def load_and_train(epochs=50, vectorlength=16):
# Set up logging
dictConfig(LOG_CONFIG)
games = Games()
# Use the following to load already parsed data
games.load()
games.flatten(vectorlength=vectorlength)
# Network structure
# Structure: [input_layer, hidden_layer, hidden_layer ... , output_layer]
# Example: [784, 620, 100, 10]
layer_structure = [vectorlength, 192, 128, 96, 4]
# Example: [rectify, rectify, softmax]
activation_functions = [rectify, rectify, rectify, rectify, softmax]
# Remeber to change num_labels to 4, since we have Up, Right, Down, Left
# Also we normalize the values. Don't know if it will affect anything,
# but not taking any chances.
cfg = {
'learning_rate': 0.00001,
'num_labels': 4,
'training_batch_size': 512,
}
# Create a network using the default parameters
net = ANN(layer_structure, activation_functions, config=cfg)
net.load_input_data(normalize=False, module6_file=True)
net.train(epochs=epochs, include_test_set=False)
return net
def load_raw_and_save(alternate=False, num_games=16, only_successful=False, vectorlength=16):
"""
Loads raw data from game_data.txt, parses, transforms to alternate repr if set to True,
then flattens the data, and pickles and saves the data to disk.
"""
games = Games()
# Use the following to parse 512 games from scratch, that have at least reached 2048,
# then add an additional filter of min_maxtile 4096, flatten to Nx16 np.ndArray and gzip pickle
games.parse_from_raw_game_data(num_games=num_games, only_successful=only_successful)
if alternate:
games.transform_to_alternate_representation()
boards, labels = games.flatten(vectorlength=vectorlength)
print('Total labels: %d' % len(labels))
print('Total board states: %d' % len(boards))
games.save()