@@ -78,33 +78,33 @@ def __init__(self,
         self.rp_state_in = nested_placeholders(self.ob_space.shape, batch_dim=None, name='rp_state_in')
 
         # Placeholders for previous step action[multi-categorical vector encoding]  and reward [scalar]:
-        self.on_last_a_in = tf.placeholder(
+        self.on_last_a_in = tf.compat.v1.placeholder(
             tf.float32,
             [None, self.ac_space.encoded_depth],
             name='on_policy_last_action_in_pl'
         )
-        self.on_last_reward_in = tf.placeholder(tf.float32, [None], name='on_policy_last_reward_in_pl')
+        self.on_last_reward_in = tf.compat.v1.placeholder(tf.float32, [None], name='on_policy_last_reward_in_pl')
 
-        self.off_last_a_in = tf.placeholder(
+        self.off_last_a_in = tf.compat.v1.placeholder(
             tf.float32,
             [None, self.ac_space.encoded_depth],
             name='off_policy_last_action_in_pl'
         )
-        self.off_last_reward_in = tf.placeholder(tf.float32, [None], name='off_policy_last_reward_in_pl')
+        self.off_last_reward_in = tf.compat.v1.placeholder(tf.float32, [None], name='off_policy_last_reward_in_pl')
 
         # Placeholders for rnn batch and time-step dimensions:
-        self.on_batch_size = tf.placeholder(tf.int32, name='on_policy_batch_size')
-        self.on_time_length = tf.placeholder(tf.int32, name='on_policy_sequence_size')
+        self.on_batch_size = tf.compat.v1.placeholder(tf.int32, name='on_policy_batch_size')
+        self.on_time_length = tf.compat.v1.placeholder(tf.int32, name='on_policy_sequence_size')
 
-        self.off_batch_size = tf.placeholder(tf.int32, name='off_policy_batch_size')
-        self.off_time_length = tf.placeholder(tf.int32, name='off_policy_sequence_size')
+        self.off_batch_size = tf.compat.v1.placeholder(tf.int32, name='off_policy_batch_size')
+        self.off_time_length = tf.compat.v1.placeholder(tf.int32, name='off_policy_sequence_size')
 
         try:
             if self.train_phase is not None:
                 pass
 
         except AttributeError:
-            self.train_phase = tf.placeholder_with_default(
+            self.train_phase = tf.compat.v1.placeholder_with_default(
                 tf.constant(False, dtype=tf.bool),
                 shape=(),
                 name='train_phase_flag_pl'
@@ -114,7 +114,7 @@ def __init__(self,
         on_aac_x = conv_2d_network(self.on_state_in['external'], self.ob_space.shape['external'], ac_space, **kwargs)
 
         # Reshape rnn inputs for  batch training as [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(on_aac_x)
+        x_shape_dynamic = tf.shape(input=on_aac_x)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.on_batch_size, tf.int32)
         x_shape_static = on_aac_x.get_shape().as_list()
 
@@ -167,7 +167,7 @@ def __init__(self,
         off_aac_x = conv_2d_network(self.off_state_in['external'], self.ob_space.shape['external'], ac_space, reuse=True, **kwargs)
 
         # Reshape rnn inputs for  batch training as [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(off_aac_x)
+        x_shape_dynamic = tf.shape(input=off_aac_x)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.off_batch_size, tf.int32)
         x_shape_static = off_aac_x.get_shape().as_list()
 
@@ -239,7 +239,7 @@ def __init__(self,
         self.vr_value = self.off_vf
 
         # Aux3: `Reward prediction` network:
-        self.rp_batch_size = tf.placeholder(tf.int32, name='rp_batch_size')
+        self.rp_batch_size = tf.compat.v1.placeholder(tf.int32, name='rp_batch_size')
 
         # Shared conv. output:
         rp_x = conv_2d_network(self.rp_state_in['external'], self.ob_space.shape['external'], ac_space, reuse=True, **kwargs)
@@ -252,13 +252,13 @@ def __init__(self,
         self.rp_logits = dense_rp_network(rp_x)
 
         # Batch-norm related :
-        self.update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        self.update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
         # Add moving averages to save list:
-        moving_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*moving.*')
-        renorm_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*renorm.*')
+        moving_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*moving.*')
+        renorm_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*renorm.*')
 
         # What to save:
-        self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, tf.get_variable_scope().name)
+        self.var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, tf.compat.v1.get_variable_scope().name)
         self.var_list += moving_var_list + renorm_var_list
 
         # Callbacks:
@@ -273,7 +273,7 @@ def get_initial_features(self, **kwargs):
             LSTM zero-state tuple.
         """
         # TODO: rework as in: AacStackedMetaPolicy --> base runner, verbose runner; synchro_runner ok
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         return sess.run(self.on_lstm_init_state)
 
     def act(self, observation, lstm_state, last_action, last_reward, deterministic=False):
@@ -293,7 +293,7 @@ def act(self, observation, lstm_state, last_action, last_reward, deterministic=F
             Action as dictionary of several action encodings, actions logits, V-fn value, output RNN state
         """
         try:
-            sess = tf.get_default_session()
+            sess = tf.compat.v1.get_default_session()
             feeder = {pl: value for pl, value in zip(self.on_lstm_state_pl_flatten, flatten_nested(lstm_state))}
             feeder.update(feed_dict_from_nested(self.on_state_in, observation, expand_batch=True))
             feeder.update(
@@ -353,7 +353,7 @@ def get_value(self, observation, lstm_state, last_action, last_reward):
         Returns:
             V-function value
         """
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         feeder = feed_dict_rnn_context(self.on_lstm_state_pl_flatten, lstm_state)
         feeder.update(feed_dict_from_nested(self.on_state_in, observation, expand_batch=True))
         feeder.update(
@@ -380,7 +380,7 @@ def get_pc_target(self, state, last_state, **kwargs):
         Returns:
             Estimated absolute difference between two subsampled states.
         """
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         feeder = {self.pc_change_state_in: state['external'], self.pc_change_last_state_in: last_state['external']}
 
         return sess.run(self.pc_target, feeder)[0,...,0]
diff --git a/btgym/algorithms/policy/meta.py b/btgym/algorithms/policy/meta.py
index d84ff43e..37ea0fdd 100644
--- a/btgym/algorithms/policy/meta.py
+++ b/btgym/algorithms/policy/meta.py
@@ -19,14 +19,14 @@ def __init__(self, task, num_host_policies, learn_rate, name='SubMetaPolicy'):
             learn_rate:             meta-policy learning rate
             name:                   name scope
         """
-        with tf.variable_scope(name_or_scope=name):
+        with tf.compat.v1.variable_scope(name_or_scope=name):
             self.task = task
             self.learn_rate = learn_rate
             self.num_host_policies = num_host_policies
 
-            self.input_stat_pl = tf.placeholder(dtype=tf.float32, name='in_stat_pl')
+            self.input_stat_pl = tf.compat.v1.placeholder(dtype=tf.float32, name='in_stat_pl')
 
-            self.input_stat = tf.reduce_mean(self.input_stat_pl)
+            self.input_stat = tf.reduce_mean(input_tensor=self.input_stat_pl)
 
             self.initial_cluster_value = tf.concat(
                 [
@@ -43,38 +43,38 @@ def __init__(self, task, num_host_policies, learn_rate, name='SubMetaPolicy'):
                 name='cluster_wide_averages_slot'
             )
 
-            update_task_iteration = tf.scatter_nd_add(self.cluster_averages_slot, [[0, task]], [1])
+            update_task_iteration = tf.compat.v1.scatter_nd_add(self.cluster_averages_slot, [[0, task]], [1])
 
             with tf.control_dependencies([update_task_iteration]):
                 avg_prev = self.cluster_averages_slot[1, task]
                 k = self.cluster_averages_slot[0, task]
                 avg = avg_prev + (self.input_stat - avg_prev) / k
-                self.update_op = tf.scatter_nd_update(self.cluster_averages_slot, [[1, task]], [avg])
+                self.update_op = tf.compat.v1.scatter_nd_update(self.cluster_averages_slot, [[1, task]], [avg])
 
-            self.reset_op = tf.assign(
+            self.reset_op = tf.compat.v1.assign(
                 self.cluster_averages_slot,
                 self.initial_cluster_value
             )
 
             # Toy network:
-            prob = tf.layers.dense(
+            prob = tf.compat.v1.layers.dense(
                 tf.expand_dims(self.cluster_averages_slot[1, :], axis=-1),
                 units=10,
                 activation=tf.nn.sigmoid,
                 use_bias=False,
             )
-            self.next_step_prob = tf.layers.dense(
+            self.next_step_prob = tf.compat.v1.layers.dense(
                 prob,
                 units=1,
                 activation=tf.nn.sigmoid,
                 use_bias=False,
             )
-            self.distribution = tf.distributions.Bernoulli(
-                probs=tf.reduce_max(self.next_step_prob)
+            self.distribution = tf.compat.v1.distributions.Bernoulli(
+                probs=tf.reduce_max(input_tensor=self.next_step_prob)
             )
             self.sample = self.distribution.sample()
 
-            self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, tf.get_variable_scope().name)
+            self.var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, tf.compat.v1.get_variable_scope().name)
 
             self.cluster_stat = tf.clip_by_value(
                 # tf.reduce_mean(
@@ -86,26 +86,26 @@ def __init__(self, task, num_host_policies, learn_rate, name='SubMetaPolicy'):
             )
             bound_avg = tf.sigmoid(- self.cluster_stat)
             self.loss = tf.reduce_mean(
-                bound_avg * (1 - self.next_step_prob) + (1 - bound_avg) * self.next_step_prob
+                input_tensor=bound_avg * (1 - self.next_step_prob) + (1 - bound_avg) * self.next_step_prob
             )
-            self.grads = tf.gradients(self.loss, self.var_list)
+            self.grads = tf.gradients(ys=self.loss, xs=self.var_list)
 
             self.summaries = [
-                tf.summary.scalar('worker_avg_stat', self.cluster_averages_slot[1, task]),
-                tf.summary.scalar('worker_iterations', self.cluster_averages_slot[0, task]),
+                tf.compat.v1.summary.scalar('worker_avg_stat', self.cluster_averages_slot[1, task]),
+                tf.compat.v1.summary.scalar('worker_iterations', self.cluster_averages_slot[0, task]),
                 #tf.summary.histogram('clipped_cluster_stat', self.cluster_stat),
-                tf.summary.scalar('loss', self.loss),
-                tf.summary.histogram('next_step_prob', self.next_step_prob),
-                tf.summary.scalar('grads_norm', tf.global_norm(self.grads))
+                tf.compat.v1.summary.scalar('loss', self.loss),
+                tf.compat.v1.summary.histogram('next_step_prob', self.next_step_prob),
+                tf.compat.v1.summary.scalar('grads_norm', tf.linalg.global_norm(self.grads))
             ]
 
     def update(self, input_stat):
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         feed_dict = {self.input_stat_pl: input_stat}
         sess.run(self.update_op, feed_dict)
 
     def reset(self):
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         sess.run(self.reset_op)
 
     def global_reset(self):
@@ -122,7 +122,7 @@ def act(self):
         Returns:
 
         """
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         fetched = sess.run([self.sample])
 
         return fetched[-1]
diff --git a/btgym/algorithms/policy/stacked_lstm.py b/btgym/algorithms/policy/stacked_lstm.py
index c4750a3a..160728e8 100644
--- a/btgym/algorithms/policy/stacked_lstm.py
+++ b/btgym/algorithms/policy/stacked_lstm.py
@@ -77,7 +77,7 @@ def __init__(self,
         # self.encode_internal_state = encode_internal_state
         self.share_encoder_params = share_encoder_params
         if self.share_encoder_params:
-            self.reuse_encoder_params = tf.AUTO_REUSE
+            self.reuse_encoder_params = tf.compat.v1.AUTO_REUSE
 
         else:
             self.reuse_encoder_params = False
@@ -93,26 +93,26 @@ def __init__(self,
         self.rp_state_in = nested_placeholders(self.ob_space.shape, batch_dim=None, name='rp_state_in')
 
         # Placeholders for previous step action[multi-categorical vector encoding]  and reward [scalar]:
-        self.on_last_a_in = tf.placeholder(
+        self.on_last_a_in = tf.compat.v1.placeholder(
             tf.float32,
             [None, self.ac_space.encoded_depth],
             name='on_policy_last_action_in_pl'
         )
-        self.on_last_reward_in = tf.placeholder(tf.float32, [None], name='on_policy_last_reward_in_pl')
+        self.on_last_reward_in = tf.compat.v1.placeholder(tf.float32, [None], name='on_policy_last_reward_in_pl')
 
-        self.off_last_a_in = tf.placeholder(
+        self.off_last_a_in = tf.compat.v1.placeholder(
             tf.float32,
             [None, self.ac_space.encoded_depth],
             name='off_policy_last_action_in_pl'
         )
-        self.off_last_reward_in = tf.placeholder(tf.float32, [None], name='off_policy_last_reward_in_pl')
+        self.off_last_reward_in = tf.compat.v1.placeholder(tf.float32, [None], name='off_policy_last_reward_in_pl')
 
         # Placeholders for rnn batch and time-step dimensions:
-        self.on_batch_size = tf.placeholder(tf.int32, name='on_policy_batch_size')
-        self.on_time_length = tf.placeholder(tf.int32, name='on_policy_sequence_size')
+        self.on_batch_size = tf.compat.v1.placeholder(tf.int32, name='on_policy_batch_size')
+        self.on_time_length = tf.compat.v1.placeholder(tf.int32, name='on_policy_sequence_size')
 
-        self.off_batch_size = tf.placeholder(tf.int32, name='off_policy_batch_size')
-        self.off_time_length = tf.placeholder(tf.int32, name='off_policy_sequence_size')
+        self.off_batch_size = tf.compat.v1.placeholder(tf.int32, name='off_policy_batch_size')
+        self.off_time_length = tf.compat.v1.placeholder(tf.int32, name='off_policy_sequence_size')
 
         self.debug['on_state_in_keys'] = list(self.on_state_in.keys())
 
@@ -122,7 +122,7 @@ def __init__(self,
                 pass
 
         except AttributeError:
-            self.train_phase = tf.placeholder_with_default(
+            self.train_phase = tf.compat.v1.placeholder_with_default(
                 tf.constant(False, dtype=tf.bool),
                 shape=(),
                 name='train_phase_flag_pl'
@@ -161,7 +161,7 @@ def __init__(self,
                 else:
                     layer_name_template = 'encoded_{}_{}'
                 encoded_streams = {
-                    name: tf.layers.flatten(
+                    name: tf.compat.v1.layers.flatten(
                         self.state_encoder_class_ref(
                             x=stream,
                             ob_space=self.ob_space.shape[key][name],
@@ -180,7 +180,7 @@ def __init__(self,
                 )
             else:
                 # Got single data stream:
-                encoded_mode = tf.layers.flatten(
+                encoded_mode = tf.compat.v1.layers.flatten(
                     self.state_encoder_class_ref(
                         x=self.on_state_in[key],
                         ob_space=self.ob_space.shape[key],
@@ -200,14 +200,14 @@ def __init__(self,
 
         # TODO: for encoder prediction test, output `naive` estimates for logits and value directly from encoder:
         [self.on_simple_logits, self.on_simple_value, _] = dense_aac_network(
-            tf.layers.flatten(on_aac_x),
+            tf.compat.v1.layers.flatten(on_aac_x),
             ac_space_depth=self.ac_space.one_hot_depth,
             linear_layer_ref=linear_layer_ref,
             name='aac_dense_simple_pi_v'
         )
 
         # Reshape rnn inputs for batch training as: [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(on_aac_x)
+        x_shape_dynamic = tf.shape(input=on_aac_x)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.on_batch_size, tf.int32)
         x_shape_static = on_aac_x.get_shape().as_list()
 
@@ -308,7 +308,7 @@ def __init__(self,
         self.debug['self.on_lstm_1_state_pl_flatten'] = self.on_lstm_1_state_pl_flatten
 
         # For time_flat only: Reshape on_lstm_1_state_out from [1,2,20,size] -->[20,1,2,size] --> [20,1, 2xsize]:
-        reshape_lstm_1_state_out = tf.transpose(self.on_lstm_1_state_out, [2, 0, 1, 3])
+        reshape_lstm_1_state_out = tf.transpose(a=self.on_lstm_1_state_out, perm=[2, 0, 1, 3])
         reshape_lstm_1_state_out_shape_static = reshape_lstm_1_state_out.get_shape().as_list()
 
         # Take policy logits off first LSTM-dense layer:
@@ -390,7 +390,7 @@ def __init__(self,
                 else:
                     layer_name_template = 'encoded_{}_{}'
                 encoded_streams = {
-                    name: tf.layers.flatten(
+                    name: tf.compat.v1.layers.flatten(
                         self.state_encoder_class_ref(
                             x=stream,
                             ob_space=self.ob_space.shape[key][name],
@@ -409,7 +409,7 @@ def __init__(self,
                 )
             else:
                 # Got single data stream:
-                encoded_mode = tf.layers.flatten(
+                encoded_mode = tf.compat.v1.layers.flatten(
                     self.state_encoder_class_ref(
                         x=self.off_state_in[key],
                         ob_space=self.ob_space.shape[key],
@@ -426,7 +426,7 @@ def __init__(self,
         off_aac_x = self.off_aac_x_encoded['external']
 
         # Reshape rnn inputs for  batch training as [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(off_aac_x)
+        x_shape_dynamic = tf.shape(input=off_aac_x)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.off_batch_size, tf.int32)
         x_shape_static = off_aac_x.get_shape().as_list()
 
@@ -602,7 +602,7 @@ def __init__(self,
 
         # Aux3:
         # `Reward prediction` network.
-        self.rp_batch_size = tf.placeholder(tf.int32, name='rp_batch_size')
+        self.rp_batch_size = tf.compat.v1.placeholder(tf.int32, name='rp_batch_size')
 
         # Shared encoded output:
         rp_x = {}
@@ -614,7 +614,7 @@ def __init__(self,
                     else:
                         layer_name_template = 'encoded_{}_{}'
                     encoded_streams = {
-                        name: tf.layers.flatten(
+                        name: tf.compat.v1.layers.flatten(
                             self.state_encoder_class_ref(
                                 x=stream,
                                 ob_space=self.ob_space.shape[key][name],
@@ -633,7 +633,7 @@ def __init__(self,
                     )
                 else:
                     # Got single data stream:
-                    encoded_mode = tf.layers.flatten(
+                    encoded_mode = tf.compat.v1.layers.flatten(
                         self.state_encoder_class_ref(
                             x=self.rp_state_in[key],
                             ob_space=self.ob_space.shape,
@@ -655,13 +655,13 @@ def __init__(self,
         self.rp_logits = dense_rp_network(rp_x, linear_layer_ref=linear_layer_ref)
 
         # Batch-norm related:
-        self.update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        self.update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
         # Add moving averages to save list:
-        moving_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*moving.*')
-        renorm_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*renorm.*')
+        moving_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*moving.*')
+        renorm_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*renorm.*')
 
         # What to save:
-        self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, tf.get_variable_scope().name)
+        self.var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, tf.compat.v1.get_variable_scope().name)
         self.var_list += moving_var_list + renorm_var_list
 
         # Callbacks:
@@ -721,7 +721,7 @@ def get_initial_features(self, state, context=None):
             KeyError if [`metadata`]:[`trial_num`,`type`] keys not found
         """
         try:
-            sess = tf.get_default_session()
+            sess = tf.compat.v1.get_default_session()
             new_context = list(sess.run(self.on_lstm_init_state))
             if state['metadata']['trial_num'] != self.current_trial_num\
                     or context is None\
diff --git a/btgym/algorithms/utils.py b/btgym/algorithms/utils.py
index 34aae914..da323aea 100644
--- a/btgym/algorithms/utils.py
+++ b/btgym/algorithms/utils.py
@@ -19,14 +19,14 @@ def rnn_placeholders(state):
 
