-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathRecurrentCNN.py
437 lines (339 loc) · 17.4 KB
/
RecurrentCNN.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
import tensorflow as tf
import numpy as np
import os
from models import Config, Model
import math
XAVIER_INIT = tf.contrib.layers.xavier_initializer
class RecurrentCNNConfig(Config):
def __init__(self):
self.batch_size = 64
self.lr = 1e-3
self.l2_lambda = 0.0000001
self.hidden_size = 512
self.num_epochs = 50
self.num_layers = 3
self.num_classes = 4 # Mean vector of size 4
self.features_shape = (100,100,3) #TO FIX!!!!
self.targets_shape = (4,)
self.init_loc_size = (4,)
self.max_norm = 10
self.keep_prob = 0.8
self.init_state_out_size = 128
self.cnn_out_shape = 128
self.variance = 1e-1
self.num_samples = 5
class RecurrentCNN(Model):
def __init__(self, features_shape, num_classes, cell_type='lstm', seq_len=8, reuse=False, add_bn=False,
add_reg=False, deeper = False, loss_type = 'negative_l1_dist', cum_sum=False, scope='RCNN'):
self.config = RecurrentCNNConfig()
self.config.features_shape = features_shape
self.config.num_classes = num_classes
self.reuse = reuse
self.inputs_placeholder = tf.placeholder(tf.float32, shape=tuple((None,None,)+ self.config.features_shape ))
self.init_loc = tf.placeholder(tf.float32, shape=tuple((None,)+ self.config.init_loc_size))
self.targets_placeholder = tf.placeholder(tf.float32, shape=tuple((None,None,) + self.config.targets_shape))
self.config.seq_len = seq_len
self.seq_len_placeholder = tf.placeholder(tf.int32, shape=tuple((None,) ))
self.deeper = deeper
self.loss_type = loss_type
self.cumsum = cum_sum
self.scope = scope
if add_bn:
self.norm_fn = tf.contrib.layers.batch_norm
else:
self.norm_fn = None
if add_reg:
self.reg_fn = tf.nn.l2_loss
else:
self.reg_fn = None
if cell_type == 'rnn':
self.cell = tf.contrib.rnn.RNNCell
elif cell_type == 'gru':
self.cell = tf.contrib.rnn.GRUCell
elif cell_type == 'lstm':
self.cell = tf.contrib.rnn.LSTMCell
else:
raise ValueError('Input correct cell type')
def conv_layer(self, inputs, outputs, kernel_size, stride, reuse, scope):
return tf.contrib.layers.conv2d(inputs=inputs, num_outputs=outputs, kernel_size=kernel_size,
stride=stride,padding='SAME',rate=1,activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse, scope=scope, trainable=True)
def build_cnn(self, cur_inputs, reuse=False, scope=None):
with tf.variable_scope(scope):
conv_out1 = self.conv_layer(cur_inputs, 32, [3,3], [1,1], reuse, 'conv1')
conv_out2 = self.conv_layer(conv_out1, 32, [3,3], [1,1], reuse, 'conv2')
max_pool1 = tf.contrib.layers.max_pool2d(inputs=conv_out2, kernel_size=[3,3],stride=[2,2],
scope='maxpool1', padding='SAME')
conv_out3 = self.conv_layer(max_pool1, 32, [3,3], [1,1], reuse, 'conv3')
conv_out4 = self.conv_layer(conv_out3, 32, [3,3], [1,1], reuse, 'conv4')
max_pool2 = tf.contrib.layers.max_pool2d(inputs=conv_out4, kernel_size=[3,3],stride=[2,2],
scope='maxpool1',padding='SAME')
flatten_out = tf.contrib.layers.flatten(max_pool2,scope='flatten')
fc1 = tf.contrib.layers.fully_connected(inputs=flatten_out, num_outputs=self.config.cnn_out_shape,activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse,scope='fc1',trainable=True)
