Some work in progress

scripts/training/bidding/bidding_nn_tf2.py

@@ -1,5 +1,6 @@
 import sys
 sys.path.append('../../../src')
+import os 
 
 import numpy as np
 import datetime
@@ -16,7 +17,7 @@
 
 model_path = 'model/bidding.h5'
 
-batch_size = 64
+batch_size = 1
 display_step = 1000
 epochs = 1
 
@@ -26,10 +27,12 @@
 print("Input shape:  ", X_train.shape)
 print("Output shape: ", y_train.shape)
 n_examples = y_train.shape[0]
-n_ftrs = X_train.shape[1]
-n_bids = y_train.shape[1]
+n_sequence_length = X_train.shape[1]
+n_ftrs = X_train.shape[2]
+n_bids = y_train.shape[2]
 
-print("Size input hand:         ", n_ftrs)
+print("Size input:              ", n_ftrs)
+print("Sequence length:         ", n_sequence_length)
 print("Examples for training:   ", n_examples)
 print("Batch size:              ", batch_size)
 n_iterations = round(((n_examples / batch_size) * epochs) / 1000) * 1000
@@ -38,19 +41,23 @@
 
 lstm_size = 128
 n_layers = 3
+keep_prob = 0.8
 
+input_shape = (n_sequence_length, n_ftrs)
+# Create a Sequential model with stateful LSTM and specify batch_input_shape
 model = Sequential()
+model.add(LSTM(lstm_size, dropout=1 - keep_prob, stateful=True, return_sequences=True, batch_input_shape=(batch_size,) + input_shape))
 
-for _ in range(n_layers):
-    model.add(LSTM(lstm_size, dropout=0.2, stateful=True))
+for _ in range(n_layers - 1):  # Subtract 1 to account for the first LSTM layer
+    model.add(LSTM(lstm_size, dropout=1 - keep_prob, stateful=True, return_sequences=True))
 
 model.add(Dense(n_bids, activation='softmax'))
 
 model.compile(loss='categorical_crossentropy', optimizer=Adam(learning_rate=0.001))
 
 # Create Batcher instances
 batch = Batcher(n_examples, batch_size)
-cost_batch = Batcher(n_examples, display_step)
+cost_batch = Batcher(n_examples, batch_size)
 
 for i in range(n_iterations):
     x_batch, y_batch = batch.next_batch([X_train, y_train])

scripts/training/bidding/bidding_nn_tf2xx.py

@@ -0,0 +1,76 @@
+import sys
+import os
+sys.path.append('../../../src')
+
+import numpy as np
+import datetime
+from keras.models import Sequential
+from keras.layers import LSTM, Dropout, Dense
+from keras.optimizers import Adam
+from batcher import Batcher
+
+if len(sys.argv) < 2:
+    print("Usage: python bidding_nn inputdirectory")
+    sys.exit(1)
+
+bin_dir = sys.argv[1]
+
+model_path = './model/bidding.h5'
+
+# Batch size is 1 so we can reset state between each sequence
+batch_size = 32
+display_step = 1000
+epochs = 1
+
+X_train = np.load(os.path.join(bin_dir, 'x.npy'))
+y_train = np.load(os.path.join(bin_dir, 'y.npy'))
+
+print("Input shape:  ", X_train.shape)
+print("Output shape: ", y_train.shape)
+n_examples = y_train.shape[0]
+n_sequence_length = X_train.shape[1]
+n_ftrs = X_train.shape[2]
+n_bids = y_train.shape[2]
+
+print("Size input hand:         ", n_ftrs)
+print("Sequence length:         ", n_sequence_length)
+print("Examples for training:   ", n_examples)
+print("Batch size:              ", batch_size)
+n_iterations = round(((n_examples / batch_size) * epochs) / 1000) * 1000
+print("Iterations               ", n_iterations)
+print("Model path:              ", model_path)
+
+lstm_size = 128
+n_layers = 3
+keep_prob = 0.8  # Adjust the dropout rate as needed
+
+input_shape = (n_sequence_length, n_ftrs)
+# Create a Sequential model with stateful LSTM and specify batch_input_shape
+model = Sequential()
+model.add(LSTM(lstm_size, dropout=1 - keep_prob, stateful=True, return_sequences=True, batch_input_shape=(batch_size,) + input_shape))
+
+for _ in range(n_layers - 1):  # Subtract 1 to account for the first LSTM layer
+    model.add(LSTM(lstm_size, dropout=1 - keep_prob, stateful=True, return_sequences=True))
+
+model.add(Dense(n_bids, activation='softmax'))
+model.compile(loss='categorical_crossentropy', optimizer=Adam(learning_rate=0.001))
+
+model.summary()
+
+# Create Batcher instances
+batch = Batcher(n_examples, batch_size)
+cost_batch = Batcher(n_examples, batch_size)
+
+for i in range(n_iterations):
+    x_batch, y_batch = batch.next_batch([X_train, y_train])
+    if (i != 0) and i % display_step == 0:
+        # Evaluate the model on the test data
+        x_cost, y_cost = cost_batch.next_batch([X_train, y_train])
+        c_train = model.evaluate(x_cost, y_cost, verbose=0)
+        print('{} {}. c_train={}'.format(datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"), i, c_train))
+        sys.stdout.flush()
+        model.save(model_path, overwrite=True)  
+
+    model.train_on_batch(x_batch, y_batch)
+
+model.save(model_path, overwrite=True)

scripts/training/bidding/default.conf

@@ -3,7 +3,7 @@ include_system = False
 tf_version = 1
 
 [bidding]
-bidder = models/gib21_model/gib21-1000000
+bidder = scripts/training/bidding/model/bidding-1248000
 info = models/gib21_info_model/gib21_info-500000
 search_threshold = 0.1