     Returns:    tuple of placeholders
     """
-    if isinstance(state, tf.contrib.rnn.LSTMStateTuple):
+    if isinstance(state, tf.nn.rnn_cell.LSTMStateTuple):
         c, h = state
-        c = tf.placeholder(tf.float32, tf.TensorShape([None]).concatenate(c.get_shape()[1:]), c.op.name + '_c_pl')
-        h = tf.placeholder(tf.float32, tf.TensorShape([None]).concatenate(h.get_shape()[1:]), h.op.name + '_h_pl')
-        return tf.contrib.rnn.LSTMStateTuple(c, h)
+        c = tf.compat.v1.placeholder(tf.float32, tf.TensorShape([None]).concatenate(c.get_shape()[1:]), c.op.name + '_c_pl')
+        h = tf.compat.v1.placeholder(tf.float32, tf.TensorShape([None]).concatenate(h.get_shape()[1:]), h.op.name + '_h_pl')
+        return tf.nn.rnn_cell.LSTMStateTuple(c, h)
     elif isinstance(state, tf.Tensor):
         h = state
-        h = tf.placeholder(tf.float32, tf.TensorShape([None]).concatenate(h.get_shape()[1:]), h.op.name + '_h_pl')
+        h = tf.compat.v1.placeholder(tf.float32, tf.TensorShape([None]).concatenate(h.get_shape()[1:]), h.op.name + '_h_pl')
         return h
     else:
         structure = [rnn_placeholders(x) for x in state]
@@ -49,7 +49,7 @@ def nested_placeholders(ob_space, batch_dim=None, name='nested'):
         out = {key: nested_placeholders(value, batch_dim, name + '_' + key) for key, value in ob_space.items()}
         return out
     else:
-        out = tf.placeholder(tf.float32, [batch_dim] + list(ob_space), name + '_pl')
+        out = tf.compat.v1.placeholder(tf.float32, [batch_dim] + list(ob_space), name + '_pl')
         return out
 
 
diff --git a/btgym/algorithms/worker.py b/btgym/algorithms/worker.py
index 5eec019a..76bf0865 100644
--- a/btgym/algorithms/worker.py
+++ b/btgym/algorithms/worker.py
@@ -16,10 +16,10 @@
 import tensorflow as tf
 
 sys.path.insert(0, '..')
-tf.logging.set_verbosity(tf.logging.INFO)
+tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.INFO)
 
 
-class FastSaver(tf.train.Saver):
+class FastSaver(tf.compat.v1.train.Saver):
     """
     Disables write_meta_graph argument,
     which freezes entire process and is mostly useless.
@@ -188,7 +188,7 @@ def run(self):
         StreamHandler(sys.stdout).push_application()
         self.log = Logger('Worker_{}'.format(self.task), level=self.log_level)
         try:
-            tf.reset_default_graph()
+            tf.compat.v1.reset_default_graph()
 
             if self.test_mode:
                 import gym
@@ -198,22 +198,22 @@ def run(self):
 
             # Start tf.server:
             if self.job_name in 'ps':
-                server = tf.train.Server(
+                server = tf.distribute.Server(
                     cluster,
                     job_name=self.job_name,
                     task_index=self.task,
-                    config=tf.ConfigProto(device_filters=["/job:ps"])
+                    config=tf.compat.v1.ConfigProto(device_filters=["/job:ps"])
                 )
                 self.log.debug('parameters_server started.')
                 # Just block here:
                 server.join()
 
             else:
-                server = tf.train.Server(
+                server = tf.distribute.Server(
                     cluster,
                     job_name='worker',
                     task_index=self.task,
-                    config=tf.ConfigProto(
+                    config=tf.compat.v1.ConfigProto(
                         intra_op_parallelism_threads=4,  # original was: 1
                         inter_op_parallelism_threads=4,  # original was: 2
                     )
@@ -308,11 +308,11 @@ def run(self):
                 self.log.debug('trainer ok.')
 
                 # Saver-related:
-                variables_to_save = [v for v in tf.global_variables() if not 'local' in v.name]
-                local_variables = [v for v in tf.global_variables() if 'local' in v.name] + tf.local_variables()
-                init_op = tf.initializers.variables(variables_to_save)
-                local_init_op = tf.initializers.variables(local_variables)
-                init_all_op = tf.global_variables_initializer()
+                variables_to_save = [v for v in tf.compat.v1.global_variables() if not 'local' in v.name]
+                local_variables = [v for v in tf.compat.v1.global_variables() if 'local' in v.name] + tf.compat.v1.local_variables()
+                init_op = tf.compat.v1.initializers.variables(variables_to_save)
+                local_init_op = tf.compat.v1.initializers.variables(local_variables)
+                init_all_op = tf.compat.v1.global_variables_initializer()
 
                 def init_fn(_sess):
                     self.log.notice("initializing all parameters...")
@@ -332,12 +332,12 @@ def init_fn(_sess):
 
                 self.saver = FastSaver(var_list=variables_to_save, max_to_keep=1, save_relative_paths=True)
 
-                self.config = tf.ConfigProto(device_filters=["/job:ps", "/job:worker/task:{}/cpu:0".format(self.task)])
+                self.config = tf.compat.v1.ConfigProto(device_filters=["/job:ps", "/job:worker/task:{}/cpu:0".format(self.task)])
 
-                sess_manager = tf.train.SessionManager(
+                sess_manager = tf.compat.v1.train.SessionManager(
                     local_init_op=local_init_op,
                     ready_op=None,
-                    ready_for_local_init_op=tf.report_uninitialized_variables(variables_to_save),
+                    ready_for_local_init_op=tf.compat.v1.report_uninitialized_variables(variables_to_save),
                     graph=None,
                     recovery_wait_secs=90,
                 )
@@ -364,7 +364,7 @@ def init_fn(_sess):
 
                     self.log.info("connecting to the parameter server... ")
 
-                    self.summary_writer = tf.summary.FileWriter(self.summary_dir, sess.graph)
+                    self.summary_writer = tf.compat.v1.summary.FileWriter(self.summary_dir, sess.graph)
                     trainer.start(sess, self.summary_writer)
 
                     # Note: `self.global_step` refers to number of environment steps
diff --git a/btgym/monitor/tensorboard2.py b/btgym/monitor/tensorboard2.py
index e1cad240..93021bd3 100644
--- a/btgym/monitor/tensorboard2.py
+++ b/btgym/monitor/tensorboard2.py
@@ -69,7 +69,7 @@ def __init__(self,
             p = psutil.Popen(['rm', '-R', ] + files, stdout=PIPE, stderr=PIPE)
 
         # Prepare writer:
-        self.writer = tf.summary.FileWriter(self.logdir, graph=tf.get_default_graph())
+        self.writer = tf.compat.v1.summary.FileWriter(self.logdir, graph=tf.compat.v1.get_default_graph())
 
 
         # Create summary:
@@ -77,25 +77,25 @@ def __init__(self,
 
         for entry in scalars:
             assert type(entry) == str
-            self.feed_holder[entry] = tf.placeholder(tf.float32)
-            summaries += [tf.summary.scalar(entry, self.feed_holder[entry],)]
+            self.feed_holder[entry] = tf.compat.v1.placeholder(tf.float32)
+            summaries += [tf.compat.v1.summary.scalar(entry, self.feed_holder[entry],)]
 
         for entry in images:
             assert type(entry) == str
-            self.feed_holder[entry] = tf.placeholder(tf.uint8, [None, None, None, 3])
-            summaries += [tf.summary.image(entry, self.feed_holder[entry], )]
+            self.feed_holder[entry] = tf.compat.v1.placeholder(tf.uint8, [None, None, None, 3])
+            summaries += [tf.compat.v1.summary.image(entry, self.feed_holder[entry], )]
 
         for entry in histograms:
             assert type(entry) == str
-            self.feed_holder[entry] = tf.placeholder(tf.float32,[None, None],)
-            summaries += [tf.summary.histogram(entry, self.feed_holder[entry], )]
+            self.feed_holder[entry] = tf.compat.v1.placeholder(tf.float32,[None, None],)
+            summaries += [tf.compat.v1.summary.histogram(entry, self.feed_holder[entry], )]
 
         for entry in text:
             assert type(entry) == str
-            self.feed_holder[entry] = tf.placeholder(tf.string)
-            summaries += [tf.summary.histogram(entry, self.feed_holder[entry], )]
+            self.feed_holder[entry] = tf.compat.v1.placeholder(tf.string)
+            summaries += [tf.compat.v1.summary.histogram(entry, self.feed_holder[entry], )]
 
-        self.summary = tf.summary.merge(summaries)
+        self.summary = tf.compat.v1.summary.merge(summaries)
 
     def write(self, sess, feed_dict, global_step):
         """
diff --git a/btgym/research/b_vae_a3c.py b/btgym/research/b_vae_a3c.py
index d6e13097..b1dc90fb 100644
--- a/btgym/research/b_vae_a3c.py
+++ b/btgym/research/b_vae_a3c.py
@@ -107,15 +107,15 @@ def __init__(self,
         self.rp_state_in = nested_placeholders(ob_space, batch_dim=None, name='rp_state_in')
 
         # Placeholders for concatenated action [one-hot] and reward [scalar]:
-        self.on_a_r_in = tf.placeholder(tf.float32, [None, ac_space + 1], name='on_policy_action_reward_in_pl')
-        self.off_a_r_in = tf.placeholder(tf.float32, [None, ac_space + 1], name='off_policy_action_reward_in_pl')
+        self.on_a_r_in = tf.compat.v1.placeholder(tf.float32, [None, ac_space + 1], name='on_policy_action_reward_in_pl')
+        self.off_a_r_in = tf.compat.v1.placeholder(tf.float32, [None, ac_space + 1], name='off_policy_action_reward_in_pl')
 
         # Placeholders for rnn batch and time-step dimensions:
-        self.on_batch_size = tf.placeholder(tf.int32, name='on_policy_batch_size')
-        self.on_time_length = tf.placeholder(tf.int32, name='on_policy_sequence_size')
+        self.on_batch_size = tf.compat.v1.placeholder(tf.int32, name='on_policy_batch_size')
+        self.on_time_length = tf.compat.v1.placeholder(tf.int32, name='on_policy_sequence_size')
 
-        self.off_batch_size = tf.placeholder(tf.int32, name='off_policy_batch_size')
-        self.off_time_length = tf.placeholder(tf.int32, name='off_policy_sequence_size')
+        self.off_batch_size = tf.compat.v1.placeholder(tf.int32, name='off_policy_batch_size')
+        self.off_time_length = tf.compat.v1.placeholder(tf.int32, name='off_policy_sequence_size')
 
         # ============= Base on-policy AAC network ===========
 
@@ -133,7 +133,7 @@ def __init__(self,
         self.on_vae_d_kl_ext = on_d_kl_ext
 
         # Reshape rnn inputs for  batch training as [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(on_aac_x_ext)
+        x_shape_dynamic = tf.shape(input=on_aac_x_ext)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.on_batch_size, tf.int32)
         x_shape_static = on_aac_x_ext.get_shape().as_list()
 
@@ -158,7 +158,7 @@ def __init__(self,
                 x_int_shape_static = on_x_int.get_shape().as_list()
                 on_x_int = [
                     tf.reshape(on_x_int, [self.on_batch_size, max_seq_len, np.prod(x_int_shape_static[1:])])]
-                self.debug['state_internal_enc'] = tf.shape(on_x_int)
+                self.debug['state_internal_enc'] = tf.shape(input=on_x_int)
 
             else:
                 # Feed as is:
@@ -167,7 +167,7 @@ def __init__(self,
                     self.on_state_in['internal'],
                     [self.on_batch_size, max_seq_len, np.prod(x_int_shape_static[1:])]
                 )
-                self.debug['state_internal'] = tf.shape(self.on_state_in['internal'])
+                self.debug['state_internal'] = tf.shape(input=self.on_state_in['internal'])
                 on_x_int = [on_x_int]
                 self.on_state_decoded_int = None
                 self.on_vae_d_kl_int = None
@@ -177,7 +177,7 @@ def __init__(self,
             self.on_state_decoded_int = None
             self.on_vae_d_kl_int = None
 
-        self.debug['conv_input_to_lstm1'] = tf.shape(on_aac_x_ext)
+        self.debug['conv_input_to_lstm1'] = tf.shape(input=on_aac_x_ext)
 
         # Feed last last_reward into LSTM_1 layer along with encoded `external` state features:
         on_stage2_1_input = [on_aac_x_ext, on_a_r_in[..., -1][..., None]] #+ on_x_internal
@@ -188,18 +188,18 @@ def __init__(self,
         # LSTM_1 full input:
         on_aac_x_ext = tf.concat(on_stage2_1_input, axis=-1)
 
-        self.debug['concat_input_to_lstm1'] = tf.shape(on_aac_x_ext)
+        self.debug['concat_input_to_lstm1'] = tf.shape(input=on_aac_x_ext)
 
         # First LSTM layer takes encoded `external` state:
         [on_x_lstm_1_out, self.on_lstm_1_init_state, self.on_lstm_1_state_out, self.on_lstm_1_state_pl_flatten] =\
             lstm_network(on_aac_x_ext, self.on_time_length, lstm_class_ref, (lstm_layers[0],), name='lstm_1')
 
-        self.debug['on_x_lstm_1_out'] = tf.shape(on_x_lstm_1_out)
-        self.debug['self.on_lstm_1_state_out'] = tf.shape(self.on_lstm_1_state_out)
-        self.debug['self.on_lstm_1_state_pl_flatten'] = tf.shape(self.on_lstm_1_state_pl_flatten)
+        self.debug['on_x_lstm_1_out'] = tf.shape(input=on_x_lstm_1_out)
+        self.debug['self.on_lstm_1_state_out'] = tf.shape(input=self.on_lstm_1_state_out)
+        self.debug['self.on_lstm_1_state_pl_flatten'] = tf.shape(input=self.on_lstm_1_state_pl_flatten)
 
         # For time_flat only: Reshape on_lstm_1_state_out from [1,2,20,size] -->[20,1,2,size] --> [20,1, 2xsize]:
-        reshape_lstm_1_state_out = tf.transpose(self.on_lstm_1_state_out, [2, 0, 1, 3])
+        reshape_lstm_1_state_out = tf.transpose(a=self.on_lstm_1_state_out, perm=[2, 0, 1, 3])
         reshape_lstm_1_state_out_shape_static = reshape_lstm_1_state_out.get_shape().as_list()
         reshape_lstm_1_state_out = tf.reshape(
             reshape_lstm_1_state_out,
@@ -212,7 +212,7 @@ def __init__(self,
         x_shape_static = on_x_lstm_1_out.get_shape().as_list()
         rsh_on_x_lstm_1_out = tf.reshape(on_x_lstm_1_out, [x_shape_dynamic[0], x_shape_static[-1]])
 
-        self.debug['reshaped_on_x_lstm_1_out'] = tf.shape(rsh_on_x_lstm_1_out)
+        self.debug['reshaped_on_x_lstm_1_out'] = tf.shape(input=rsh_on_x_lstm_1_out)
 
         # Aac policy output and action-sampling function:
         [self.on_logits, _, self.on_sample] = dense_aac_network(
@@ -231,20 +231,20 @@ def __init__(self,
         # LSTM_2 full input:
         on_aac_x_ext = tf.concat(on_stage2_2_input, axis=-1)
 
-        self.debug['on_stage2_2_input'] = tf.shape(on_aac_x_ext)
+        self.debug['on_stage2_2_input'] = tf.shape(input=on_aac_x_ext)
 
         [on_x_lstm_2_out, self.on_lstm_2_init_state, self.on_lstm_2_state_out, self.on_lstm_2_state_pl_flatten] = \
             lstm_network(on_aac_x_ext, self.on_time_length, lstm_class_ref, (lstm_layers[-1],), name='lstm_2')
 
-        self.debug['on_x_lstm_2_out'] = tf.shape(on_x_lstm_2_out)
-        self.debug['self.on_lstm_2_state_out'] = tf.shape(self.on_lstm_2_state_out)
-        self.debug['self.on_lstm_2_state_pl_flatten'] = tf.shape(self.on_lstm_2_state_pl_flatten)
+        self.debug['on_x_lstm_2_out'] = tf.shape(input=on_x_lstm_2_out)
+        self.debug['self.on_lstm_2_state_out'] = tf.shape(input=self.on_lstm_2_state_out)
+        self.debug['self.on_lstm_2_state_pl_flatten'] = tf.shape(input=self.on_lstm_2_state_pl_flatten)
 
         # Reshape back to [batch, flattened_depth], where batch = rnn_batch_dim * rnn_time_dim:
         x_shape_static = on_x_lstm_2_out.get_shape().as_list()
         on_x_lstm_out = tf.reshape(on_x_lstm_2_out, [x_shape_dynamic[0], x_shape_static[-1]])
 
-        self.debug['reshaped_on_x_lstm_out'] = tf.shape(on_x_lstm_out)
+        self.debug['reshaped_on_x_lstm_out'] = tf.shape(input=on_x_lstm_out)
 
         # Aac value function:
         [_, self.on_vf, _] = dense_aac_network(
@@ -274,7 +274,7 @@ def __init__(self,
         self.off_vae_d_kl_ext = off_d_kl_ext
 
         # Reshape rnn inputs for  batch training as [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(off_aac_x)
+        x_shape_dynamic = tf.shape(input=off_aac_x)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.off_batch_size, tf.int32)
         x_shape_static = off_aac_x.get_shape().as_list()
 
@@ -397,7 +397,7 @@ def __init__(self,
 
         # Aux3:
         # `Reward prediction` network.
-        self.rp_batch_size = tf.placeholder(tf.int32, name='rp_batch_size')
+        self.rp_batch_size = tf.compat.v1.placeholder(tf.int32, name='rp_batch_size')
 
         # Shared conv. output:
         rp_encoded_layers, rp_x, rp_decoded_layers, _, rp_d_kl = encoder_class_ref(
@@ -421,18 +421,18 @@ def __init__(self,
                 pass
 
         except AttributeError:
-            self.train_phase = tf.placeholder_with_default(
+            self.train_phase = tf.compat.v1.placeholder_with_default(
                 tf.constant(False, dtype=tf.bool),
                 shape=(),
                 name='train_phase_flag_pl'
             )
-        self.update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        self.update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
         # Add moving averages to save list:
-        moving_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*moving.*')
-        renorm_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*renorm.*')
+        moving_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*moving.*')
+        renorm_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*renorm.*')
 
         # What to save:
-        self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, tf.get_variable_scope().name)
+        self.var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, tf.compat.v1.get_variable_scope().name)
         self.var_list += moving_var_list + renorm_var_list
 
         # Callbacks:
@@ -465,7 +465,7 @@ def get_initial_features(self, state, context=None):
             KeyError if [`metadata`]:[`trial_num`,`type`] keys not found
         """
         try:
-            sess = tf.get_default_session()
+            sess = tf.compat.v1.get_default_session()
             new_context = list(sess.run(self.on_lstm_init_state))
             if state['metadata']['trial_num'] != self.current_trial_num\
                     or context is None\
@@ -503,7 +503,7 @@ def __init__(self, ae_loss=beta_vae_loss_def, ae_alpha=1.0, ae_beta=1.0, _log_na
         try:
             super(bVAENA3C, self).__init__(name=_log_name, **kwargs)
             with tf.device(self.worker_device):
-                with tf.variable_scope('local'):
+                with tf.compat.v1.variable_scope('local'):
                     on_vae_loss_ext, on_ae_summary_ext = ae_loss(
                         targets=self.local_network.on_state_in['external'],
                         logits=self.local_network.on_state_decoded_ext,
@@ -531,7 +531,7 @@ def __init__(self, ae_loss=beta_vae_loss_def, ae_alpha=1.0, ae_beta=1.0, _log_na
 
                     # Override train op def:
                     self.grads, _ = tf.clip_by_global_norm(
-                        tf.gradients(self.loss, self.local_network.var_list),
+                        tf.gradients(ys=self.loss, xs=self.local_network.var_list),
                         40.0
                     )
                     grads_and_vars = list(zip(self.grads, self.network.var_list))
@@ -539,7 +539,7 @@ def __init__(self, ae_loss=beta_vae_loss_def, ae_alpha=1.0, ae_beta=1.0, _log_na
 