fc2 = tf.contrib.layers.fully_connected(inputs=fc1, num_outputs=self.config.cnn_out_shape,activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse,scope='fc2',trainable=True)
cnn_out = fc2
return cnn_out
def build_deeper_cnn(self, cur_inputs, reuse=False, scope=None):
with tf.variable_scope(scope):
conv_out1 = self.conv_layer(cur_inputs, 32, [3,3], [1,1], reuse, 'conv1')
conv_out2 = self.conv_layer(conv_out1, 32, [3,3], [1,1], reuse, 'conv2')
max_pool1 = tf.contrib.layers.max_pool2d(inputs=conv_out2, kernel_size=[3,3],stride=[1,1],
scope='maxpool1', padding='SAME')
conv_out3 = self.conv_layer(max_pool1, 32, [3,3], [1,1], reuse, 'conv3')
conv_out4 = self.conv_layer(conv_out3, 32, [3,3], [1,1], reuse, 'conv4')
max_pool2 = tf.contrib.layers.max_pool2d(inputs=conv_out4, kernel_size=[3,3],stride=[2,2],
scope='maxpool2',padding='SAME')
conv_out5 = self.conv_layer(max_pool2, 32, [3,3], [1,1], reuse, 'conv5')
conv_out6 = self.conv_layer(conv_out5, 32, [3,3], [1,1], reuse, 'conv6')
max_pool3 = tf.contrib.layers.max_pool2d(inputs=conv_out6, kernel_size=[3,3],stride=[1,1],
scope='maxpool3',padding='SAME')
conv_out7 = self.conv_layer(max_pool3, 32, [3,3], [1,1], reuse, 'conv7')
conv_out8 = self.conv_layer(conv_out7, 32, [3,3], [1,1], reuse, 'conv8')
max_pool4 = tf.contrib.layers.max_pool2d(inputs=conv_out8, kernel_size=[3,3],stride=[2,2],
scope='maxpool4',padding='SAME')
flatten_out = tf.contrib.layers.flatten(max_pool4,scope='flatten')
fc1 = tf.contrib.layers.fully_connected(inputs=flatten_out, num_outputs=self.config.cnn_out_shape,activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse,scope='fc1',trainable=True)
fc2 = tf.contrib.layers.fully_connected(inputs=fc1, num_outputs=self.config.cnn_out_shape,activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse,scope='fc2',trainable=True)
cnn_out = fc2
return cnn_out
def build_rnn(self, rnn_inputs):
W = tf.get_variable("Weights", shape=[self.config.hidden_size, self.config.num_classes],
initializer=XAVIER_INIT(uniform=True))
b = tf.get_variable("Bias", shape=[self.config.num_classes])
rnnNet = tf.contrib.rnn.MultiRNNCell([self.cell(num_units = self.config.hidden_size) for _ in
range(self.config.num_layers)], state_is_tuple=True)
(rnnNet_out, rnnNet_state) = tf.nn.dynamic_rnn(cell = rnnNet, inputs=rnn_inputs,
sequence_length=self.seq_len_placeholder,dtype=tf.float32)
cur_shape = tf.shape(rnnNet_out)
rnnOut_2d = tf.reshape(rnnNet_out, [-1, cur_shape[2]])
logits_2d = tf.matmul(rnnOut_2d, W) + b
rnn_out = tf.reshape(logits_2d,[cur_shape[0], cur_shape[1], self.config.num_classes])
return rnn_out
def build_initial_state(self, loc_inputs, reuse=False, scope=None):
with tf.variable_scope(scope):
fc1 = tf.contrib.layers.fully_connected(inputs=loc_inputs, num_outputs=self.config.init_state_out_size,
activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse, scope='fc1', trainable=True)
fc2 = tf.contrib.layers.fully_connected(inputs=fc1, num_outputs=self.config.init_state_out_size,
activation_fn=tf.nn.relu,
normalizer_fn=self.norm_fn, weights_initializer=XAVIER_INIT(uniform=True) ,
weights_regularizer=self.reg_fn , biases_regularizer=self.reg_fn ,
reuse = reuse,scope='fc2',trainable=True)
init_state_out = fc2
