def train(args):
"""
Train the CNN.
:param args:
"""
# get our working device
device = get_device(args)
# read training samples from disk, and prepare the dataset (i.e., shuffle, scaling, and moving the data to tensors.)
dts = dataset(device)
dts.read_training_dataset(args['train_data'])
train_dts, test_dts = dts, dts
if args['D'] != None:
train_dts, test_dts = dts.train_test_splits(args['D'], td=False)
classifier = seq2seq(dts.words_converter.no_entries(),
dts.slots_converter.no_entries(),
dts.intent_converter.no_entries(), device,
train_dts.words_converter.T2id('<PAD>'),
train_dts.slots_converter.T2id('<PAD>'), args)
if args['E'] != None:
embeddings = gensim.models.KeyedVectors.load_word2vec_format(
args['E'][0], binary=True)
classifier.pretrained_embeddings(train_dts, embeddings)
try:
classifier.fit(train_dts,
test_dts,
args['e'],
args['b'],
args['lr'],
args['save_every'],
args['o'],
do_predict=not args['no_training_predictions'])
except KeyboardInterrupt:
pass
intent_true, intent_pred, slots_true, slots_pred = classifier.predict_and_get_labels_batch(
test_dts, args['b'])
assess(dts, intent_true, intent_pred, slots_true, slots_pred)
if args['o'] != None:
# write it into a file.
classifier.dump(args['o'])
def add_level(node, play, level):
if level <= 0: return
gameState = node.contents[0]
all_play_moves = {
'w': white_moves.all_white_moves,
'b': black_moves.all_black_moves
}
positions, moves = all_play_moves[play](gameState)
play_moves = {'w': white_moves.move_white, 'b': black_moves.move_black}
if play == 'w': next_play = 'b'
elif play == 'b': next_play = 'w'
for ind in range(len(positions)):
for piece_move in moves[ind]:
new_state = gameState.copy()
play_moves[play](new_state, positions[ind], piece_move)
new_node = node.addNode([new_state, positions[ind], piece_move],
assessment.assess(new_state))
add_level(new_node, next_play, level - 1)
# Prune the node
if play == 'w':
node.pruneSetMax()
elif play == 'b':
node.pruneSetMin()
def move(gameState, play, depth=1):
trunk = tree.Node(0, [gameState, 0, 0], assessment.assess(gameState))
start_play = play
add_level(trunk, play, depth)
if play == 'w':
trunk.pruneSetMax()
if play == 'b':
trunk.pruneSetMin()
return trunk.nodes[0].contents
def fit(self, train, dev, epochs, batch_size, learning_rate, save_every, save_model_path, print_every=50, do_predict=True):
for i in range(epochs):
print(f"******Epoch: {i}********")
self.trainIters(train, batch_size, learning_rate=learning_rate, print_every=print_every, no_epochs=epochs)
if do_predict:
print(' predicting:')
intent_true, intent_pred, slots_true, slots_pred = self.predict_and_get_labels_batch(dev, batch_size)
intent_assessment, slots_assessment = assess(dev, intent_true, intent_pred, slots_true, slots_pred,
plot=False)
if save_every != None and i % save_every == 0:
self.dump(save_model_path + '.epoch' + str(i))
state = state.State()
print(state)
for itt in range(80):
# White move
positions, moves = white_moves.all_white_moves(state)
ind = random.randrange(len(positions))
position = positions[ind]
dst = random.choice(moves[ind])
white_moves.move_white(state, position, dst)
print(
str(itt) + ": White moves " + convert_position.coord2basic(position) +
" --> " + convert_position.coord2basic(dst))
print(state)
print("Assessment: " + str(assessment.assess(state)))
if state.isEnd(): break
print()
# Black move
positions, moves = black_moves.all_black_moves(state)
ind = random.randrange(len(positions))
position = positions[ind]
dst = random.choice(moves[ind])
black_moves.move_black(state, position, dst)
print(
str(itt) + ": Black moves " + convert_position.coord2basic(position) +
" --> " + convert_position.coord2basic(dst))
print(state)
print("Assessment: " + str(assessment.assess(state)))
if state.isEnd(): break