                     # Merge summary:
                     extended_summary.append(self.model_summary_op)
-                    self.model_summary_op = tf.summary.merge(extended_summary, name='extended_summary')
+                    self.model_summary_op = tf.compat.v1.summary.merge(extended_summary, name='extended_summary')
 
         except:
             msg = 'Child 0.0 class __init()__ exception occurred' + \
diff --git a/btgym/research/casual/aac.py b/btgym/research/casual/aac.py
index 7f27de5a..e389e959 100644
--- a/btgym/research/casual/aac.py
+++ b/btgym/research/casual/aac.py
@@ -279,7 +279,7 @@ def _make_loss(self, **kwargs):
 
         # Guidance annealing:
         if self.guided_decay_steps is not None:
-            self.guided_lambda_decayed = tf.train.polynomial_decay(
+            self.guided_lambda_decayed = tf.compat.v1.train.polynomial_decay(
                 self.guided_lambda,
                 self.global_step + 1,
                 self.guided_decay_steps,
@@ -327,14 +327,14 @@ def _make_train_op(self, pi, pi_prime, pi_global):
         """
 
         # Each worker gets a different set of adam optimizer parameters:
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
 
         # Clipped gradients:
         self.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.loss, pi.var_list),
+            tf.gradients(ys=self.loss, xs=pi.var_list),
             40.0
         )
-        self.grads_global_norm = tf.global_norm(self.grads)
+        self.grads_global_norm = tf.linalg.global_norm(self.grads)
         # Copy weights from the parameter server to the local model
         self.sync = self.sync_pi = tf.group(
             *[v1.assign(v2) for v1, v2 in zip(pi.var_list, pi_global.var_list)]
@@ -351,21 +351,21 @@ def _make_train_op(self, pi, pi_prime, pi_global):
 
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
-        self.inc_step = self.global_step.assign_add(tf.shape(pi.on_state_in['external'])[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=pi.on_state_in['external'])[0])
 
         self.local_network.meta.grads_and_vars = list(
             zip(self.local_network.meta.grads, self.network.meta.var_list)
         )
-        self.meta_opt = tf.train.GradientDescentOptimizer(self.local_network.meta.learn_rate)
+        self.meta_opt = tf.compat.v1.train.GradientDescentOptimizer(self.local_network.meta.learn_rate)
 
         self.meta_train_op = self.meta_opt.apply_gradients(self.local_network.meta.grads_and_vars)
 
-        self.local_network.meta.sync_slot_op = tf.assign(
+        self.local_network.meta.sync_slot_op = tf.compat.v1.assign(
             self.local_network.meta.cluster_averages_slot,
             self.network.meta.cluster_averages_slot,
         )
 
-        self.local_network.meta.send_stat_op = tf.scatter_nd_update(
+        self.local_network.meta.send_stat_op = tf.compat.v1.scatter_nd_update(
             self.network.meta.cluster_averages_slot,
             [[0, self.task], [1, self.task]],
             [
diff --git a/btgym/research/casual_conv/layers.py b/btgym/research/casual_conv/layers.py
index 8fa01ca4..d47e75a8 100644
--- a/btgym/research/casual_conv/layers.py
+++ b/btgym/research/casual_conv/layers.py
@@ -4,22 +4,22 @@
 
 
 def time_to_batch(value, dilation, name=None):
-    with tf.name_scope('time_to_batch'):
-        shape = tf.shape(value)
+    with tf.compat.v1.name_scope('time_to_batch'):
+        shape = tf.shape(input=value)
         pad_elements = dilation - 1 - (shape[1] + dilation - 1) % dilation
-        padded = tf.pad(value, [[0, 0], [0, pad_elements], [0, 0]])
+        padded = tf.pad(tensor=value, paddings=[[0, 0], [0, pad_elements], [0, 0]])
         reshaped = tf.reshape(padded, [-1, dilation, shape[2]])
-        transposed = tf.transpose(reshaped, perm=[1, 0, 2])
+        transposed = tf.transpose(a=reshaped, perm=[1, 0, 2])
         return tf.reshape(transposed, [shape[0] * dilation, -1, shape[2]])
 
 
 def batch_to_time(value, dilation, name=None):
-    with tf.name_scope('batch_to_time'):
-        shape = tf.shape(value)
+    with tf.compat.v1.name_scope('batch_to_time'):
+        shape = tf.shape(input=value)
         prepared = tf.reshape(value, [dilation, -1, shape[2]])
-        transposed = tf.transpose(prepared, perm=[1, 0, 2])
+        transposed = tf.transpose(a=prepared, perm=[1, 0, 2])
         return tf.reshape(transposed,
-                          [tf.div(shape[0], dilation), -1, shape[2]])
+                          [tf.compat.v1.div(shape[0], dilation), -1, shape[2]])
 
 
 def dilated_conv1d(
@@ -32,7 +32,7 @@ def dilated_conv1d(
         name='dialted_conv_1d',
         reuse=False
 ):
-    with tf.name_scope(name):
+    with tf.compat.v1.name_scope(name):
         if dilation_rate > 1:
             transformed = time_to_batch(inputs, dilation_rate)
             conv = conv1d(
@@ -56,7 +56,7 @@ def dilated_conv1d(
                 reuse=reuse
             )
         # Remove excess elements at the end.
-        out_width = tf.shape(inputs)[1] - (filter_width - 1) * dilation_rate
+        out_width = tf.shape(input=inputs)[1] - (filter_width - 1) * dilation_rate
         result = tf.slice(restored,
                           [0, 0, 0],
                           [-1, out_width, -1])
diff --git a/btgym/research/casual_conv/networks.py b/btgym/research/casual_conv/networks.py
index cef1c333..ac48c641 100644
--- a/btgym/research/casual_conv/networks.py
+++ b/btgym/research/casual_conv/networks.py
@@ -32,7 +32,7 @@ def conv_1d_casual_encoder(
         tensor holding state features;
     """
 
-    with tf.variable_scope(name_or_scope=name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name_or_scope=name, reuse=reuse):
         shape = x.get_shape().as_list()
         if len(shape) > 3:  # remove pseudo 2d dimension
             x = x[:, :, 0, :]
@@ -53,7 +53,7 @@ def conv_1d_casual_encoder(
             if tail != 0:
                 pad = conv_1d_filter_size - tail
                 paddings = [[0, 0], [pad, 0], [0, 0]]
-                y = tf.pad(y, paddings)
+                y = tf.pad(tensor=y, paddings=paddings)
                 length += pad
 
             # print('padded_length: ', length)
@@ -83,7 +83,7 @@ def conv_1d_casual_encoder(
                 y = conv_1d_activation(y)
 
             if keep_prob is not None:
-                y = tf.nn.dropout(y, keep_prob=keep_prob, name="_layer_{}_with_dropout".format(i))
+                y = tf.nn.dropout(y, rate=1 - (keep_prob), name="_layer_{}_with_dropout".format(i))
 
             layers.append(y)
 
@@ -171,7 +171,7 @@ def attention_layer(inputs, attention_ref=tf.contrib.seq2seq.LuongAttention, nam
     # Suppose there is no previous context for attention (uhm?):
     alignments = attention_mechanism(
         query_state,
-        attention_mechanism.initial_alignments(tf.shape(inputs)[0], dtype=tf.float32)
+        attention_mechanism.initial_alignments(tf.shape(input=inputs)[0], dtype=tf.float32)
     )
     # Somehow attention call returns tuple of twin tensors (wtf?):
     if isinstance(alignments, tuple):
@@ -219,7 +219,7 @@ def conv_1d_casual_attention_encoder(
         tensor holding state features;
     """
 
-    with tf.variable_scope(name_or_scope=name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name_or_scope=name, reuse=reuse):
         shape = x.get_shape().as_list()
         if len(shape) > 3:  # remove pseudo 2d dimension
             x = x[:, :, 0, :]
@@ -240,7 +240,7 @@ def conv_1d_casual_attention_encoder(
             if tail != 0:
                 pad = conv_1d_filter_size - tail
                 paddings = [[0, 0], [pad, 0], [0, 0]]
-                y = tf.pad(y, paddings)
+                y = tf.pad(tensor=y, paddings=paddings)
                 length += pad
 
             # print('padded_length: ', length)
@@ -268,7 +268,7 @@ def conv_1d_casual_attention_encoder(
                 y = conv_1d_activation(y)
 
             if keep_prob is not None:
-                y = tf.nn.dropout(y, keep_prob=keep_prob, name="_layer_{}_with_dropout".format(i))
+                y = tf.nn.dropout(y, rate=1 - (keep_prob), name="_layer_{}_with_dropout".format(i))
 
             if conv_1d_gated:
                 split_size = int(conv_1d_num_filters / 2)
diff --git a/btgym/research/encoder_test/aac.py b/btgym/research/encoder_test/aac.py
index d27d82dc..a9997d3e 100644
--- a/btgym/research/encoder_test/aac.py
+++ b/btgym/research/encoder_test/aac.py
@@ -553,13 +553,13 @@ def _make_loss(self, pi, pi_prime, name='base', verbose=True, **kwargs):
             tensor holding estimated loss graph
             list of related summaries
         """
-        with tf.name_scope(name):
+        with tf.compat.v1.name_scope(name):
             # On-policy AAC loss definition:
-            pi.on_pi_act_target = tf.placeholder(
+            pi.on_pi_act_target = tf.compat.v1.placeholder(
                 tf.float32, [None, self.ref_env.action_space.one_hot_depth], name="on_policy_action_pl"
             )
-            pi.on_pi_adv_target = tf.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
-            pi.on_pi_r_target = tf.placeholder(tf.float32, [None], name="on_policy_return_pl")
+            pi.on_pi_adv_target = tf.compat.v1.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
+            pi.on_pi_r_target = tf.compat.v1.placeholder(tf.float32, [None], name="on_policy_return_pl")
 
             # clip_epsilon = tf.cast(self.clip_epsilon * self.learn_rate_decayed / self.opt_learn_rate, tf.float32)
             #
@@ -576,20 +576,20 @@ def _make_loss(self, pi, pi_prime, name='base', verbose=True, **kwargs):
             #     verbose=verbose
             # )
             pi_regression = tf.exp(pi.regression)
-            regress_loss = tf.losses.mean_squared_error(
+            regress_loss = tf.compat.v1.losses.mean_squared_error(
                 labels=pi.regression_targets,
                 predictions=pi_regression,
                 weights=self.regress_lambda,
             )
 
-            self.mse = tf.metrics.mean_squared_error(
+            self.mse = tf.compat.v1.metrics.mean_squared_error(
                 labels=pi.regression_targets,
                 predictions=pi_regression
             )
 
             model_summaries = [
-                tf.summary.scalar('regress_loss', regress_loss),
-                tf.summary.scalar('mse_metric', self.mse[0])
+                tf.compat.v1.summary.scalar('regress_loss', regress_loss),
+                tf.compat.v1.summary.scalar('mse_metric', self.mse[0])
             ]
             # Accumulate total loss:
             # loss = float(self.class_lambda) * regress_loss + float(self.aac_lambda) * on_pi_loss\
@@ -615,14 +615,14 @@ def _make_train_op(self, pi, pi_prime, pi_global):
         """
 
         # Each worker gets a different set of adam optimizer parameters:
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
 
         # Clipped gradients:
         self.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.loss, pi.var_list),
+            tf.gradients(ys=self.loss, xs=pi.var_list),
             40.0
         )
-        self.grads_global_norm = tf.global_norm(self.grads)
+        self.grads_global_norm = tf.linalg.global_norm(self.grads)
 
         # Copy weights from the parameter server to the local model:
         self.sync = self.sync_pi = tf.group(
@@ -647,7 +647,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
             stream = pi.on_state_in['external'][list(pi.on_state_in['external'].keys())[0]]
         else:
             stream = pi.on_state_in['external']
-        self.inc_step = self.global_step.assign_add(tf.shape(stream)[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=stream)[0])
 
         train_op = [self.optimizer.apply_gradients(grads_and_vars), self.mse]
 
@@ -665,64 +665,64 @@ def _combine_summaries(self, policy=None, model_summaries=None):
         if model_summaries is not None:
             if self.use_global_network:
                 # Model-wide statistics:
-                with tf.name_scope('model'):
+                with tf.compat.v1.name_scope('model'):
                     model_summaries += [
-                        tf.summary.scalar("grad_global_norm", self.grads_global_norm),
-                        tf.summary.scalar("learn_rate", self.learn_rate_decayed),
+                        tf.compat.v1.summary.scalar("grad_global_norm", self.grads_global_norm),
+                        tf.compat.v1.summary.scalar("learn_rate", self.learn_rate_decayed),
                         # cause actual rate is a jaggy due to test freezes
-                        tf.summary.scalar("total_loss", self.loss),
+                        tf.compat.v1.summary.scalar("total_loss", self.loss),
                     ]
                     if policy is not None:
-                        model_summaries += [tf.summary.scalar("var_global_norm", tf.global_norm(policy.var_list))]
+                        model_summaries += [tf.compat.v1.summary.scalar("var_global_norm", tf.linalg.global_norm(policy.var_list))]
         else:
             model_summaries = []
         # Model stat. summary:
-        model_summary = tf.summary.merge(model_summaries, name='model_summary')
+        model_summary = tf.compat.v1.summary.merge(model_summaries, name='model_summary')
 
         # Episode-related summaries:
         ep_summary = dict(
             # Summary placeholders
-            render_atari=tf.placeholder(tf.uint8, [None, None, None, 1]),
-            total_r=tf.placeholder(tf.float32, ),
-            cpu_time=tf.placeholder(tf.float32, ),
-            final_value=tf.placeholder(tf.float32, ),
-            steps=tf.placeholder(tf.int32, ),
+            render_atari=tf.compat.v1.placeholder(tf.uint8, [None, None, None, 1]),
+            total_r=tf.compat.v1.placeholder(tf.float32, ),
+            cpu_time=tf.compat.v1.placeholder(tf.float32, ),
+            final_value=tf.compat.v1.placeholder(tf.float32, ),
+            steps=tf.compat.v1.placeholder(tf.int32, ),
         )
         if self.test_mode:
             # For Atari:
-            ep_summary['render_op'] = tf.summary.image("model/state", ep_summary['render_atari'])
+            ep_summary['render_op'] = tf.compat.v1.summary.image("model/state", ep_summary['render_atari'])
 
         else:
             # BTGym rendering:
             ep_summary.update(
                 {
-                    mode: tf.placeholder(tf.uint8, [None, None, None, None], name=mode + '_pl')
+                    mode: tf.compat.v1.placeholder(tf.uint8, [None, None, None, None], name=mode + '_pl')
                     for mode in self.env_list[0].render_modes + self.aux_render_modes
                 }
             )
-            ep_summary['render_op'] = tf.summary.merge(
-                [tf.summary.image(mode, ep_summary[mode])
+            ep_summary['render_op'] = tf.compat.v1.summary.merge(
+                [tf.compat.v1.summary.image(mode, ep_summary[mode])
                  for mode in self.env_list[0].render_modes + self.aux_render_modes]
             )
         # Episode stat. summary:
-        ep_summary['btgym_stat_op'] = tf.summary.merge(
+        ep_summary['btgym_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode_train/cpu_time_sec', ep_summary['cpu_time']),
-                tf.summary.scalar('episode_train/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode_train/cpu_time_sec', ep_summary['cpu_time']),
+                tf.compat.v1.summary.scalar('episode_train/total_reward', ep_summary['total_r']),
             ],
             name='episode_train_btgym'
         )
         # Test episode stat. summary:
-        ep_summary['test_btgym_stat_op'] = tf.summary.merge(
+        ep_summary['test_btgym_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode_test/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode_test/total_reward', ep_summary['total_r']),
             ],
             name='episode_test_btgym'
         )
-        ep_summary['atari_stat_op'] = tf.summary.merge(
+        ep_summary['atari_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode/steps', ep_summary['steps'])
+                tf.compat.v1.summary.scalar('episode/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode/steps', ep_summary['steps'])
             ],
             name='episode_atari'
         )
diff --git a/btgym/research/encoder_test/networks.py b/btgym/research/encoder_test/networks.py
index b115cc82..24852510 100644
--- a/btgym/research/encoder_test/networks.py
+++ b/btgym/research/encoder_test/networks.py
@@ -33,7 +33,7 @@ def conv_2d_network_skip(x,
     """
     assert conv_2d_num_filters[-1] % 2 == 0
     layers = []
-    with tf.variable_scope(name, reuse=reuse):
+    with tf.compat.v1.variable_scope(name, reuse=reuse):
         for i, num_filters in enumerate(conv_2d_num_filters):
             x = tf.nn.elu(
                 norm_layer(
@@ -52,7 +52,7 @@ def conv_2d_network_skip(x,
                 )
             )
             if keep_prob is not None:
-                x = tf.nn.dropout(x, keep_prob=keep_prob, name="_layer_{}_with_dropout".format(i + 1))
+                x = tf.nn.dropout(x, rate=1 - (keep_prob), name="_layer_{}_with_dropout".format(i + 1))
 
             layers.append(x)
 
@@ -72,7 +72,7 @@ def conv_2d_network_skip(x,
             # print('{}.shape = {}'.format(x.name, x.get_shape().as_list()))
 
         if conv_2d_enable_skip:
-            x = tf.concat([tf.layers.flatten(l) for l in layers], axis=-1, name='flattened_encoded_state')
+            x = tf.concat([tf.compat.v1.layers.flatten(l) for l in layers], axis=-1, name='flattened_encoded_state')
 
         # print('{}.shape = {}'.format(x.name, x.get_shape().as_list()))
         return x
@@ -85,7 +85,7 @@ def identity_encoder(x, name='identity_encoder', **kwargs):
     Returns:
         tensor holding state features;
     """
-    with tf.variable_scope(name,):
-        x = tf.layers.flatten(x)
+    with tf.compat.v1.variable_scope(name,):
+        x = tf.compat.v1.layers.flatten(x)
 
         return x
\ No newline at end of file
diff --git a/btgym/research/encoder_test/policy.py b/btgym/research/encoder_test/policy.py
index 9383595b..e5fb05d0 100644
--- a/btgym/research/encoder_test/policy.py
+++ b/btgym/research/encoder_test/policy.py
@@ -72,7 +72,7 @@ def __init__(self,
         self.encode_internal_state = encode_internal_state
         self.share_encoder_params = share_encoder_params
         if self.share_encoder_params:
-            self.reuse_encoder_params = tf.AUTO_REUSE
+            self.reuse_encoder_params = tf.compat.v1.AUTO_REUSE
 
         else:
             self.reuse_encoder_params = False
@@ -88,16 +88,16 @@ def __init__(self,
         self.on_state_in = nested_placeholders(self.ob_space.shape, batch_dim=None, name='on_policy_state_in')
 
         # Placeholders for previous step action[multi-categorical vector encoding]  and reward [scalar]:
-        self.on_last_a_in = tf.placeholder(
+        self.on_last_a_in = tf.compat.v1.placeholder(
             tf.float32,
             [None, self.ac_space.encoded_depth],
             name='on_policy_last_action_in_pl'
         )
-        self.on_last_reward_in = tf.placeholder(tf.float32, [None], name='on_policy_last_reward_in_pl')
+        self.on_last_reward_in = tf.compat.v1.placeholder(tf.float32, [None], name='on_policy_last_reward_in_pl')
 
         # Placeholders for rnn batch and time-step dimensions:
-        self.on_batch_size = tf.placeholder(tf.int32, name='on_policy_batch_size')
-        self.on_time_length = tf.placeholder(tf.int32, name='on_policy_sequence_size')
+        self.on_batch_size = tf.compat.v1.placeholder(tf.int32, name='on_policy_batch_size')
+        self.on_time_length = tf.compat.v1.placeholder(tf.int32, name='on_policy_sequence_size')
 
         self.debug['on_state_in_keys'] = list(self.on_state_in.keys())
 
@@ -111,7 +111,7 @@ def __init__(self,
                 pass
 
         except AttributeError:
-            self.train_phase = tf.placeholder_with_default(
+            self.train_phase = tf.compat.v1.placeholder_with_default(
                 tf.constant(False, dtype=tf.bool),
                 shape=(),
                 name='train_phase_flag_pl'
@@ -146,7 +146,7 @@ def __init__(self,
                     else:
                         layer_name_template = 'encoded_{}_{}'
                     encoded_streams = {
-                        name: tf.layers.flatten(
+                        name: tf.compat.v1.layers.flatten(
                             self.state_encoder_class_ref(
                                 x=stream,
                                 ob_space=self.ob_space.shape[key][name],
@@ -165,7 +165,7 @@ def __init__(self,
                     )
                 else:
                     # Got single data stream:
-                    encoded_mode = tf.layers.flatten(
+                    encoded_mode = tf.compat.v1.layers.flatten(
                         self.state_encoder_class_ref(
                             x=self.on_state_in[key],
                             ob_space=self.ob_space.shape[key],
@@ -184,14 +184,14 @@ def __init__(self,
 