return init_state_out
def build_model(self):
self.cnn_scope = 'CNN'
self.fc_scope = 'FC'
self.rnn_scope = 'RNN'
obs_outputs = []
reuse = False
with tf.variable_scope(self.scope):
for t in xrange(self.config.seq_len):
print("Current iteration: {0}".format(t))
x = tf.placeholder(tf.float32, shape=[None, self.config.init_state_out_size])
st_state = tf.zeros_like(x)
if t == 0:
reuse = False
st_state = self.build_initial_state(tf.zeros_like(self.init_loc), reuse, self.fc_scope)
if t > 0:
# tf.get_variable_scope().reuse_variables()
reuse = True
st_state = self.build_initial_state(tf.zeros_like(self.init_loc), reuse, self.fc_scope)
if not self.deeper:
concat_result = tf.concat([self.build_cnn(self.inputs_placeholder[:,t,:,:,:], reuse, self.cnn_scope),st_state],
axis=1)
obs_outputs.append(concat_result)
else:
concat_result = tf.concat([self.build_deeper_cnn(self.inputs_placeholder[:,t,:,:,:], reuse, self.cnn_scope),st_state],
axis=1)
obs_outputs.append(concat_result)
obs_outputs = tf.stack(obs_outputs, axis=1)
rnn_output = None
with tf.variable_scope(self.rnn_scope):
rnn_output = self.build_rnn(obs_outputs)
self.logits = tf.nn.sigmoid(rnn_output)
# self.logits = rnn_output
def get_iou_loss(self):
p_left = self.location_samples[:, :, :, 1]
g_left = self.targets_placeholder[:, :, 1]
left = tf.maximum(p_left, g_left)
p_right = self.location_samples[:, :, :, 1] + self.location_samples[:, :, :, 3]
g_right = self.targets_placeholder[:, :, 1] + self.targets_placeholder[:, :, 3]
right = tf.minimum(p_right, g_right)
p_top = self.location_samples[:, :, :, 0]
g_top = self.targets_placeholder[:, :, 0]
top = tf.maximum(p_top, g_top)
p_bottom = self.location_samples[:, :, :, 0] + self.location_samples[:, :, :, 2]
g_bottom = self.targets_placeholder[:, :, 0] + self.targets_placeholder[:, :, 2]
bottom = tf.minimum(p_bottom, g_bottom)
intersection = tf.maximum((right - left), 0) * tf.maximum((bottom - top), 0)
p_area = self.location_samples[:, :, :, 3] * self.location_samples[:, :, :, 2]
g_area = self.targets_placeholder[:, :, 3] * self.targets_placeholder[:, :, 2]
union = p_area + g_area - intersection
return intersection/union
def add_loss_op(self, loss_type='negative_l1_dist'):
self.loss_type = loss_type
logits_shape = tf.shape(self.logits)
logits_flat = tf.reshape(self.logits, [-1])
location_dist = tf.contrib.distributions.MultivariateNormalDiag(mu=logits_flat,
diag_stdev=self.config.variance*tf.ones_like(logits_flat))
location_samples = location_dist.sample([self.config.num_samples])
new_logits_shape = tf.concat([[self.config.num_samples,] , logits_shape], axis=0)
location_samples = tf.reshape(location_samples, new_logits_shape)
self.location_samples = location_samples
if self.loss_type == 'negative_l1_dist':
rewards = -tf.reduce_mean(tf.abs(location_samples - tf.cast(self.targets_placeholder,tf.float32)),axis=3,keep_dims=True) - \
tf.reduce_max(tf.abs(location_samples - tf.cast(self.targets_placeholder,tf.float32)), axis=3,keep_dims=True)
elif self.loss_type == 'iou':
rewards = self.get_iou_loss()
rewards = tf.expand_dims(rewards,axis=-1)
timestep_rewards = tf.reduce_mean(rewards, axis=0, keep_dims=True)
self.timestep_rewards = timestep_rewards
if self.cumsum:
tot_cum_rewards = tf.cumsum(rewards, axis=2, reverse=True)
else:
tot_cum_rewards = tf.tile(tf.reduce_sum(rewards, axis=2, keep_dims = True),multiples=[1,1,self.config.seq_len, 1])
self.tot_cum_rewards = tot_cum_rewards
timestep_rewards_grad_op = tf.stop_gradient(timestep_rewards)
rewards_grad_op = tf.stop_gradient(rewards)
location_samples_op = tf.stop_gradient(location_samples)
tot_cum_rewards_op = tf.stop_gradient(tot_cum_rewards)
const1 = 1.0 / (np.sqrt(2.0 * math.pi) * self.config.variance)
const2 = 2.0 * self.config.variance**2
squared_diff = tf.square(self.targets_placeholder - self.logits)
density_func = tf.log(const1 * tf.exp(-squared_diff / const2))
self.density_func = density_func
self.loss = tf.reduce_mean(tf.reduce_sum(density_func*(tot_cum_rewards_op - timestep_rewards_grad_op), axis=2),
axis=[1, 0])
self.total_rewards = tf.reduce_mean(tf.reduce_sum(timestep_rewards, axis=2), axis=1)
tf.summary.scalar('Total Rewards', self.total_rewards[0][0])
def add_optimizer_op(self):
tvars = tf.trainable_variables()
grads = tf.gradients(self.loss, tvars)
optimizer = tf.train.AdamOptimizer(self.config.lr)
self.train_op = optimizer.apply_gradients(zip(grads, tvars))
def add_error_op(self):
# VOT metrics (MOT only makes sense for multiple object)
# Accuracy:
# intersection / union
# Robustness
# average count of number of resets (0 overlap in predicted and actual)
# y, x, height, width
# left = x
# right = x + width
# top = y
# bottom = y + height
# IoU Metric calculation
p_left = self.logits[:, :, 1]
g_left = self.targets_placeholder[:, :, 1]
left = tf.maximum(p_left, g_left)
self.left = left
p_right = self.logits[:, :, 1] + self.logits[:, :, 3]
g_right = self.targets_placeholder[:, :, 1] + self.targets_placeholder[:, :, 3]
right = tf.minimum(p_right, g_right)
self.right = right
p_top = self.logits[:, :, 0]
g_top = self.targets_placeholder[:, :, 0]
top = tf.maximum(p_top, g_top)
self.top = top
p_bottom = self.logits[:, :, 0] + self.logits[:, :, 2]
g_bottom = self.targets_placeholder[:, :, 0] + self.targets_placeholder[:, :, 2]
bottom = tf.minimum(p_bottom, g_bottom)
self.bottom = bottom
intersection = tf.maximum((right - left), 0) * tf.maximum((bottom - top), 0)
self.intersection = intersection
p_area = self.logits[:, :, 3] * self.logits[:, :, 2]
g_area = self.targets_placeholder[:, :, 3] * self.targets_placeholder[:, :, 2]
union = p_area + g_area - intersection
self.union = union
self.area_accuracy = tf.reduce_mean(intersection / union)
tf.summary.scalar('IOU Area Accuracy', self.area_accuracy)
# Bounding box summaries
p_seq_image_bboxes = []
g_seq_image_bboxes = []
for i in xrange(self.config.seq_len):
p_left_i = self.logits[:, i, 1]
g_left_i = self.targets_placeholder[:, i, 1]
p_right_i = self.logits[:, i, 1] + self.logits[:, i, 3]
g_right_i = self.targets_placeholder[:, i, 1] + self.targets_placeholder[:, i, 3]
p_top_i = self.logits[:, i, 0]
g_top_i = self.targets_placeholder[:, i, 0]
p_bottom_i = self.logits[:, i, 0] + self.logits[:, i, 2]
g_bottom_i = self.targets_placeholder[:, i, 0] + self.targets_placeholder[:, i, 2]
p_top_i = tf.expand_dims(p_top_i, axis=-1)
p_left_i = tf.expand_dims(p_left_i, axis=-1)
p_bottom_i = tf.expand_dims(p_bottom_i, axis=-1)