         # TODO: for encoder prediction test, output `naive` estimates for logits and value directly from encoder:
         [self.on_simple_logits, self.on_simple_value, _] = dense_aac_network(
-            tf.layers.flatten(on_aac_x),
+            tf.compat.v1.layers.flatten(on_aac_x),
             ac_space_depth=self.ac_space.one_hot_depth,
             linear_layer_ref=linear_layer_ref,
             name='aac_dense_simple_pi_v'
         )
 
         # Reshape rnn inputs for batch training as: [rnn_batch_dim, rnn_time_dim, flattened_depth]:
-        x_shape_dynamic = tf.shape(on_aac_x)
+        x_shape_dynamic = tf.shape(input=on_aac_x)
         max_seq_len = tf.cast(x_shape_dynamic[0] / self.on_batch_size, tf.int32)
         x_shape_static = on_aac_x.get_shape().as_list()
 
@@ -280,7 +280,7 @@ def __init__(self,
         self.debug['self.on_lstm_1_state_pl_flatten'] = self.on_lstm_1_state_pl_flatten
 
         # For time_flat only: Reshape on_lstm_1_state_out from [1,2,20,size] -->[20,1,2,size] --> [20,1, 2xsize]:
-        reshape_lstm_1_state_out = tf.transpose(self.on_lstm_1_state_out, [2, 0, 1, 3])
+        reshape_lstm_1_state_out = tf.transpose(a=self.on_lstm_1_state_out, perm=[2, 0, 1, 3])
         reshape_lstm_1_state_out_shape_static = reshape_lstm_1_state_out.get_shape().as_list()
 
         # Take policy logits off first LSTM-dense layer:
@@ -363,7 +363,7 @@ def __init__(self,
             # )
             #self.regression = tf.layers.flatten(self.debug['on_state_external_encoded'])
             self.regression = linear(
-                x=tf.layers.flatten(self.debug['on_state_external_encoded']),
+                x=tf.compat.v1.layers.flatten(self.debug['on_state_external_encoded']),
                 size=self.regression_targets.shape.as_list()[-1],
                 initializer=normalized_columns_initializer(0.1),
                 name='on_dense_simple_regression',
@@ -379,7 +379,7 @@ def __init__(self,
             #     name='on_dense_rnn_regression'
             # )
             self.regression = linear(
-                x=tf.layers.flatten(self.debug['reshaped_on_x_lstm_2_out']),
+                x=tf.compat.v1.layers.flatten(self.debug['reshaped_on_x_lstm_2_out']),
                 size=self.regression_targets.shape.as_list()[-1],
                 initializer=normalized_columns_initializer(0.1),
                 name='on_dense_rnn_regression',
@@ -394,13 +394,13 @@ def __init__(self,
         self.on_lstm_state_pl_flatten = self.on_lstm_1_state_pl_flatten + self.on_lstm_2_state_pl_flatten
 
         # Batch-norm related:
-        self.update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        self.update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
         # Add moving averages to save list:
-        moving_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*moving.*')
-        renorm_var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, tf.get_variable_scope().name + '.*renorm.*')
+        moving_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*moving.*')
+        renorm_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, tf.compat.v1.get_variable_scope().name + '.*renorm.*')
 
         # What to save:
-        self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, tf.get_variable_scope().name)
+        self.var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, tf.compat.v1.get_variable_scope().name)
         self.var_list += moving_var_list + renorm_var_list
 
         # RL2 related:
@@ -434,7 +434,7 @@ def get_initial_features(self, state, context=None):
             KeyError if [`metadata`]:[`trial_num`,`type`] keys not found
         """
         try:
-            sess = tf.get_default_session()
+            sess = tf.compat.v1.get_default_session()
             new_context = list(sess.run(self.on_lstm_init_state))
             if state['metadata']['trial_num'] != self.current_trial_num\
                     or context is None\
@@ -473,7 +473,7 @@ def act(self, observation, lstm_state, last_action, last_reward):
         Returns:
             Action as dictionary of several action encodings, actions logits, V-fn value, output RNN state
         """
-        sess = tf.get_default_session()
+        sess = tf.compat.v1.get_default_session()
         feeder = {pl: value for pl, value in zip(self.on_lstm_state_pl_flatten, flatten_nested(lstm_state))}
         feeder.update(feed_dict_from_nested(self.on_state_in, observation, expand_batch=True))
         feeder.update(
diff --git a/btgym/research/gps/aac.py b/btgym/research/gps/aac.py
index 73cac65b..68c6ae54 100644
--- a/btgym/research/gps/aac.py
+++ b/btgym/research/gps/aac.py
@@ -93,7 +93,7 @@ def _make_loss(self, **kwargs):
 
         # Guidance annealing:
         if self.guided_decay_steps is not None:
-            self.guided_lambda_decayed = tf.train.polynomial_decay(
+            self.guided_lambda_decayed = tf.compat.v1.train.polynomial_decay(
                 self.guided_lambda,
                 self.global_step + 1,
                 self.guided_decay_steps,
diff --git a/btgym/research/gps/loss.py b/btgym/research/gps/loss.py
index cf221afa..056c081a 100644
--- a/btgym/research/gps/loss.py
+++ b/btgym/research/gps/loss.py
@@ -14,17 +14,17 @@ def guided_aac_loss_def_0_0(pi_actions, expert_actions, name='on_policy/aac', ve
         tensor holding estimated imitation loss;
         list of related tensorboard summaries.
     """
-    with tf.name_scope(name + '/guided_loss'):
+    with tf.compat.v1.name_scope(name + '/guided_loss'):
         # Loss over expert action's distribution:
 
-        neg_pi_log_prob = tf.nn.softmax_cross_entropy_with_logits_v2(
+        neg_pi_log_prob = tf.nn.softmax_cross_entropy_with_logits(
             logits=pi_actions,
-            labels=tf.argmax(expert_actions, axis=-1)
+            labels=tf.argmax(input=expert_actions, axis=-1)
         )
-        loss = tf.reduce_mean(neg_pi_log_prob)
+        loss = tf.reduce_mean(input_tensor=neg_pi_log_prob)
 
         if verbose:
-            summaries = [tf.summary.scalar('actions_ce', loss)]
+            summaries = [tf.compat.v1.summary.scalar('actions_ce', loss)]
         else:
             summaries = []
 
@@ -44,18 +44,18 @@ def guided_aac_loss_def_0_1(pi_actions, expert_actions, name='on_policy/aac', ve
         tensor holding estimated imitation loss;
         list of related tensorboard summaries.
     """
-    with tf.name_scope(name + '/guided_loss'):
+    with tf.compat.v1.name_scope(name + '/guided_loss'):
         # Loss over expert buy/ sell:
         # Cross-entropy on subset?...
 
-        neg_pi_log_prob = tf.nn.softmax_cross_entropy_with_logits_v2(
+        neg_pi_log_prob = tf.nn.softmax_cross_entropy_with_logits(
             logits=pi_actions[..., 1:3],
             labels=expert_actions[..., 1:3]
         )
-        loss = tf.reduce_mean(neg_pi_log_prob)
+        loss = tf.reduce_mean(input_tensor=neg_pi_log_prob)
 
         if verbose:
-            summaries = [tf.summary.scalar('actions_ce', loss)]
+            summaries = [tf.compat.v1.summary.scalar('actions_ce', loss)]
         else:
             summaries = []
 
@@ -75,20 +75,20 @@ def guided_aac_loss_def_0_3(pi_actions, expert_actions, name='on_policy/aac', ve
         tensor holding estimated imitation loss;
         list of related tensorboard summaries.
     """
-    with tf.name_scope(name + '/guided_loss'):
+    with tf.compat.v1.name_scope(name + '/guided_loss'):
         if 'guided_lambda' in kwargs.keys():
             guided_lambda = kwargs['guided_lambda']
         else:
             guided_lambda = 1.0
 
         # Loss over expert buy/ sell:
-        loss = tf.losses.mean_squared_error(
+        loss = tf.compat.v1.losses.mean_squared_error(
             labels=expert_actions[..., 1:3],
             predictions=tf.nn.softmax(pi_actions)[..., 1:3],
         ) * guided_lambda
 
         if verbose:
-            summaries = [tf.summary.scalar('actions_mse', loss)]
+            summaries = [tf.compat.v1.summary.scalar('actions_mse', loss)]
         else:
             summaries = []
 
diff --git a/btgym/research/metalearn_2/_fwrnn_aac.py b/btgym/research/metalearn_2/_fwrnn_aac.py
index 52cf1a54..e69de29b 100644
--- a/btgym/research/metalearn_2/_fwrnn_aac.py
+++ b/btgym/research/metalearn_2/_fwrnn_aac.py
@@ -1,188 +0,0 @@
-import tensorflow as tf
-import time
-
-import sys
-from logbook import Logger, StreamHandler
-
-from btgym.research.gps.aac import GuidedAAC
-from btgym.algorithms.runner.synchro import BaseSynchroRunner
-
-
-class MetaAAC_2_0(GuidedAAC):
-    """
-    RNN adaptation experiment
-    """
-
-    def __init__(
-            self,
-            trial_source_target_cycle=(1, 0),
-            num_episodes_per_trial=1,
-            name='AAC_FWRNN_Ada',
-            **kwargs
-    ):
-        runner_config = {
-            'class_ref': BaseSynchroRunner,
-            'kwargs': {
-                'data_sample_config': {'mode': 0},
-                'name': '',
-            },
-        }
-        super(MetaAAC_2_0, self).__init__(
-            runner_config=runner_config,
-            name=name,
-            **kwargs
-        )
-
-        self.current_data = None
-        self.current_feed_dict = None
-
-        # Trials sampling control:
-        self.num_source_trials = trial_source_target_cycle[0]
-        self.num_target_trials = trial_source_target_cycle[-1]
-        self.num_episodes_per_trial = num_episodes_per_trial
-
-        # Note that only master (test runner) is requesting trials
-
-        self.current_source_trial = 0
-        self.current_target_trial = 0
-        self.current_trial_mode = 0  # source
-        self.current_episode = 0
-
-    def get_sample_config(self, mode=0, **kwargs):
-        """
-        Returns environment configuration parameters for next episode to sample.
-
-        Args:
-              mode:     bool, False for slave (train data), True for master (test data)
-
-        Returns:
-            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
-        """
-
-        new_trial = 0
-
-        # Only master environment updates counters:
-        if self.current_episode >= self.num_episodes_per_trial:
-            # Reset episode counter:
-            self.current_episode = 0
-
-            # Request new trial:
-            new_trial = 1
-            # Decide on trial type (source/target):
-            if self.current_source_trial >= self.num_source_trials:
-                # Time to switch to target mode:
-                self.current_trial_mode = 1
-                # Reset counters:
-                self.current_source_trial = 0
-                self.current_target_trial = 0
-
-            if self.current_target_trial >= self.num_target_trials:
-                # Vise versa:
-                self.current_trial_mode = 0
-                self.current_source_trial = 0
-                self.current_target_trial = 0
-
-            # Update counter:
-            if self.current_trial_mode:
-                self.current_target_trial += 1
-            else:
-                self.current_source_trial += 1
-
-        self.current_episode += 1
-
-
-        # Compose btgym.datafeed.base.EnvResetConfig-consistent dict:
-        sample_config = dict(
-            episode_config=dict(
-                get_new=True,
-                sample_type=mode,
-                b_alpha=1.0,
-                b_beta=1.0
-            ),
-            trial_config=dict(
-                get_new=new_trial,
-                sample_type=self.current_trial_mode,
-                b_alpha=1.0,
-                b_beta=1.0
-            )
-        )
-        return sample_config
-
-    def get_episode(self, **kwargs):
-        data_streams = [runner.get_episode(**kwargs) for runner in self.runners]
-        return {key: [stream[key] for stream in data_streams] for key in data_streams[0].keys()}
-
-    def process(self, sess, **kwargs):
-        if self.task < 1:
-            self.process_test(sess)
-
-        else:
-            self.process_train(sess)
-
-    def process_test(self, sess):
-        """
-         test step.
-
-         Args:
-             sess (tensorflow.Session):   tf session obj.
-
-         """
-        # Copy from parameter server:
-        sess.run(self.sync_pi)
-        for i in range(1):
-            test_data = self.get_episode(init_context=0)
-            self.process_summary(sess, test_data)
-            #self.log.warning('self.current_episode: {}'.format(self.current_episode))
-
-        #time.sleep(5)
-
-    def process_train(self, sess):
-        """
-        Train step.
-
-        Args:
-            sess (tensorflow.Session):   tf session obj.
-
-        """
-        try:
-            # Say `No` to redundant summaries:
-            wirte_model_summary = \
-                self.local_steps % self.model_summary_freq == 0
-
-            # Collect train trajectory:
-            train_data = self.get_data()
-            feed_dict = self.process_data(sess,,,,, train_data,,
-                        # self.log.warning('Train data ok.')
-
-                        # Copy from parameter server:
-                        sess.run(self.sync_pi)
-            # self.log.warning('Sync ok.')
-
-            # Update pi_prime parameters wrt collected data:
-            if wirte_model_summary:
-                fetches = [self.train_op, self.model_summary_op, self.inc_step]
-            else:
-                fetches = [self.train_op, self.inc_step]
-
-            fetched = sess.run(fetches, feed_dict=feed_dict)
-
-            # self.log.warning('Train gradients ok.')
-
-            if wirte_model_summary:
-                model_summary = fetched[-2]
-
-            else:
-                model_summary = None
-
-            # Write down summaries:
-            self.process_summary(sess, train_data, model_summary)
-            self.local_steps += 1
-
-        except:
-            msg = 'process() exception occurred' + \
-                  '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
-            self.log.exception(msg)
-            raise RuntimeError(msg)
-
-
-
diff --git a/btgym/research/metalearn_2/_mldg_batch.py b/btgym/research/metalearn_2/_mldg_batch.py
index f3e5ae11..e69de29b 100644
--- a/btgym/research/metalearn_2/_mldg_batch.py
+++ b/btgym/research/metalearn_2/_mldg_batch.py
@@ -1,530 +0,0 @@
-import tensorflow as tf
-import numpy as np
-
-import sys
-from logbook import Logger, StreamHandler
-
-from btgym.research.mldg.aac import SubAAC
-from btgym.algorithms.runner.synchro import BaseSynchroRunner
-
-
-class MLDG():
-    """
-    Asynchronous implementation of MLDG algorithm
-    for continuous adaptation in dynamically changing environments.
-
-    Papers:
-        Da Li et al.,
-         "Learning to Generalize: Meta-Learning for Domain Generalization"
-         https://arxiv.org/abs/1710.03463
-
-        Maruan Al-Shedivat et al.,
-        "Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments"
-        https://arxiv.org/abs/1710.03641
-
-    """
-    def __init__(
-            self,
-            env,
-            task,
-            log_level,
-            aac_class_ref=SubAAC,
-            runner_config=None,
-            aac_lambda=1.0,
-            guided_lambda=1.0,
-            rollout_length=20,
-            train_support=300,
-            fast_adapt_num_steps=10,
-            fast_adapt_batch_size=32,
-            trial_source_target_cycle=(1, 0),
-            num_episodes_per_trial=1,  # one-shot adaptation
-            _aux_render_modes=('action_prob', 'value_fn', 'lstm_1_h', 'lstm_2_h'),
-            name='MLDG',
-            **kwargs
-    ):
-        try:
-            self.aac_class_ref = aac_class_ref
-            self.task = task
-            self.name = name
-            self.summary_writer = None
-
-            StreamHandler(sys.stdout).push_application()
-            self.log = Logger('{}_{}'.format(name, task), level=log_level)
-
-            self.rollout_length = rollout_length
-            self.train_support = train_support  # number of train experiences to collect
-            self.train_batch_size = int(self.train_support / self.rollout_length)
-            self.fast_adapt_num_steps = fast_adapt_num_steps
-            self.fast_adapt_batch_size = fast_adapt_batch_size
-
-            if runner_config is None:
-                self.runner_config = {
-                    'class_ref': BaseSynchroRunner,
-                    'kwargs': {},
-                }
-            else:
-                self.runner_config = runner_config
-
-            self.env_list = env
-
-            assert isinstance(self.env_list, list) and len(self.env_list) == 2, \
-                'Expected pair of environments, got: {}'.format(self.env_list)
-
-            # Instantiate two sub-trainers: one for test and one for train environments:
-
-            self.runner_config['kwargs']['data_sample_config'] = {'mode': 1}  # master
-            self.runner_config['kwargs']['name'] = 'master'
-
-            self.train_aac = aac_class_ref(
-                env=self.env_list[0],  # train data will be master environment TODO: really dumb data control. improve.
-                task=self.task,
-                log_level=log_level,
-                runner_config=self.runner_config,
-                aac_lambda=aac_lambda,
-                guided_lambda=guided_lambda,
-                rollout_length=self.rollout_length,
-                trial_source_target_cycle=trial_source_target_cycle,
-                num_episodes_per_trial=num_episodes_per_trial,
-                _use_target_policy=False,
-                _use_global_network=True,
-                _aux_render_modes=_aux_render_modes,
-                name=self.name + '_sub_Train',
-                **kwargs
-            )
-
-            self.runner_config['kwargs']['data_sample_config'] = {'mode': 0}  # master
-            self.runner_config['kwargs']['name'] = 'slave'
-
-            self.test_aac = aac_class_ref(
-                env=self.env_list[-1],  # test data -> slave env.
-                task=self.task,
-                log_level=log_level,
-                runner_config=self.runner_config,
-                aac_lambda=aac_lambda,
-                guided_lambda=guided_lambda,
-                rollout_length=self.rollout_length,
-                trial_source_target_cycle=trial_source_target_cycle,
-                num_episodes_per_trial=num_episodes_per_trial,
-                _use_target_policy=False,
-                _use_global_network=False,
-                global_step_op=self.train_aac.global_step,
-                global_episode_op=self.train_aac.global_episode,
-                inc_episode_op=self.train_aac.inc_episode,
-                _aux_render_modes=_aux_render_modes,
-                name=self.name + '_sub_Test',
-                **kwargs
-            )
-
-            self.local_steps = self.train_aac.local_steps
-            self.model_summary_freq = self.train_aac.model_summary_freq
-            #self.model_summary_op = self.train_aac.model_summary_op
-
-            self._make_train_op()
-            self.test_aac.model_summary_op = tf.summary.merge(
-                [self.test_aac.model_summary_op, self._combine_meta_summaries()],
-                name='meta_model_summary'
-            )
-
-        except:
-            msg = 'MLDG.__init()__ exception occurred' + \
-                  '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
-            self.log.exception(msg)
-            raise RuntimeError(msg)
-
-    def _make_train_op(self):
-        """
-
-        Defines:
-            tensors holding training op graph for sub trainers and self;
-        """
-        pi = self.train_aac.local_network
-        pi_prime = self.test_aac.local_network
-
-        self.test_aac.sync = self.test_aac.sync_pi = tf.group(
-            *[v1.assign(v2) for v1, v2 in zip(pi_prime.var_list, pi.var_list)]
-        )
-
-        self.global_step = self.train_aac.global_step
-        self.global_episode = self.train_aac.global_episode
-
-        self.test_aac.global_step = self.train_aac.global_step
-        self.test_aac.global_episode = self.train_aac.global_episode
-        self.test_aac.inc_episode = self.train_aac.inc_episode
-        self.train_aac.inc_episode = None
-        self.inc_step = self.train_aac.inc_step
-
-        # Meta-loss:
-        self.loss = 0.5 * self.train_aac.loss + 0.5 * self.test_aac.loss
-
-        # Clipped gradients:
-        self.train_aac.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.train_aac.loss, pi.var_list),
-            40.0
-        )
-        self.log.warning('self.train_aac.grads: {}'.format(len(list(self.train_aac.grads))))
-
-        # self.test_aac.grads, _ = tf.clip_by_global_norm(
-        #     tf.gradients(self.test_aac.loss, pi_prime.var_list),
-        #     40.0
-        # )
-        # Meta-gradient:
-        grads_i, _ = tf.clip_by_global_norm(
-            tf.gradients(self.train_aac.loss, pi.var_list),
-            40.0
-        )
-
-        grads_i_next, _ = tf.clip_by_global_norm(
-            tf.gradients(self.test_aac.loss, pi_prime.var_list),
-            40.0
-        )
-
-        self.grads = []
-        for g1, g2 in zip(grads_i, grads_i_next):
-            if g1 is not None and g2 is not None:
-                meta_g = 0.5 * g1 + 0.5 * g2
-            else:
-                meta_g = None
-
-            self.grads.append(meta_g)
-
-        #self.log.warning('self.grads_len: {}'.format(len(list(self.grads))))
-
-        # Gradients to update local copy of pi_prime (from train data):
-        train_grads_and_vars = list(zip(self.train_aac.grads, pi_prime.var_list))
-
-        # self.log.warning('train_grads_and_vars_len: {}'.format(len(train_grads_and_vars)))
-
-        # Meta-gradients to be sent to parameter server:
-        meta_grads_and_vars = list(zip(self.grads, self.train_aac.network.var_list))
-
-        # self.log.warning('meta_grads_and_vars_len: {}'.format(len(meta_grads_and_vars)))
-
-        # Set global_step increment equal to observation space batch size:
-        obs_space_keys = list(self.train_aac.local_network.on_state_in.keys())
-
-        assert 'external' in obs_space_keys, \
-            'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
-        self.train_aac.inc_step = self.train_aac.global_step.assign_add(
-            tf.shape(self.train_aac.local_network.on_state_in['external'])[0]
-        )
-
-        self.train_op = self.train_aac.optimizer.apply_gradients(train_grads_and_vars)
-
-        # Optimizer for meta-update:
-        self.optimizer = tf.train.AdamOptimizer(self.train_aac.train_learn_rate, epsilon=1e-5)
-        # TODO: own alpha-leran rate
-        self.meta_train_op = self.optimizer.apply_gradients(meta_grads_and_vars)
-
-        self.log.debug('meta_train_op defined')
-
-    def _combine_meta_summaries(self):
-
-        meta_model_summaries = [
-            tf.summary.scalar("meta_grad_global_norm", tf.global_norm(self.grads)),
-            tf.summary.scalar("total_meta_loss", self.loss),
-        ]
-
-        return meta_model_summaries
-
-    def start(self, sess, summary_writer, **kwargs):
-        """
-        Executes all initializing operations,
-        starts environment runner[s].
-        Supposed to be called by parent worker just before training loop starts.
-
-        Args:
-            sess:           tf session object.
-            kwargs:         not used by default.
-        """
-        try:
-            # Copy weights from global to local:
-            sess.run(self.train_aac.sync_pi)
-            sess.run(self.test_aac.sync_pi)
-
-            # Start thread_runners:
-            self.train_aac._start_runners(   # master first
-                sess,
-                summary_writer,
-                init_context=None,
-                data_sample_config=self.train_aac.get_sample_config(mode=1)
-            )
-            self.test_aac._start_runners(
-                sess,
-                summary_writer,
-                init_context=None,
-                data_sample_config=self.test_aac.get_sample_config(mode=0)
-            )
-
-            self.summary_writer = summary_writer
-            self.log.notice('Runners started.')
-
-        except:
-            msg = 'start() exception occurred' + \
-                '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
-            self.log.exception(msg)
-            raise RuntimeError(msg)
-
-    def fast_adapt_step(self, sess, batch_size, on_policy_batch, off_policy_batch, rp_batch, make_summary=False):
-        """
-        One step of test_policy adaptation.
-
-        Args:
-            sess:                   tensorflow.Session obj.
-            batch_size:             train mini-batch size
-            on_policy_batch:        `on_policy` train data
-            off_policy_batch:       `off_policy` train data or None
-            rp_batch:               'reward_prediction` train data or None
-            make_summary:           bool, if True - compute model summary
-
-        Returns:
-            model summary or None
-        """
-        # Sample from train distribution:
-        on_mini_batch = self.train_aac.sample_batch(on_policy_batch, batch_size)
-        off_mini_batch = self.train_aac.sample_batch(off_policy_batch, batch_size)
-        rp_mini_batch = self.train_aac.sample_batch(rp_batch, batch_size)
-
-        feed_dict = self.train_aac._get_main_feeder(sess, on_mini_batch, off_mini_batch, rp_mini_batch, True)
-
-        if make_summary:
-            fetches = [self.train_op, self.train_aac.model_summary_op]
-        else:
-            fetches = [self.train_op]
-
-        # Update pi_prime parameters wrt sampled data:
-        fetched = sess.run(fetches, feed_dict=feed_dict)
-
-        # self.log.warning('Train gradients ok.')
-
-        if make_summary:
-            summary =  fetched[-1]
-
-        else:
-            summary = None
-
-        return summary
-
-    def train_step(self, sess, data_config):
-        """
-        Collects train task data and updates test policy parameters (fast adaptation).
-
-        Args:
-            sess:                   tensorflow.Session obj.
-            data_config:            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
-
-        Returns:
-            batched train data
-
-        """
-        # Collect train distribution:
-        train_batch = self.train_aac.get_batch(
-            size=self.train_batch_size,
-            require_terminal=True,
-            same_trial=True,
-            data_sample_config=data_config
-        )
-
-        # for rollout in train_batch['on_policy']:
-        #     self.log.warning(
-        #         'Train data trial_num: {}'.format(
-        #             np.asarray(rollout['state']['metadata']['trial_num'])
-        #         )
-        #     )
-
-        # Process time-flat-alike (~iid) to treat as empirical data distribution over train task:
-        on_policy_batch, off_policy_batch, rp_batch = self.train_aac.process_batch(sess, train_batch)
-
-        # self.log.warning('Train data ok.')
-
-        local_step = sess.run(self.global_step)
-        local_episode = sess.run(self.global_episode)
-        model_summary = None
-
-        # Extract all non-empty summaries:
-        ep_summary = [summary for summary in train_batch['ep_summary'] if summary is not None]
-
-        # Perform number of test policy updates wrt. collected train data:
-        for i in range(self.fast_adapt_num_steps):
-            model_summary = self.fast_adapt_step(
-                sess,
-                batch_size=self.fast_adapt_batch_size,
-                on_policy_batch=on_policy_batch,
-                off_policy_batch=off_policy_batch,
-                rp_batch=rp_batch,
-                make_summary=(local_step + i) % self.model_summary_freq == 0
-            )
-            # self.log.warning('Batch {} Train gradients ok.'.format(i))
-
-            # Write down summaries:
-            train_summary = dict(
-                render_summary=[None],
-                test_ep_summary=[None],
-                ep_summary=[ep_summary.pop() if len(ep_summary) > 0 else None]
-            )
-            self.train_aac.process_summary(
-                sess,
-                train_summary,
-                model_summary,
-                step=local_step + i,
-                episode=local_episode + i
-            )
-
-        return on_policy_batch, off_policy_batch, rp_batch
-
-    def meta_train_step(self, sess, data_config, on_policy_batch, off_policy_batch, rp_batch):
-        """
-        Collects data from source domain test task and performs meta-update to shared parameters vector.
-        Writes down relevant summaries.
-
-        Args:
-            sess:                   tensorflow.Session obj.
-            data_config:            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
-            on_policy_batch:        `on_policy` train data
-            off_policy_batch:       `off_policy` train data or None
-            rp_batch:               'reward_prediction` train data or None
-
-        """
-        done = False
-        while not done:
-            # Say `No` to redundant summaries:
-            wirte_model_summary = \
-                self.local_steps % self.model_summary_freq == 0
-
-            # Collect test trajectory wrt updated test_policy parameters:
-            test_data = self.test_aac.get_data(
-                init_context=None,
-                data_sample_config=data_config
-            )
-            test_batch_size = 0  # TODO: adjust on/off/rp sizes
-            for rollout in test_data['on_policy']:
-                test_batch_size += len(rollout['position'])
-
-            test_feed_dict = self.test_aac.process_data(sess,,,,, test_data,,
-
-                             # self.log.warning('Test data rollout for step {} ok.'.format(self.local_steps))
-                             #
-                             # self.log.warning(
-                             #     'Test data trial_num: {}'.format(
-                             #         np.asarray(test_data['on_policy'][0]['state']['metadata']['trial_num'])
-                             #     )
-                             # )
-
-                             # Sample train data of same size:
-                             feed_dict = self.train_aac._get_main_feeder(
-                sess,
-                self.train_aac.sample_batch(on_policy_batch, test_batch_size),
-                self.train_aac.sample_batch(off_policy_batch, test_batch_size),
-                self.train_aac.sample_batch(rp_batch, test_batch_size),
-                True
-            )
-            # Add test trajectory:
-            feed_dict.update(test_feed_dict)
-
-            # Perform meta-update:
-            if wirte_model_summary:
-                meta_fetches = [self.meta_train_op, self.test_aac.model_summary_op, self.inc_step]
-            else:
-                meta_fetches = [self.meta_train_op, self.inc_step]
-
-            meta_fetched = sess.run(meta_fetches, feed_dict=feed_dict)
-
-            # self.log.warning('Meta-gradients ok.')
-
-            if wirte_model_summary:
-                meta_model_summary = meta_fetched[-2]
-
-            else:
-                meta_model_summary = None
-
-            # Write down summaries:
-            self.test_aac.process_summary(sess, test_data, meta_model_summary)
-            self.local_steps += 1
-
-            # If test episode ended?
-            done = np.asarray(test_data['terminal']).any()
-
-    def meta_test_step(self, sess, data_config, on_policy_batch, off_policy_batch, rp_batch):
-        """
-        Validates adapted policy on data from target domain test task.
-        Writes down relevant summaries.
-
-        Args:
-            sess:                   tensorflow.Session obj.
-            data_config:            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
-            on_policy_batch:        `on_policy` train data
-            off_policy_batch:       `off_policy` train data or None
-            rp_batch:               'reward_prediction` train data or None
-
-        """
-        done = False
-        while not done:
-            # Collect test trajectory:
-            test_data = self.test_aac.get_data(
-                init_context=None,
-                data_sample_config=data_config
-            )
-
-            # self.log.warning('Target test rollout ok.')
-            # self.log.warning(
-            #     'Test data target trial_num: {}'.format(
-            #         np.asarray(test_data['on_policy'][0]['state']['metadata']['trial_num'])
-            #     )
-            # )
-            # self.log.warning('target_render_ep_summary: {}'.format(test_data['render_summary']))
-
-            # Write down summaries:
-            self.test_aac.process_summary(sess, test_data)
-
-            # If test episode ended?
-            done = np.asarray(test_data['terminal']).any()
-
-    def process(self, sess):
-        """
-        Meta-train procedure for one-shot learning/
-
-        Args:
-            sess (tensorflow.Session):   tf session obj.
-
-        """
-        try:
-            # Copy from parameter server:
-            sess.run(self.train_aac.sync_pi)
-            sess.run(self.test_aac.sync_pi)
-
-            #self.log.warning('Sync ok.')
-
-            # Decide on data configuration for train/test trajectories,
-            # such as all data will come from same trial (maybe different episodes)
-            # and trial type as well (~from source or target domain):
-            # note: data_config counters get updated once per process() call
-            train_data_config = self.train_aac.get_sample_config(mode=1)  # master env., draws trial
-            test_data_config = self.train_aac.get_sample_config(mode=0)   # slave env, catches up with same trial
-
-            # If data comes from source or target domain:
-            is_target = train_data_config['trial_config']['sample_type']
-
-            # self.log.warning('PROCESS_train_data_config: {}'.format(train_data_config))
-            # self.log.warning('PROCESS_test_data_config: {}'.format(test_data_config))
-
-            # Fast adaptation step:
-            # collect train trajectories, process time-flat-alike (~iid) to treat as empirical data distribution
-            # over train task and adapt test_policy wrt. train experience:
-            on_policy_batch, off_policy_batch, rp_batch = self.train_step(sess, train_data_config)
-
-            # Slow adaptation step:
-            if is_target:
-                # Meta-test:
-                # self.log.warning('Running meta-test episode...')
-                self.meta_test_step(sess,test_data_config, on_policy_batch, off_policy_batch, rp_batch)
-
-            else:
-                # Meta-train:
-                # self.log.warning('Running meta-train episode...')
-                self.meta_train_step(sess,test_data_config, on_policy_batch, off_policy_batch, rp_batch)
-
-        except:
-            msg = 'process() exception occurred' + \
-                  '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
-            self.log.exception(msg)
-            raise RuntimeError(msg)
-
diff --git a/btgym/research/metalearn_2/loss.py b/btgym/research/metalearn_2/loss.py
index cb845436..15f709bf 100644
--- a/btgym/research/metalearn_2/loss.py
+++ b/btgym/research/metalearn_2/loss.py
@@ -19,37 +19,37 @@ def meta_loss_def_1_0(
         name='_meta_',
         verbose=False
 ):
-    with tf.name_scope(name + '/meta'):
-        neg_pi_log_prob_train = tf.nn.softmax_cross_entropy_with_logits_v2(
+    with tf.compat.v1.name_scope(name + '/meta'):
+        neg_pi_log_prob_train = tf.nn.softmax_cross_entropy_with_logits(
             logits=pi_logits_train,
             labels=act_target_train
         )
-        neg_pi_log_prob_test = tf.nn.softmax_cross_entropy_with_logits_v2(
+        neg_pi_log_prob_test = tf.nn.softmax_cross_entropy_with_logits(
             logits=pi_logits_test,
             labels=act_target_test
         )
         pi_loss = tf.reduce_mean(
-            (neg_pi_log_prob_train + neg_pi_log_prob_test) * adv_target_test
+            input_tensor=(neg_pi_log_prob_train + neg_pi_log_prob_test) * adv_target_test
         )
-        vf_loss_train = 0.5 * tf.losses.mean_squared_error(r_target_test, pi_vf_train)
-        vf_loss_test = 0.5 * tf.losses.mean_squared_error(r_target_test, pi_vf_test)
+        vf_loss_train = 0.5 * tf.compat.v1.losses.mean_squared_error(r_target_test, pi_vf_train)
+        vf_loss_test = 0.5 * tf.compat.v1.losses.mean_squared_error(r_target_test, pi_vf_test)
 
-        entropy = tf.reduce_mean(cat_entropy(pi_logits_test))
+        entropy = tf.reduce_mean(input_tensor=cat_entropy(pi_logits_test))
 
         loss = pi_loss + vf_loss_test + vf_loss_train - entropy * entropy_beta
 
-        mean_vf_test = tf.reduce_mean(pi_vf_test)
-        mean_vf_train = tf.reduce_mean(pi_vf_train)
+        mean_vf_test = tf.reduce_mean(input_tensor=pi_vf_test)
+        mean_vf_train = tf.reduce_mean(input_tensor=pi_vf_train)
 
         summaries = [
-            tf.summary.scalar('meta_policy_loss', pi_loss),
-            tf.summary.scalar('meta_value_loss_test', vf_loss_test),
+            tf.compat.v1.summary.scalar('meta_policy_loss', pi_loss),
+            tf.compat.v1.summary.scalar('meta_value_loss_test', vf_loss_test),
         ]
         if verbose:
             summaries += [
-                tf.summary.scalar('entropy', entropy),
-                tf.summary.scalar('value_fn_test', mean_vf_test),
-                tf.summary.scalar('value_fn_train', mean_vf_train)
+                tf.compat.v1.summary.scalar('entropy', entropy),
+                tf.compat.v1.summary.scalar('value_fn_test', mean_vf_test),
+                tf.compat.v1.summary.scalar('value_fn_train', mean_vf_train)
             ]
 
     return loss, summaries
diff --git a/btgym/research/mldg/aac.py b/btgym/research/mldg/aac.py
index 93aedad1..8c52d5a3 100644
--- a/btgym/research/mldg/aac.py
+++ b/btgym/research/mldg/aac.py
@@ -235,7 +235,7 @@ def __init__(
 
             self._make_train_op()
 
-            self.test_aac.model_summary_op = tf.summary.merge(
+            self.test_aac.model_summary_op = tf.compat.v1.summary.merge(
                 [self.test_aac.model_summary_op, self._combine_meta_summaries()],
                 name='meta_model_summary'
             )
@@ -279,11 +279,11 @@ def _make_train_op(self):
 
         # Clipped gradients:
         self.train_aac.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.train_aac.loss, pi.var_list),
+            tf.gradients(ys=self.train_aac.loss, xs=pi.var_list),
             40.0
         )
         self.test_aac.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.test_aac.loss, pi_prime.var_list),
+            tf.gradients(ys=self.test_aac.loss, xs=pi_prime.var_list),
             40.0
         )
         # Aliases:
@@ -320,18 +320,18 @@ def _make_train_op(self):
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
         self.train_aac.inc_step = self.train_aac.global_step.assign_add(
-            tf.shape(self.test_aac.local_network.on_state_in['external'])[0]
+            tf.shape(input=self.test_aac.local_network.on_state_in['external'])[0]
         )
         self.inc_step = self.train_aac.inc_step
         # Pi to pi_prime local adaptation op:
         # self.train_op = self.train_aac.optimizer.apply_gradients(train_grads_and_vars)
 
         # self.fast_opt = tf.train.GradientDescentOptimizer(self.alpha_rate)
-        self.fast_opt = tf.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
+        self.fast_opt = tf.compat.v1.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
         self.train_op = self.fast_opt.apply_gradients(train_grads_and_vars)
 
         #  Learning rate annealing:
-        self.learn_rate_decayed = tf.train.polynomial_decay(
+        self.learn_rate_decayed = tf.compat.v1.train.polynomial_decay(
             self.opt_learn_rate,
             self.global_step + 1,
             self.opt_decay_steps,
@@ -341,7 +341,7 @@ def _make_train_op(self):
         )
 
         # Optimizer for meta-update, sharing same learn rate (change?):
-        self.optimizer = tf.train.AdamOptimizer(self.learn_rate_decayed, epsilon=1e-5)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.learn_rate_decayed, epsilon=1e-5)
 
         # Global meta-optimisation op:
         self.meta_train_op = self.optimizer.apply_gradients(meta_grads_and_vars)
@@ -353,8 +353,8 @@ def _combine_meta_summaries(self):
         Additional summaries here.
         """
         meta_model_summaries = [
-            tf.summary.scalar('meta_grad_global_norm', tf.global_norm(self.grads)),
-            tf.summary.scalar('total_meta_loss', self.loss),
+            tf.compat.v1.summary.scalar('meta_grad_global_norm', tf.linalg.global_norm(self.grads)),
+            tf.compat.v1.summary.scalar('total_meta_loss', self.loss),
             #tf.summary.scalar('alpha_learn_rate', self.alpha_rate),
             #tf.summary.scalar('alpha_learn_rate_loss', self.alpha_rate_loss)
         ]
@@ -848,11 +848,11 @@ def _make_train_op(self):
 
         # Clipped gradients:
         self.train_aac.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.train_aac.loss, pi.var_list),
+            tf.gradients(ys=self.train_aac.loss, xs=pi.var_list),
             40.0
         )
         self.test_aac.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.test_aac.loss, pi_prime.var_list),
+            tf.gradients(ys=self.test_aac.loss, xs=pi_prime.var_list),
             40.0
         )
         # Aliases:
@@ -889,17 +889,17 @@ def _make_train_op(self):
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
         self.train_aac.inc_step = self.train_aac.global_step.assign_add(
-            tf.shape(self.train_aac.local_network.on_state_in['external'])[0]
+            tf.shape(input=self.train_aac.local_network.on_state_in['external'])[0]
         )
         # Simple SGD, no average statisitics:
-        self.fast_optimizer_train = tf.train.GradientDescentOptimizer(self.fast_learn_rate_train)
-        self.fast_optimizer_test = tf.train.GradientDescentOptimizer(self.fast_learn_rate_test)
+        self.fast_optimizer_train = tf.compat.v1.train.GradientDescentOptimizer(self.fast_learn_rate_train)
+        self.fast_optimizer_test = tf.compat.v1.train.GradientDescentOptimizer(self.fast_learn_rate_test)
 