p_right_i = tf.expand_dims(p_right_i, axis=-1)
g_top_i = tf.expand_dims(g_top_i, axis=-1)
g_left_i = tf.expand_dims(g_left_i, axis=-1)
g_bottom_i = tf.expand_dims(g_bottom_i, axis=-1)
g_right_i = tf.expand_dims(g_right_i, axis=-1)
p_bboxes_i = tf.expand_dims(tf.concat([p_top_i, p_left_i, p_bottom_i, p_right_i], axis=-1), axis=1)
g_bboxes_i = tf.expand_dims(tf.concat([g_top_i, g_left_i, g_bottom_i, g_right_i], axis=-1), axis=1)
# squeezed_seq_input = tf.squeeze(self.inputs_placeholder[:, i, :, :, :], axis=1)
# print p_bboxes_i.get_shape().as_list()
# print g_bboxes_i.get_shape().as_list()
p_image_bboxes = tf.image.draw_bounding_boxes(self.inputs_placeholder[:, i, :, :, :], p_bboxes_i)
g_image_bboxes = tf.image.draw_bounding_boxes(self.inputs_placeholder[:, i, :, :, :], g_bboxes_i)
p_seq_image_bboxes.append(p_image_bboxes)
g_seq_image_bboxes.append(g_image_bboxes)
p_image_bboxes = tf.concat(p_seq_image_bboxes, axis=2)
g_image_bboxes = tf.concat(g_seq_image_bboxes, axis=2)
bbox_summary = tf.concat([p_image_bboxes, g_image_bboxes], axis=1)
tf.summary.image('bounding boxes', bbox_summary, max_outputs=10)
def add_summary_op(self):
self.summary_op = tf.summary.merge_all()
def add_feed_dict(self, input_batch, target_batch, seq_len_batch , init_locations_batch):
feed_dict = {self.inputs_placeholder:input_batch, self.targets_placeholder:target_batch,
self.init_loc:init_locations_batch, self.seq_len_placeholder:seq_len_batch}
return feed_dict
def train_one_batch(self, session, input_batch, target_batch, seq_len_batch , init_locations_batch):
feed_dict = self.add_feed_dict(input_batch, target_batch, seq_len_batch , init_locations_batch)
_, loss, summary, density_func, total_rewards, area_accuracy = session.run([
self.train_op,
self.loss,
self.summary_op,
self.density_func,
self.total_rewards,
self.area_accuracy],
feed_dict)
return summary, loss, total_rewards[0][0], area_accuracy
def test_one_batch(self, session, input_batch, target_batch, seq_len_batch , init_locations_batch):
feed_dict = self.add_feed_dict(input_batch, target_batch, init_locations)
# Accuracy
loss, summary, rewards, area_accuracy = session.run([self.loss, self.summary_op, self.total_rewards, self.area_accuracy], feed_dict)
return summary, loss, rewards, area_accuracy
def run_one_batch(self, args, session, input_batch, target_batch, seq_len_batch , init_locations_batch):
if args.train == 'train':
summary, loss, rewards, area_accuracy = self.train_one_batch(session, input_batch, target_batch, seq_len_batch , init_locations_batch)
else:
summary, loss, rewards, area_accuracy = self.test_one_batch(session, input_batch, target_batch, seq_len_batch , init_locations_batch)
return summary, loss, rewards, area_accuracy
def get_config(self):
return self.config
def add_update_weights_op(self, input_model, gamma):
q_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=input_model.scope)
target_q_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.scope)
update_ops = []
for targ, orig in zip(target_q_vars, q_vars):
new_targ = tf.assign(targ, gamma*orig + (1-gamma)*targ)
update_ops.append(new_targ)
self.update_target_op = tf.group(*update_ops)
def update_weights(self, session):
session.run(self.update_target_op)