         # Pi to pi_prime local adaptation op:
         self.train_op = self.fast_optimizer_train.apply_gradients(train_grads_and_vars)
 
         # Optimizer for meta-update, sharing same learn rate (change?):
-        self.optimizer = tf.train.AdamOptimizer(self.train_aac.train_learn_rate, epsilon=1e-5)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_aac.train_learn_rate, epsilon=1e-5)
 
         # Global meta-optimisation op:
         self.meta_train_op = self.optimizer.apply_gradients(meta_grads_and_vars)
diff --git a/btgym/research/mldg/aac_1.py b/btgym/research/mldg/aac_1.py
index 45a1d7d8..4d34d1cd 100644
--- a/btgym/research/mldg/aac_1.py
+++ b/btgym/research/mldg/aac_1.py
@@ -73,7 +73,7 @@ def __init__(
                 name=name,
                 **kwargs
             )
-            self.model_summary_op = tf.summary.merge(
+            self.model_summary_op = tf.compat.v1.summary.merge(
                 [self.model_summary_op, self._combine_meta_summaries()],
                 name='meta_model_summary'
             )
@@ -98,7 +98,7 @@ def _make_loss(self, pi, pi_prime):
         )
         # Guidance annealing:
         if self.guided_decay_steps is not None:
-            self.guided_lambda_decayed = tf.train.polynomial_decay(
+            self.guided_lambda_decayed = tf.compat.v1.train.polynomial_decay(
                 self.guided_lambda,
                 self.global_step + 1,
                 self.guided_decay_steps,
@@ -152,16 +152,16 @@ def _make_train_op(self, pi, pi_prime, pi_global):
             *[v1.assign(v2) for v1, v2 in zip(pi.var_list, pi_prime.var_list)]
         )
         self.sync = [self.sync_pi, self.sync_pi_prime]
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
-        self.fast_optimizer = tf.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.fast_optimizer = tf.compat.v1.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
 
         # Clipped gradients:
         pi.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.meta_train_loss, pi.var_list),
+            tf.gradients(ys=self.meta_train_loss, xs=pi.var_list),
             40.0
         )
         pi_prime.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.meta_test_loss, pi_prime.var_list),
+            tf.gradients(ys=self.meta_test_loss, xs=pi_prime.var_list),
             40.0
         )
         # Meta_optimisation gradients as sum of meta-train and meta-test gradients:
@@ -189,7 +189,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
 
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
-        self.inc_step = self.global_step.assign_add(tf.shape(self.local_network.on_state_in['external'])[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=self.local_network.on_state_in['external'])[0])
 
         # Local fast optimisation op:
         self.fast_train_op = self.fast_optimizer.apply_gradients(train_grads_and_vars)
@@ -204,9 +204,9 @@ def _combine_meta_summaries(self):
         """
         Additional summaries here.
         """
-        with tf.name_scope(self.name):
+        with tf.compat.v1.name_scope(self.name):
             meta_model_summaries = [
-                tf.summary.scalar('meta_grad_global_norm', tf.global_norm(self.grads)),
+                tf.compat.v1.summary.scalar('meta_grad_global_norm', tf.linalg.global_norm(self.grads)),
                 # tf.summary.scalar('total_meta_loss', self.loss),
                 # tf.summary.scalar('alpha_learn_rate', self.alpha_rate),
                 # tf.summary.scalar('alpha_learn_rate_loss', self.alpha_rate_loss)
diff --git a/btgym/research/mldg/aac_1d.py b/btgym/research/mldg/aac_1d.py
index 1a24ab40..005b1023 100644
--- a/btgym/research/mldg/aac_1d.py
+++ b/btgym/research/mldg/aac_1d.py
@@ -97,16 +97,16 @@ def _make_train_op(self, pi, pi_prime, pi_global):
             *[v1.assign(v2) for v1, v2 in zip(pi.var_list, pi_prime.var_list)]
         )
         self.sync = [self.sync_pi, self.sync_pi_prime]
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
-        self.fast_optimizer = tf.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.fast_optimizer = tf.compat.v1.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
 
         # Clipped gradients:
         pi.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.meta_train_loss, pi.var_list),
+            tf.gradients(ys=self.meta_train_loss, xs=pi.var_list),
             40.0
         )
         pi_prime.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.meta_test_loss, pi_prime.var_list),
+            tf.gradients(ys=self.meta_test_loss, xs=pi_prime.var_list),
             40.0
         )
         # Meta_optimisation gradients as sum of meta-train and meta-test gradients:
@@ -134,7 +134,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
 
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
-        self.inc_step = self.global_step.assign_add(tf.shape(pi_prime.on_state_in['external'])[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=pi_prime.on_state_in['external'])[0])
 
         # Local fast optimisation op:
         self.fast_train_op = self.fast_optimizer.apply_gradients(train_grads_and_vars)
diff --git a/btgym/research/mldg/aac_1s.py b/btgym/research/mldg/aac_1s.py
index 2f3312ea..4ab55db8 100644
--- a/btgym/research/mldg/aac_1s.py
+++ b/btgym/research/mldg/aac_1s.py
@@ -223,11 +223,11 @@ def _make_loss(self, pi, pi_prime, name='base', verbose=True):
             tensor holding estimated loss graph
             list of related summaries
         """
-        with tf.name_scope(name):
+        with tf.compat.v1.name_scope(name):
 
             # Guidance annealing:
             if self.guided_decay_steps is not None:
-                self.guided_lambda_decayed = tf.train.polynomial_decay(
+                self.guided_lambda_decayed = tf.compat.v1.train.polynomial_decay(
                     self.guided_lambda,
                     self.global_step + 1,
                     self.guided_decay_steps,
@@ -249,11 +249,11 @@ def _make_loss(self, pi, pi_prime, name='base', verbose=True):
             )
 
             # On-policy AAC loss definition:
-            pi.on_pi_act_target = tf.placeholder(
+            pi.on_pi_act_target = tf.compat.v1.placeholder(
                 tf.float32, [None, self.ref_env.action_space.n], name="on_policy_action_pl"
             )
-            pi.on_pi_adv_target = tf.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
-            pi.on_pi_r_target = tf.placeholder(tf.float32, [None], name="on_policy_return_pl")
+            pi.on_pi_adv_target = tf.compat.v1.placeholder(tf.float32, [None], name="on_policy_advantage_pl")
+            pi.on_pi_r_target = tf.compat.v1.placeholder(tf.float32, [None], name="on_policy_return_pl")
 
             clip_epsilon = tf.cast(self.clip_epsilon * self.learn_rate_decayed / self.opt_learn_rate, tf.float32)
 
@@ -273,10 +273,10 @@ def _make_loss(self, pi, pi_prime, name='base', verbose=True):
             model_summaries = on_pi_summaries + g_summary
 
             # Off-policy losses:
-            pi.off_pi_act_target = tf.placeholder(
+            pi.off_pi_act_target = tf.compat.v1.placeholder(
                 tf.float32, [None, self.ref_env.action_space.n], name="off_policy_action_pl")
-            pi.off_pi_adv_target = tf.placeholder(tf.float32, [None], name="off_policy_advantage_pl")
-            pi.off_pi_r_target = tf.placeholder(tf.float32, [None], name="off_policy_return_pl")
+            pi.off_pi_adv_target = tf.compat.v1.placeholder(tf.float32, [None], name="off_policy_advantage_pl")
+            pi.off_pi_r_target = tf.compat.v1.placeholder(tf.float32, [None], name="off_policy_return_pl")
 
             if self.use_off_policy_aac:
                 # Off-policy AAC loss graph mirrors on-policy:
@@ -296,7 +296,7 @@ def _make_loss(self, pi, pi_prime, name='base', verbose=True):
 
             if self.use_value_replay:
                 # Value function replay loss:
-                pi.vr_target = tf.placeholder(tf.float32, [None], name="vr_target")
+                pi.vr_target = tf.compat.v1.placeholder(tf.float32, [None], name="vr_target")
                 self.vr_loss, self.vr_summaries = self.vr_loss(
                     r_target=pi.vr_target,
                     pi_vf=pi.vr_value,
@@ -323,16 +323,16 @@ def _make_train_op(self, pi, pi_prime, pi_global):
             *[v1.assign(v2) for v1, v2 in zip(pi.var_list, pi_global.var_list)]
         )
         self.sync = self.sync_pi
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
-        self.fast_optimizer = tf.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.fast_optimizer = tf.compat.v1.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
 
         # Clipped gradients:
         pi.on_grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.on_pi_loss, pi.var_list),
+            tf.gradients(ys=self.on_pi_loss, xs=pi.var_list),
             40.0
         )
         pi.off_grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.off_pi_loss, pi.var_list),
+            tf.gradients(ys=self.off_pi_loss, xs=pi.var_list),
             40.0
         )
 
@@ -354,7 +354,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
 
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
-        self.inc_step = self.global_step.assign_add(tf.shape(pi.on_state_in['external'])[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=pi.on_state_in['external'])[0])
 
         # Local fast optimisation op:
         self.local_train_op = self.fast_optimizer.apply_gradients(local_grads_and_vars)
@@ -501,22 +501,22 @@ def _make_train_op(self, pi, pi_prime, pi_global):
             *[v1.assign(v2) for v1, v2 in zip(pi.var_list, pi_global.var_list)]
         )
         self.sync = self.sync_pi
-        self.optimizer = tf.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
+        self.optimizer = tf.compat.v1.train.AdamOptimizer(self.train_learn_rate, epsilon=1e-5)
 
         # Clipped gradients:
         pi.on_grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.on_pi_loss, pi.var_list),
+            tf.gradients(ys=self.on_pi_loss, xs=pi.var_list),
             40.0
         )
         pi.off_grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.off_pi_loss, pi.var_list),
+            tf.gradients(ys=self.off_pi_loss, xs=pi.var_list),
             40.0
         )
         self.grads = pi.on_grads
 
         # Learnable fast rate:
         #self.fast_learn_rate = tf.reduce_mean(pi.off_learn_alpha, name='mean_alpha_rate') / 10
-        self.fast_optimizer = tf.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
+        self.fast_optimizer = tf.compat.v1.train.GradientDescentOptimizer(self.fast_opt_learn_rate)
         # self.alpha_rate_loss = tf.global_norm(pi.off_grads)
         # self.alpha_grads, _ = tf.clip_by_global_norm(
         #     tf.gradients(self.alpha_rate_loss, pi.var_list),
@@ -539,7 +539,7 @@ def _make_train_op(self, pi, pi_prime, pi_global):
 
         assert 'external' in obs_space_keys, \
             'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
-        self.inc_step = self.global_step.assign_add(tf.shape(pi.on_state_in['external'])[0])
+        self.inc_step = self.global_step.assign_add(tf.shape(input=pi.on_state_in['external'])[0])
 
         # Local fast optimisation op:
         self.local_train_op = self.fast_optimizer.apply_gradients(local_grads_and_vars)
diff --git a/btgym/research/mldg/policy.py b/btgym/research/mldg/policy.py
index 09a2e7ec..71899a59 100644
--- a/btgym/research/mldg/policy.py
+++ b/btgym/research/mldg/policy.py
@@ -46,7 +46,7 @@ def get_initial_features(self, state, context=None):
         #print('Meta_policy_init_metadata:', state['metadata'])
         #print('self.current_trial_num:', self.current_trial_num)
         try:
-            sess = tf.get_default_session()
+            sess = tf.compat.v1.get_default_session()
             new_context = list(sess.run(self.on_lstm_init_state))
             if context is not None:
                 if state['metadata']['trial_num'] == self.current_trial_num or state['metadata']['type']:
diff --git a/btgym/research/model_based/aac.py b/btgym/research/model_based/aac.py
index facb3e9a..33573776 100644
--- a/btgym/research/model_based/aac.py
+++ b/btgym/research/model_based/aac.py
@@ -44,67 +44,67 @@ def _combine_summaries(self, policy=None, model_summaries=None):
         if model_summaries is not None:
             if self.use_global_network:
                 # Model-wide statistics:
-                with tf.name_scope('model'):
+                with tf.compat.v1.name_scope('model'):
                     model_summaries += [
-                        tf.summary.scalar("grad_global_norm", self.grads_global_norm),
-                        tf.summary.scalar("learn_rate", self.learn_rate_decayed),
+                        tf.compat.v1.summary.scalar("grad_global_norm", self.grads_global_norm),
+                        tf.compat.v1.summary.scalar("learn_rate", self.learn_rate_decayed),
                         # cause actual rate is a jaggy due to test freezes
-                        tf.summary.scalar("total_loss", self.loss),
+                        tf.compat.v1.summary.scalar("total_loss", self.loss),
                     ]
                     if policy is not None:
-                        model_summaries += [tf.summary.scalar("var_global_norm", tf.global_norm(policy.var_list))]
+                        model_summaries += [tf.compat.v1.summary.scalar("var_global_norm", tf.linalg.global_norm(policy.var_list))]
         else:
             model_summaries = []
         # Model stat. summary:
-        model_summary = tf.summary.merge(model_summaries, name='model_summary')
+        model_summary = tf.compat.v1.summary.merge(model_summaries, name='model_summary')
 
         # Episode-related summaries:
         ep_summary = dict(
             # Summary placeholders
-            render_atari=tf.placeholder(tf.uint8, [None, None, None, 1]),
-            total_r=tf.placeholder(tf.float32, ),
-            cpu_time=tf.placeholder(tf.float32, ),
-            final_value=tf.placeholder(tf.float32, ),
-            steps=tf.placeholder(tf.int32, ),
-            ou_lambda=tf.placeholder(tf.float32, ),
-            ou_sigma=tf.placeholder(tf.float32, ),
-            ou_mu=tf.placeholder(tf.float32, ),
+            render_atari=tf.compat.v1.placeholder(tf.uint8, [None, None, None, 1]),
+            total_r=tf.compat.v1.placeholder(tf.float32, ),
+            cpu_time=tf.compat.v1.placeholder(tf.float32, ),
+            final_value=tf.compat.v1.placeholder(tf.float32, ),
+            steps=tf.compat.v1.placeholder(tf.int32, ),
+            ou_lambda=tf.compat.v1.placeholder(tf.float32, ),
+            ou_sigma=tf.compat.v1.placeholder(tf.float32, ),
+            ou_mu=tf.compat.v1.placeholder(tf.float32, ),
 
         )
         ep_summary.update(
             {
-                mode: tf.placeholder(tf.uint8, [None, None, None, None], name=mode + '_pl')
+                mode: tf.compat.v1.placeholder(tf.uint8, [None, None, None, None], name=mode + '_pl')
                 for mode in self.env_list[0].render_modes + self.aux_render_modes
             }
         )
-        ep_summary['render_op'] = tf.summary.merge(
-            [tf.summary.image(mode, ep_summary[mode])
+        ep_summary['render_op'] = tf.compat.v1.summary.merge(
+            [tf.compat.v1.summary.image(mode, ep_summary[mode])
              for mode in self.env_list[0].render_modes + self.aux_render_modes]
         )
         # Episode stat. summary:
-        ep_summary['btgym_stat_op'] = tf.summary.merge(
+        ep_summary['btgym_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode_train/cpu_time_sec', ep_summary['cpu_time']),
-                tf.summary.scalar('episode_train/final_value', ep_summary['final_value']),
-                tf.summary.scalar('episode_train/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode_train/ou_lambda', ep_summary['ou_lambda']),
-                tf.summary.scalar('episode_train/ou_sigma', ep_summary['ou_sigma']),
-                tf.summary.scalar('episode_train/ou_mu', ep_summary['ou_mu']),
+                tf.compat.v1.summary.scalar('episode_train/cpu_time_sec', ep_summary['cpu_time']),
+                tf.compat.v1.summary.scalar('episode_train/final_value', ep_summary['final_value']),
+                tf.compat.v1.summary.scalar('episode_train/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode_train/ou_lambda', ep_summary['ou_lambda']),
+                tf.compat.v1.summary.scalar('episode_train/ou_sigma', ep_summary['ou_sigma']),
+                tf.compat.v1.summary.scalar('episode_train/ou_mu', ep_summary['ou_mu']),
             ],
             name='episode_train_btgym'
         )
         # Test episode stat. summary:
-        ep_summary['test_btgym_stat_op'] = tf.summary.merge(
+        ep_summary['test_btgym_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode_test/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode_test/final_value', ep_summary['final_value']),
+                tf.compat.v1.summary.scalar('episode_test/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode_test/final_value', ep_summary['final_value']),
             ],
             name='episode_test_btgym'
         )
-        ep_summary['atari_stat_op'] = tf.summary.merge(
+        ep_summary['atari_stat_op'] = tf.compat.v1.summary.merge(
             [
-                tf.summary.scalar('episode/total_reward', ep_summary['total_r']),
-                tf.summary.scalar('episode/steps', ep_summary['steps'])
+                tf.compat.v1.summary.scalar('episode/total_reward', ep_summary['total_r']),
+                tf.compat.v1.summary.scalar('episode/steps', ep_summary['steps'])
             ],
             name='episode_atari'
         )
@@ -280,16 +280,16 @@ def _make_train_op(self):
         # Clipped gradients for critic (critic's train op is disabled by `_use_global_network=False`
         # to avoid actor's name scope violation):
         self.critic_aac.grads, _ = tf.clip_by_global_norm(
-            tf.gradients(self.critic_aac.loss, pi_critic.var_list),
+            tf.gradients(ys=self.critic_aac.loss, xs=pi_critic.var_list),
             40.0
         )
         # Placeholders for stored gradients values, include None's to correctly map Vars:
         self.actor_aac.grads_placeholders = [
-            tf.placeholder(shape=grad.shape, dtype=grad.dtype) if grad is not None else None
+            tf.compat.v1.placeholder(shape=grad.shape, dtype=grad.dtype) if grad is not None else None
             for grad in self.actor_aac.grads
         ]
         self.critic_aac.grads_placeholders = [
-            tf.placeholder(shape=grad.shape, dtype=grad.dtype) if grad is not None else None
+            tf.compat.v1.placeholder(shape=grad.shape, dtype=grad.dtype) if grad is not None else None
             for grad in self.critic_aac.grads
         ]
 
@@ -317,7 +317,7 @@ def _make_train_op(self):
         self.inc_step = self.actor_aac.inc_step
 
         # Op to update critic with gradients from actor:
-        self.critic_aac.optimizer = tf.train.AdamOptimizer(self.actor_aac.learn_rate_decayed, epsilon=1e-5)
+        self.critic_aac.optimizer = tf.compat.v1.train.AdamOptimizer(self.actor_aac.learn_rate_decayed, epsilon=1e-5)
         self.update_critic_op = self.critic_aac.optimizer.apply_gradients(critic_grads_and_vars)
 
         # Use actor optimizer to update global policy instance:
diff --git a/btgym/research/policy_rl2.py b/btgym/research/policy_rl2.py
index 94e058c9..3fd24c69 100644
--- a/btgym/research/policy_rl2.py
+++ b/btgym/research/policy_rl2.py
@@ -41,7 +41,7 @@ def get_initial_features(self, state, context=None):
         try:
             if state['metadata']['trial_num'] != self.current_trial_num or context is None or state['metadata']['type']:
                 # Assume new/initial trial or test, reset context:
-                sess = tf.get_default_session()
+                sess = tf.compat.v1.get_default_session()
                 new_context = sess.run(self.on_lstm_init_state)
                 print('RL^2 policy context reset')
 
diff --git a/setup.py b/setup.py
index 95b3e342..f627e0c5 100644
--- a/setup.py
+++ b/setup.py
@@ -52,7 +52,7 @@
     ],
     version='0.0.8',
     install_requires=[
-        'tensorflow>=1.5',
+        'tensorflow>=2.3',
         'opencv-python',
         'gym[atari]',
         'backtrader',

From 8addf9cead9b1b728164effd09f443835f4eac0a Mon Sep 17 00:00:00 2001
From: Wojciech Indyk <wojciechindyk@gmail.com>
Date: Fri, 28 Aug 2020 20:56:07 +0200
Subject: [PATCH 2/6] Tf1 -> Tf2: LayerNormBasicLSTMCell to
 tfa.rnn.LayerNormLSTMCell

---
 btgym/algorithms/policy/stacked_lstm.py | 6 +++---
 btgym/research/b_vae_a3c.py             | 6 +++---
 btgym/research/encoder_test/policy.py   | 6 ++++--
 btgym/research/gps/policy.py            | 4 ++--
 setup.py                                | 1 +
 5 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/btgym/algorithms/policy/stacked_lstm.py b/btgym/algorithms/policy/stacked_lstm.py
index 160728e8..fca44d41 100644
--- a/btgym/algorithms/policy/stacked_lstm.py
+++ b/btgym/algorithms/policy/stacked_lstm.py
@@ -1,4 +1,4 @@
-from tensorflow.contrib.layers import flatten as batch_flatten
+import tensorflow_addons as tfa
 
 from btgym.algorithms.policy.base import BaseAacPolicy
 from btgym.algorithms.nn.networks import *
@@ -27,7 +27,7 @@ def __init__(self,
                  ac_space,
                  rp_sequence_size,
                  state_encoder_class_ref=conv_2d_network,
-                 lstm_class_ref=tf.contrib.rnn.LayerNormBasicLSTMCell,
+                 lstm_class_ref=tfa.rnn.LayerNormLSTMCell,
                  lstm_layers=(256, 256),
                  linear_layer_ref=noisy_linear,
                  share_encoder_params=False,
@@ -47,7 +47,7 @@ def __init__(self,
             ob_space:               instance of btgym.spaces.DictSpace
             ac_space:               instance of btgym.spaces.ActionDictSpace
             rp_sequence_size:       reward prediction sample length
-            lstm_class_ref:         tf.nn.lstm class to use
+            lstm_class_ref:         tfa.rnn class to use
             lstm_layers:            tuple of LSTM layers sizes
             linear_layer_ref:       linear layer class to use
             share_encoder_params:   bool, whether to share encoder parameters for every 'external' data stream
diff --git a/btgym/research/b_vae_a3c.py b/btgym/research/b_vae_a3c.py
index b1dc90fb..99342cb5 100644
--- a/btgym/research/b_vae_a3c.py
+++ b/btgym/research/b_vae_a3c.py
@@ -1,4 +1,4 @@
-from tensorflow.contrib.layers import flatten as batch_flatten
+import tensorflow_addons as tfa
 
 from btgym.algorithms.policy.base import BaseAacPolicy
 from btgym.algorithms.policy.stacked_lstm import AacStackedRL2Policy
@@ -51,7 +51,7 @@ def __init__(self,
                      (32, (3, 1), (2, 1)),
                      (32, (3, 1), (2, 1))
                  ),
-                 lstm_class_ref=tf.contrib.rnn.LayerNormBasicLSTMCell,
+                 lstm_class_ref=tfa.rnn.LayerNormLSTMCell,
                  lstm_layers=(256, 256),
                  lstm_2_init_period=50,
                  linear_layer_ref=noisy_linear,
@@ -68,7 +68,7 @@ def __init__(self,
             ob_space:           dictionary of observation state shapes
             ac_space:           discrete action space shape (length)
             rp_sequence_size:   reward prediction sample length
-            lstm_class_ref:     tf.nn.lstm class to use
+            lstm_class_ref:     tfa.rnn class to use
             lstm_layers:        tuple of LSTM layers sizes
             lstm_2_init_period: number of `get_initial_context()` method calls before force LSTM_2 context reset.
             linear_layer_ref:   linear layer class to use
diff --git a/btgym/research/encoder_test/policy.py b/btgym/research/encoder_test/policy.py
index e5fb05d0..92de41c6 100644
--- a/btgym/research/encoder_test/policy.py
+++ b/btgym/research/encoder_test/policy.py
@@ -1,3 +1,5 @@
+import tensorflow_addons as tfa
+
 from btgym.algorithms.policy.base import BaseAacPolicy
 from btgym.algorithms.nn.networks import *
 from btgym.algorithms.utils import *
@@ -17,7 +19,7 @@ def __init__(self,
                  ac_space,
                  rp_sequence_size=4,
                  state_encoder_class_ref=conv_2d_network,
-                 lstm_class_ref=tf.contrib.rnn.LayerNormBasicLSTMCell,
+                 lstm_class_ref=tfa.rnn.LayerNormLSTMCell,
                  lstm_layers=(256, 256),
                  linear_layer_ref=noisy_linear,
                  share_encoder_params=False,
@@ -39,7 +41,7 @@ def __init__(self,
             ob_space:               instance of btgym.spaces.DictSpace
             ac_space:               instance of btgym.spaces.ActionDictSpace
             rp_sequence_size:       reward prediction sample length
-            lstm_class_ref:         tf.nn.lstm class to use
+            lstm_class_ref:         tfa.rnn class to use
             lstm_layers:            tuple of LSTM layers sizes
             linear_layer_ref:       linear layer class to use
             share_encoder_params:   bool, whether to share encoder parameters for every 'external' data stream
diff --git a/btgym/research/gps/policy.py b/btgym/research/gps/policy.py
index ca271506..e90dd2fa 100644
--- a/btgym/research/gps/policy.py
+++ b/btgym/research/gps/policy.py
@@ -1,4 +1,4 @@
-import tensorflow as tf
+import tensorflow_addons as tfa
 from btgym.algorithms.policy.stacked_lstm import AacStackedRL2Policy
 from btgym.algorithms.nn.layers import noisy_linear
 
@@ -16,7 +16,7 @@ def __init__(
                 (64, (3, 1), (2, 1)),
                 (64, (3, 1), (2, 1))
             ),
-            lstm_class_ref=tf.contrib.rnn.LayerNormBasicLSTMCell,
+            lstm_class_ref=tfa.rnn.LayerNormLSTMCell,
             lstm_layers=(256, 256),
             lstm_2_init_period=50,
             linear_layer_ref=noisy_linear,
diff --git a/setup.py b/setup.py
index f627e0c5..3c1e2715 100644
--- a/setup.py
+++ b/setup.py
@@ -53,6 +53,7 @@
     version='0.0.8',
     install_requires=[
         'tensorflow>=2.3',
+        'tensorflow_addons>=0.11',
         'opencv-python',
         'gym[atari]',
         'backtrader',

From 1032bc1085f8322de4ab7bca5649b972937ba595 Mon Sep 17 00:00:00 2001
From: Wojciech Indyk <wojciechindyk@gmail.com>
Date: Fri, 28 Aug 2020 21:36:56 +0200
Subject: [PATCH 3/6] Tf1 -> Tf2: layer_norm to tf.keras.layers, LuongAttention

---
 btgym/algorithms/nn/ae.py               | 2 +-
 btgym/algorithms/nn/networks.py         | 2 +-
 btgym/research/casual_conv/networks.py  | 9 +++++----
 btgym/research/encoder_test/networks.py | 2 +-
 4 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/btgym/algorithms/nn/ae.py b/btgym/algorithms/nn/ae.py
index f88b60b9..3c36d368 100644
--- a/btgym/algorithms/nn/ae.py
+++ b/btgym/algorithms/nn/ae.py
@@ -1,7 +1,6 @@
 import numpy as np
 import tensorflow as tf
 from tensorflow.contrib.layers import flatten as batch_flatten
-from tensorflow.contrib.layers import layer_norm as norm_layer
 
 from btgym.algorithms.nn.layers import normalized_columns_initializer, linear, conv2d
 
@@ -35,6 +34,7 @@ def conv2d_encoder(x,
         layer_shapes = [x.get_shape()]
         layer_outputs = []
         for i, layer_spec in enumerate(layer_config, 1):
+            norm_layer = tf.keras.layers.LayerNormalization()
             x = tf.nn.elu(
                 norm_layer(
                     conv2d(
diff --git a/btgym/algorithms/nn/networks.py b/btgym/algorithms/nn/networks.py
index 2d998db9..03007e09 100644
--- a/btgym/algorithms/nn/networks.py
+++ b/btgym/algorithms/nn/networks.py
@@ -7,7 +7,6 @@
 import numpy as np
 import tensorflow as tf
 import tensorflow.contrib.rnn as rnn
-from tensorflow.contrib.layers import layer_norm as norm_layer
 from tensorflow.python.util.nest import flatten as flatten_nested
 
 from btgym.algorithms.nn.layers import normalized_columns_initializer, categorical_sample
@@ -38,6 +37,7 @@ def conv_2d_network(x,
     """
     with tf.compat.v1.variable_scope(name, reuse=reuse):
         for i, num_filters in enumerate(conv_2d_num_filters):
+            norm_layer = tf.keras.layers.LayerNormalization()
             x = tf.nn.elu(
                 norm_layer(
                     conv_2d_layer_ref(
diff --git a/btgym/research/casual_conv/networks.py b/btgym/research/casual_conv/networks.py
index ac48c641..e8e1f6e3 100644
--- a/btgym/research/casual_conv/networks.py
+++ b/btgym/research/casual_conv/networks.py
@@ -1,6 +1,6 @@
 
 import tensorflow as tf
-from tensorflow.contrib.layers import layer_norm as norm_layer
+import tensorflow_addons as tfa
 
 import numpy as np
 import math
@@ -78,7 +78,8 @@ def conv_1d_casual_encoder(
             # b2t:
             y = tf.reshape(y, [-1, num_time_batches, conv_1d_num_filters], name='layer_{}_output'.format(i))
 
-            y = norm_layer(y)
+            normalization_layer = tf.keras.layers.LayerNormalization()
+            y = normalization_layer(y)
             if conv_1d_activation is not None:
                 y = conv_1d_activation(y)
 
@@ -137,7 +138,7 @@ def conv_1d_casual_encoder(
     return encoded
 
 
-def attention_layer(inputs, attention_ref=tf.contrib.seq2seq.LuongAttention, name='attention_layer', **kwargs):
+def attention_layer(inputs, attention_ref=tfa.seq2seq.LuongAttention, name='attention_layer', **kwargs):
     """
     Temporal attention layer.
     Computes attention context based on last(left) value in time dim.
@@ -201,7 +202,7 @@ def conv_1d_casual_attention_encoder(
         conv_1d_num_filters=32,
         conv_1d_filter_size=2,
         conv_1d_activation=tf.nn.elu,
-        conv_1d_attention_ref=tf.contrib.seq2seq.LuongAttention,
+        conv_1d_attention_ref=tfa.seq2seq.LuongAttention,
         name='casual_encoder',
         keep_prob=None,
         conv_1d_gated=False,
diff --git a/btgym/research/encoder_test/networks.py b/btgym/research/encoder_test/networks.py
index 24852510..7938a1c3 100644
--- a/btgym/research/encoder_test/networks.py
+++ b/btgym/research/encoder_test/networks.py
@@ -1,7 +1,6 @@
 import numpy as np
 import tensorflow as tf
 import tensorflow.contrib.rnn as rnn
-from tensorflow.contrib.layers import layer_norm as norm_layer
 from tensorflow.python.util.nest import flatten as flatten_nested
 
 from btgym.algorithms.nn.layers import normalized_columns_initializer, categorical_sample
@@ -35,6 +34,7 @@ def conv_2d_network_skip(x,
     layers = []
     with tf.compat.v1.variable_scope(name, reuse=reuse):
         for i, num_filters in enumerate(conv_2d_num_filters):
+            norm_layer = tf.keras.layers.LayerNormalization()
             x = tf.nn.elu(
                 norm_layer(
                     conv_2d_layer_ref(

From d591bea1358c7e2115c41b95baa84e012574f607 Mon Sep 17 00:00:00 2001
From: Wojciech Indyk <wojciechindyk@gmail.com>
Date: Sat, 29 Aug 2020 05:54:56 +0200
Subject: [PATCH 4/6] Tf1 -> Tf2: migrate tf.contrib

---
 btgym/algorithms/nn/ae.py               |  3 +--
 btgym/algorithms/nn/networks.py         |  7 ++++---
 btgym/algorithms/rollout.py             |  2 +-
 btgym/algorithms/utils.py               | 15 +++++++--------
 btgym/research/encoder_test/networks.py |  1 -
 5 files changed, 13 insertions(+), 15 deletions(-)

diff --git a/btgym/algorithms/nn/ae.py b/btgym/algorithms/nn/ae.py
index 3c36d368..6bb2e5dd 100644
--- a/btgym/algorithms/nn/ae.py
+++ b/btgym/algorithms/nn/ae.py
@@ -1,6 +1,5 @@
 import numpy as np
 import tensorflow as tf
-from tensorflow.contrib.layers import flatten as batch_flatten
 
 from btgym.algorithms.nn.layers import normalized_columns_initializer, linear, conv2d
 
@@ -165,7 +164,7 @@ def conv2d_autoencoder(
             reuse=reuse
         )
         # Flatten hidden state, pass through dense :
-        z = batch_flatten(encoder_layers[-1])
+        z = tf.reshape(encoder_layers[-1], [tf.shape(encoder_layers[-1])[0], -1])
         h, w, c = encoder_layers[-1].get_shape().as_list()[1:]
 
         z = linear_layer_ref(
diff --git a/btgym/algorithms/nn/networks.py b/btgym/algorithms/nn/networks.py
index 03007e09..0841ee7d 100644
--- a/btgym/algorithms/nn/networks.py
+++ b/btgym/algorithms/nn/networks.py
@@ -6,7 +6,7 @@
 
 import numpy as np
 import tensorflow as tf
-import tensorflow.contrib.rnn as rnn
+import tensorflow_addons as tfa
 from tensorflow.python.util.nest import flatten as flatten_nested
 
 from btgym.algorithms.nn.layers import normalized_columns_initializer, categorical_sample
@@ -105,7 +105,7 @@ def conv_1d_network(x,
 def lstm_network(
         x,
         lstm_sequence_length,
-        lstm_class=rnn.BasicLSTMCell,
+        lstm_class=tf.compat.v1.nn.rnn_cell.BasicLSTMCell,
         lstm_layers=(256,),
         static=False,
         keep_prob=None,
@@ -141,7 +141,7 @@ def lstm_network(
 
             lstm.append(layer)
 
-        lstm = rnn.MultiRNNCell(lstm, state_is_tuple=True)
+        lstm = tf.compat.v1.nn.rnn_cell.MultiRNNCell(lstm, state_is_tuple=True)
         # Get time_dimension as [1]-shaped tensor:
         step_size = tf.expand_dims(tf.shape(input=x)[1], [0])
 
@@ -185,6 +185,7 @@ def dense_aac_network(x, ac_space_depth, name='dense_aac', linear_layer_ref=nois
 
     with tf.compat.v1.variable_scope(name, reuse=reuse):
         # Center-logits:
+        norm_layer = tf.keras.layers.LayerNormalization()
         logits = norm_layer(
             linear_layer_ref(
                 x=x,
diff --git a/btgym/algorithms/rollout.py b/btgym/algorithms/rollout.py
index 1eea6028..4b9282ba 100644
--- a/btgym/algorithms/rollout.py
+++ b/btgym/algorithms/rollout.py
@@ -8,7 +8,7 @@
 
 import numpy as np
 
-from tensorflow.contrib.rnn import LSTMStateTuple
+from tensorflow._api.v2.compat.v1.nn.rnn_cell import LSTMStateTuple
 from btgym.algorithms.math_utils import discount
 from btgym.algorithms.utils import batch_pad
 
diff --git a/btgym/algorithms/utils.py b/btgym/algorithms/utils.py
index da323aea..82c61645 100644
--- a/btgym/algorithms/utils.py
+++ b/btgym/algorithms/utils.py
@@ -3,7 +3,6 @@
 import tensorflow as tf
 from tensorflow.python.util.nest import flatten as flatten_nested
 from tensorflow.python.util.nest import assert_same_structure
-from tensorflow.contrib.rnn import LSTMStateTuple
 
 from gym.spaces import Discrete, Dict
 
@@ -19,11 +18,11 @@ def rnn_placeholders(state):
 
     Returns:    tuple of placeholders
     """
-    if isinstance(state, tf.nn.rnn_cell.LSTMStateTuple):
+    if isinstance(state, tf.compat.v1.nn.rnn_cell.LSTMStateTuple):
         c, h = state
         c = tf.compat.v1.placeholder(tf.float32, tf.TensorShape([None]).concatenate(c.get_shape()[1:]), c.op.name + '_c_pl')
         h = tf.compat.v1.placeholder(tf.float32, tf.TensorShape([None]).concatenate(h.get_shape()[1:]), h.op.name + '_h_pl')
-        return tf.nn.rnn_cell.LSTMStateTuple(c, h)
+        return tf.compat.v1.nn.rnn_cell.LSTMStateTuple(c, h)
     elif isinstance(state, tf.Tensor):
         h = state
         h = tf.compat.v1.placeholder(tf.float32, tf.TensorShape([None]).concatenate(h.get_shape()[1:]), h.op.name + '_h_pl')
@@ -180,10 +179,10 @@ def batch_stack(dict_list, _top=True):
             value_list = [value[key] for value in dict_list]
             batch[key] = batch_stack(value_list, False)
 
-    elif isinstance(master, LSTMStateTuple):
+    elif isinstance(master, tf.compat.v1.nn.rnn_cell.LSTMStateTuple):
         c = batch_stack([state[0] for state in dict_list], False)
         h = batch_stack([state[1] for state in dict_list], False)
-        batch = LSTMStateTuple(c=c, h=h)
+        batch = tf.compat.v1.nn.rnn_cell.LSTMStateTuple(c=c, h=h)
 
     elif isinstance(master, tuple):
         batch = tuple([batch_stack([struct[i] for struct in dict_list], False) for i in range(len(master))])
@@ -220,10 +219,10 @@ def batch_gather(batch_dict, indices, _top=True):
         for key, value in batch_dict.items():
             batch[key] = batch_gather(value, indices, False)
 
-    elif isinstance(batch_dict, LSTMStateTuple):
+    elif isinstance(batch_dict, tf.compat.v1.nn.rnn_cell.LSTMStateTuple):
         c = batch_gather(batch_dict[0], indices, False)
         h = batch_gather(batch_dict[1], indices, False)
-        batch = LSTMStateTuple(c=c, h=h)
+        batch = tf.compat.v1.nn.rnn_cell.LSTMStateTuple(c=c, h=h)
 
     elif isinstance(batch_dict, tuple):
         batch = tuple([batch_gather(struct, indices, False) for struct in batch_dict])
@@ -312,7 +311,7 @@ def _show_struct(struct):
             print(key)
             _show_struct(value)
 
-    elif type(struct) in [LSTMStateTuple, tuple, list]:
+    elif type(struct) in [tf.compat.v1.nn.rnn_cell.LSTMStateTuple, tuple, list]:
         print('LSTM/tuple/list:', type(struct), len(struct))
         for i in struct:
             _show_struct(i)
diff --git a/btgym/research/encoder_test/networks.py b/btgym/research/encoder_test/networks.py
index 7938a1c3..abaae2c0 100644
--- a/btgym/research/encoder_test/networks.py
+++ b/btgym/research/encoder_test/networks.py
@@ -1,6 +1,5 @@
 import numpy as np
 import tensorflow as tf
-import tensorflow.contrib.rnn as rnn
 from tensorflow.python.util.nest import flatten as flatten_nested
 
 from btgym.algorithms.nn.layers import normalized_columns_initializer, categorical_sample

From 640fb55a77a5126a8b8d71600fbbe20aaeb3d990 Mon Sep 17 00:00:00 2001
From: Wojciech Indyk <wojciechindyk@gmail.com>
Date: Sun, 27 Sep 2020 14:21:54 +0200
Subject: [PATCH 5/6] Tf1 -> Tf2: LSTMCell

---
 btgym/algorithms/policy/base.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/btgym/algorithms/policy/base.py b/btgym/algorithms/policy/base.py
index a7dd4b93..3c0dcece 100644
--- a/btgym/algorithms/policy/base.py
+++ b/btgym/algorithms/policy/base.py
@@ -10,6 +10,7 @@
 #
 
 from gym.spaces import Discrete, Dict
+from tensorflow._api.v2.compat.v1.nn.rnn_cell import BasicLSTMCell
 
 from btgym.algorithms.nn.networks import *
 from btgym.algorithms.utils import *
@@ -34,7 +35,7 @@ def __init__(self,
                  ob_space,
                  ac_space,
                  rp_sequence_size,
-                 lstm_class=rnn.BasicLSTMCell,
+                 lstm_class=BasicLSTMCell,
                  lstm_layers=(256,),
                  action_dp_alpha=200.0,
                  aux_estimate=False,
@@ -405,7 +406,7 @@ def __init__(self,
                  ob_space,
                  ac_space,
                  rp_sequence_size,
-                 lstm_class=rnn.BasicLSTMCell,
+                 lstm_class=BasicLSTMCell,
                  lstm_layers=(256,),
                  action_dp_alpha=200.0,
                  aux_estimate=True,

From 57efd45c77f55ccd0b943ee78535050c01ee377a Mon Sep 17 00:00:00 2001
From: Wojciech Indyk <wojciechindyk@gmail.com>
Date: Sun, 27 Sep 2020 16:57:35 +0200
Subject: [PATCH 6/6] Tf1 -> Tf2: MLDG

---
 btgym/research/metalearn_2/_mldg_batch.py | 531 ++++++++++++++++++++++
 1 file changed, 531 insertions(+)

diff --git a/btgym/research/metalearn_2/_mldg_batch.py b/btgym/research/metalearn_2/_mldg_batch.py
index e69de29b..3a22701a 100644
--- a/btgym/research/metalearn_2/_mldg_batch.py
+++ b/btgym/research/metalearn_2/_mldg_batch.py
@@ -0,0 +1,531 @@
+import tensorflow as tf
+import numpy as np
+
+import sys
+from logbook import Logger, StreamHandler
+
+from btgym.research.mldg.aac import SubAAC
+from btgym.algorithms.runner.synchro import BaseSynchroRunner
+
+
+class MLDG():
+    """
+    Asynchronous implementation of MLDG algorithm
+    for continuous adaptation in dynamically changing environments.
+
+    Papers:
+        Da Li et al.,
+         "Learning to Generalize: Meta-Learning for Domain Generalization"
+         https://arxiv.org/abs/1710.03463
+
+        Maruan Al-Shedivat et al.,
+        "Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments"
+        https://arxiv.org/abs/1710.03641
+
+    """
+    def __init__(
+            self,
+            env,
+            task,
+            log_level,
+            aac_class_ref=SubAAC,
+            runner_config=None,
+            aac_lambda=1.0,
+            guided_lambda=1.0,
+            rollout_length=20,
+            train_support=300,
+            fast_adapt_num_steps=10,
+            fast_adapt_batch_size=32,
+            trial_source_target_cycle=(1, 0),
+            num_episodes_per_trial=1,  # one-shot adaptation
+            _aux_render_modes=('action_prob', 'value_fn', 'lstm_1_h', 'lstm_2_h'),
+            name='MLDG',
+            **kwargs
+    ):
+        try:
+            self.aac_class_ref = aac_class_ref
+            self.task = task
+            self.name = name
+            self.summary_writer = None
+
+            StreamHandler(sys.stdout).push_application()
+            self.log = Logger('{}_{}'.format(name, task), level=log_level)
+
+            self.rollout_length = rollout_length
+            self.train_support = train_support  # number of train experiences to collect
+            self.train_batch_size = int(self.train_support / self.rollout_length)
+            self.fast_adapt_num_steps = fast_adapt_num_steps
+            self.fast_adapt_batch_size = fast_adapt_batch_size
+
+            if runner_config is None:
+                self.runner_config = {
+                    'class_ref': BaseSynchroRunner,
+                    'kwargs': {},
+                }
+            else:
+                self.runner_config = runner_config
+
+            self.env_list = env
+
+            assert isinstance(self.env_list, list) and len(self.env_list) == 2, \
+                'Expected pair of environments, got: {}'.format(self.env_list)
+
+            # Instantiate two sub-trainers: one for test and one for train environments:
+
+            self.runner_config['kwargs']['data_sample_config'] = {'mode': 1}  # master
+            self.runner_config['kwargs']['name'] = 'master'
+
+            self.train_aac = aac_class_ref(
+                env=self.env_list[0],  # train data will be master environment TODO: really dumb data control. improve.
+                task=self.task,
+                log_level=log_level,
+                runner_config=self.runner_config,
+                aac_lambda=aac_lambda,
+                guided_lambda=guided_lambda,
+                rollout_length=self.rollout_length,
+                trial_source_target_cycle=trial_source_target_cycle,
+                num_episodes_per_trial=num_episodes_per_trial,
+                _use_target_policy=False,
+                _use_global_network=True,
+                _aux_render_modes=_aux_render_modes,
+                name=self.name + '_sub_Train',
+                **kwargs
+            )
+
+            self.runner_config['kwargs']['data_sample_config'] = {'mode': 0}  # master
+            self.runner_config['kwargs']['name'] = 'slave'
+
+            self.test_aac = aac_class_ref(
+                env=self.env_list[-1],  # test data -> slave env.
+                task=self.task,
+                log_level=log_level,
+                runner_config=self.runner_config,
+                aac_lambda=aac_lambda,
+                guided_lambda=guided_lambda,
+                rollout_length=self.rollout_length,
+                trial_source_target_cycle=trial_source_target_cycle,
+                num_episodes_per_trial=num_episodes_per_trial,
+                _use_target_policy=False,
+                _use_global_network=False,
+                global_step_op=self.train_aac.global_step,
+                global_episode_op=self.train_aac.global_episode,
+                inc_episode_op=self.train_aac.inc_episode,
+                _aux_render_modes=_aux_render_modes,
+                name=self.name + '_sub_Test',
+                **kwargs
+            )
+
+            self.local_steps = self.train_aac.local_steps
+            self.model_summary_freq = self.train_aac.model_summary_freq
+            #self.model_summary_op = self.train_aac.model_summary_op
+
+            self._make_train_op()
+            self.test_aac.model_summary_op = tf.summary.merge(
+                [self.test_aac.model_summary_op, self._combine_meta_summaries()],
+                name='meta_model_summary'
+            )
+
+        except:
+            msg = 'MLDG.__init()__ exception occurred' + \
+                  '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
+            self.log.exception(msg)
+            raise RuntimeError(msg)
+
+    def _make_train_op(self):
+        """
+
+        Defines:
+            tensors holding training op graph for sub trainers and self;
+        """
+        pi = self.train_aac.local_network
+        pi_prime = self.test_aac.local_network
+
+        self.test_aac.sync = self.test_aac.sync_pi = tf.group(
+            *[v1.assign(v2) for v1, v2 in zip(pi_prime.var_list, pi.var_list)]
+        )
+
+        self.global_step = self.train_aac.global_step
+        self.global_episode = self.train_aac.global_episode
+
+        self.test_aac.global_step = self.train_aac.global_step
+        self.test_aac.global_episode = self.train_aac.global_episode
+        self.test_aac.inc_episode = self.train_aac.inc_episode
+        self.train_aac.inc_episode = None
+        self.inc_step = self.train_aac.inc_step
+
+        # Meta-loss:
+        self.loss = 0.5 * self.train_aac.loss + 0.5 * self.test_aac.loss
+
+        # Clipped gradients:
+        self.train_aac.grads, _ = tf.clip_by_global_norm(
+            tf.gradients(self.train_aac.loss, pi.var_list),
+            40.0
+        )
+        self.log.warning('self.train_aac.grads: {}'.format(len(list(self.train_aac.grads))))
+
+        # self.test_aac.grads, _ = tf.clip_by_global_norm(
+        #     tf.gradients(self.test_aac.loss, pi_prime.var_list),
+        #     40.0
+        # )
+        # Meta-gradient:
+        grads_i, _ = tf.clip_by_global_norm(
+            tf.gradients(self.train_aac.loss, pi.var_list),
+            40.0
+        )
+
+        grads_i_next, _ = tf.clip_by_global_norm(
+            tf.gradients(self.test_aac.loss, pi_prime.var_list),
+            40.0
+        )
+
+        self.grads = []
+        for g1, g2 in zip(grads_i, grads_i_next):
+            if g1 is not None and g2 is not None:
+                meta_g = 0.5 * g1 + 0.5 * g2
+            else:
+                meta_g = None
+
+            self.grads.append(meta_g)
+
+        #self.log.warning('self.grads_len: {}'.format(len(list(self.grads))))
+
+        # Gradients to update local copy of pi_prime (from train data):
+        train_grads_and_vars = list(zip(self.train_aac.grads, pi_prime.var_list))
+
+        # self.log.warning('train_grads_and_vars_len: {}'.format(len(train_grads_and_vars)))
+
+        # Meta-gradients to be sent to parameter server:
+        meta_grads_and_vars = list(zip(self.grads, self.train_aac.network.var_list))
+
+        # self.log.warning('meta_grads_and_vars_len: {}'.format(len(meta_grads_and_vars)))
+
+        # Set global_step increment equal to observation space batch size:
+        obs_space_keys = list(self.train_aac.local_network.on_state_in.keys())
+
+        assert 'external' in obs_space_keys, \
+            'Expected observation space to contain `external` mode, got: {}'.format(obs_space_keys)
+        self.train_aac.inc_step = self.train_aac.global_step.assign_add(
+            tf.shape(self.train_aac.local_network.on_state_in['external'])[0]
+        )
+
+        self.train_op = self.train_aac.optimizer.apply_gradients(train_grads_and_vars)
+
+        # Optimizer for meta-update:
+        self.optimizer = tf.train.AdamOptimizer(self.train_aac.train_learn_rate, epsilon=1e-5)
+        # TODO: own alpha-leran rate
+        self.meta_train_op = self.optimizer.apply_gradients(meta_grads_and_vars)
+
+        self.log.debug('meta_train_op defined')
+
+    def _combine_meta_summaries(self):
+
+        meta_model_summaries = [
+            tf.summary.scalar("meta_grad_global_norm", tf.global_norm(self.grads)),
+            tf.summary.scalar("total_meta_loss", self.loss),
+        ]
+
+        return meta_model_summaries
+
+    def start(self, sess, summary_writer, **kwargs):
+        """
+        Executes all initializing operations,
+        starts environment runner[s].
+        Supposed to be called by parent worker just before training loop starts.
+
+        Args:
+            sess:           tf session object.
+            kwargs:         not used by default.
+        """
+        try:
+            # Copy weights from global to local:
+            sess.run(self.train_aac.sync_pi)
+            sess.run(self.test_aac.sync_pi)
+
+            # Start thread_runners:
+            self.train_aac._start_runners(   # master first
+                sess,
+                summary_writer,
+                init_context=None,
+                data_sample_config=self.train_aac.get_sample_config(mode=1)
+            )
+            self.test_aac._start_runners(
+                sess,
+                summary_writer,
+                init_context=None,
+                data_sample_config=self.test_aac.get_sample_config(mode=0)
+            )
+
+            self.summary_writer = summary_writer
+            self.log.notice('Runners started.')
+
+        except:
+            msg = 'start() exception occurred' + \
+                  '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
+            self.log.exception(msg)
+            raise RuntimeError(msg)
+
+    def fast_adapt_step(self, sess, batch_size, on_policy_batch, off_policy_batch, rp_batch, make_summary=False):
+        """
+        One step of test_policy adaptation.
+
+        Args:
+            sess:                   tensorflow.Session obj.
+            batch_size:             train mini-batch size
+            on_policy_batch:        `on_policy` train data
+            off_policy_batch:       `off_policy` train data or None
+            rp_batch:               'reward_prediction` train data or None
+            make_summary:           bool, if True - compute model summary
+
+        Returns:
+            model summary or None
+        """
+        # Sample from train distribution:
+        on_mini_batch = self.train_aac.sample_batch(on_policy_batch, batch_size)
+        off_mini_batch = self.train_aac.sample_batch(off_policy_batch, batch_size)
+        rp_mini_batch = self.train_aac.sample_batch(rp_batch, batch_size)
+
+        feed_dict = self.train_aac._get_main_feeder(sess, on_mini_batch, off_mini_batch, rp_mini_batch, True)
+
+        if make_summary:
+            fetches = [self.train_op, self.train_aac.model_summary_op]
+        else:
+            fetches = [self.train_op]
+
+        # Update pi_prime parameters wrt sampled data:
+        fetched = sess.run(fetches, feed_dict=feed_dict)
+
+        # self.log.warning('Train gradients ok.')
+
+        if make_summary:
+            summary =  fetched[-1]
+
+        else:
+            summary = None
+
+        return summary
+
+    def train_step(self, sess, data_config):
+        """
+        Collects train task data and updates test policy parameters (fast adaptation).
+
+        Args:
+            sess:                   tensorflow.Session obj.
+            data_config:            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
+
+        Returns:
+            batched train data
+
+        """
+        # Collect train distribution:
+        train_batch = self.train_aac.get_batch(
+            size=self.train_batch_size,
+            require_terminal=True,
+            same_trial=True,
+            data_sample_config=data_config
+        )
+
+        # for rollout in train_batch['on_policy']:
+        #     self.log.warning(
+        #         'Train data trial_num: {}'.format(
+        #             np.asarray(rollout['state']['metadata']['trial_num'])
+        #         )
+        #     )
+
+        # Process time-flat-alike (~iid) to treat as empirical data distribution over train task:
+        on_policy_batch, off_policy_batch, rp_batch = self.train_aac.process_batch(sess, train_batch)
+
+        # self.log.warning('Train data ok.')
+
+        local_step = sess.run(self.global_step)
+        local_episode = sess.run(self.global_episode)
+        model_summary = None
+
+        # Extract all non-empty summaries:
+        ep_summary = [summary for summary in train_batch['ep_summary'] if summary is not None]
+
+        # Perform number of test policy updates wrt. collected train data:
+        for i in range(self.fast_adapt_num_steps):
+            model_summary = self.fast_adapt_step(
+                sess,
+                batch_size=self.fast_adapt_batch_size,
+                on_policy_batch=on_policy_batch,
+                off_policy_batch=off_policy_batch,
+                rp_batch=rp_batch,
+                make_summary=(local_step + i) % self.model_summary_freq == 0
+            )
+            # self.log.warning('Batch {} Train gradients ok.'.format(i))
+
+            # Write down summaries:
+            train_summary = dict(
+                render_summary=[None],
+                test_ep_summary=[None],
+                ep_summary=[ep_summary.pop() if len(ep_summary) > 0 else None]
+            )
+            self.train_aac.process_summary(
+                sess,
+                train_summary,
+                model_summary,
+                step=local_step + i,
+                episode=local_episode + i
+            )
+
+        return on_policy_batch, off_policy_batch, rp_batch
+
+    def meta_train_step(self, sess, data_config, on_policy_batch, off_policy_batch, rp_batch):
+        """
+        Collects data from source domain test task and performs meta-update to shared parameters vector.
+        Writes down relevant summaries.
+
+        Args:
+            sess:                   tensorflow.Session obj.
+            data_config:            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
+            on_policy_batch:        `on_policy` train data
+            off_policy_batch:       `off_policy` train data or None
+            rp_batch:               'reward_prediction` train data or None
+
+        """
+        done = False
+        while not done:
+            # Say `No` to redundant summaries:
+            wirte_model_summary = \
+                self.local_steps % self.model_summary_freq == 0
+
+            # Collect test trajectory wrt updated test_policy parameters:
+            test_data = self.test_aac.get_data(
+                init_context=None,
+                data_sample_config=data_config
+            )
+            test_batch_size = 0  # TODO: adjust on/off/rp sizes
+            for rollout in test_data['on_policy']:
+                test_batch_size += len(rollout['position'])
+
+            test_feed_dict = self.test_aac.process_data(sess,,,,, test_data,,
+
+            # self.log.warning('Test data rollout for step {} ok.'.format(self.local_steps))
+            #
+            # self.log.warning(
+            #     'Test data trial_num: {}'.format(
+            #         np.asarray(test_data['on_policy'][0]['state']['metadata']['trial_num'])
+            #     )
+            # )
+
+            # Sample train data of same size:
+            feed_dict = self.train_aac._get_main_feeder(
+                sess,
+                self.train_aac.sample_batch(on_policy_batch, test_batch_size),
+                self.train_aac.sample_batch(off_policy_batch, test_batch_size),
+                self.train_aac.sample_batch(rp_batch, test_batch_size),
+                True
+            )
+            # Add test trajectory:
+            feed_dict.update(test_feed_dict)
+
+            # Perform meta-update:
+            if wirte_model_summary:
+                meta_fetches = [self.meta_train_op, self.test_aac.model_summary_op, self.inc_step]
+            else:
+                meta_fetches = [self.meta_train_op, self.inc_step]
+
+            meta_fetched = sess.run(meta_fetches, feed_dict=feed_dict)
+
+            # self.log.warning('Meta-gradients ok.')
+
+            if wirte_model_summary:
+                meta_model_summary = meta_fetched[-2]
+
+            else:
+                meta_model_summary = None
+
+            # Write down summaries:
+            self.test_aac.process_summary(sess, test_data, meta_model_summary)
+            self.local_steps += 1
+
+            # If test episode ended?
+            done = np.asarray(test_data['terminal']).any()
+
+    def meta_test_step(self, sess, data_config, on_policy_batch, off_policy_batch, rp_batch):
+        """
+        Validates adapted policy on data from target domain test task.
+        Writes down relevant summaries.
+
+        Args:
+            sess:                   tensorflow.Session obj.
+            data_config:            configuration dictionary of type `btgym.datafeed.base.EnvResetConfig`
+            on_policy_batch:        `on_policy` train data
+            off_policy_batch:       `off_policy` train data or None
+            rp_batch:               'reward_prediction` train data or None
+
+        """
+        done = False
+        while not done:
+            # Collect test trajectory:
+            test_data = self.test_aac.get_data(
+                init_context=None,
+                data_sample_config=data_config
+            )
+
+            # self.log.warning('Target test rollout ok.')
+            # self.log.warning(
+            #     'Test data target trial_num: {}'.format(
+            #         np.asarray(test_data['on_policy'][0]['state']['metadata']['trial_num'])
+            #     )
+            # )
+            # self.log.warning('target_render_ep_summary: {}'.format(test_data['render_summary']))
+
+            # Write down summaries:
+            self.test_aac.process_summary(sess, test_data)
+
+            # If test episode ended?
+            done = np.asarray(test_data['terminal']).any()
+
+    def process(self, sess):
+        """
+        Meta-train procedure for one-shot learning/
+
+        Args:
+            sess (tensorflow.Session):   tf session obj.
+
+        """
+        try:
+            # Copy from parameter server:
+            sess.run(self.train_aac.sync_pi)
+            sess.run(self.test_aac.sync_pi)
+
+            #self.log.warning('Sync ok.')
+
+            # Decide on data configuration for train/test trajectories,
+            # such as all data will come from same trial (maybe different episodes)
+            # and trial type as well (~from source or target domain):
+            # note: data_config counters get updated once per process() call
+            train_data_config = self.train_aac.get_sample_config(mode=1)  # master env., draws trial
+            test_data_config = self.train_aac.get_sample_config(mode=0)   # slave env, catches up with same trial
+
+            # If data comes from source or target domain:
+            is_target = train_data_config['trial_config']['sample_type']
+
+            # self.log.warning('PROCESS_train_data_config: {}'.format(train_data_config))
+            # self.log.warning('PROCESS_test_data_config: {}'.format(test_data_config))
+
+            # Fast adaptation step:
+            # collect train trajectories, process time-flat-alike (~iid) to treat as empirical data distribution
+            # over train task and adapt test_policy wrt. train experience:
+            on_policy_batch, off_policy_batch, rp_batch = self.train_step(sess, train_data_config)
+
+            # Slow adaptation step:
+            if is_target:
+                # Meta-test:
+                # self.log.warning('Running meta-test episode...')
+                self.meta_test_step(sess,test_data_config, on_policy_batch, off_policy_batch, rp_batch)
+
+            else:
+                # Meta-train:
+                # self.log.warning('Running meta-train episode...')
+                self.meta_train_step(sess,test_data_config, on_policy_batch, off_policy_batch, rp_batch)
+
+        except:
+            msg = 'process() exception occurred' + \
+                  '\n\nPress `Ctrl-C` or jupyter:[Kernel]->[Interrupt] for clean exit.\n'
+            self.log.exception(msg)
+            raise RuntimeError(msg)
+
+