nonlinearity=args.act,
seed=args.seed,
debug=False)
total_params = sum([np.prod(x[1].shape) for x in model.params.items()])
print('RNN type: ' + args.rnn + " Grammar: " + args.data + " Seed: " +
str(args.seed))
print('Model total parameters: {}'.format(total_params))
try:
if args.continue_train:
model.params, h_init = load_params(params_file, hinit_file,
model.params)
model.reload_hidden(h_init, args.batch)
model.init_tparams()
model.build_model()
else:
model.init_hidden(args.batch, args.seed)
save_hinit(model.h_init[0], hinit_file)
model.build_model()
model_test.reload_hidden(model.h_init[0], args.test_batch)
model_test.build_model()
# train the model
if args.curriculum:
model = curriculum_train(model=model,
x=train_x,
m=train_m,
y=train_y,
x_v=val_x,
def train(args):
positive_data_dir = args.positive_data_dir
negative_data_dir = args.negative_data_dir
emb_size = args.emb_size
hidden_size = args.hidden_size
dropout = args.dropout
lr = args.learning_rate
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
result_dir = 'wiki_' + timestamp
result_dir = os.path.join(args.result_dir, result_dir)
if not os.path.exists(result_dir):
os.makedirs(result_dir)
# Output to Stream
handler1 = logging.StreamHandler()
handler1.setLevel(logging.DEBUG)
handler1.setFormatter(
logging.Formatter("%(asctime)s %(levelname)8s %(message)s"))
# Output to log file
handler2 = logging.FileHandler(
filename=os.path.join(result_dir, "log_{}.log".format(timestamp)))
handler2.setLevel(logging.DEBUG)
handler2.setFormatter(
logging.Formatter("%(asctime)s %(levelname)8s %(message)s"))
logger = logging.getLogger("StyleLearning.wiki")
logger.addHandler(handler1)
logger.addHandler(handler2)
logger.info("results data will be saved at {}".format(result_dir))
positive_data_path = [
p for p in glob.glob(os.path.join(positive_data_dir, "**/*"),
recursive=True) if os.path.isfile(p)
]
negative_data_path = [
p for p in glob.glob(os.path.join(negative_data_dir, "**/*"),
recursive=True) if os.path.isfile(p)
]
positive_data_path.sort()
negative_data_path.sort()
epochs = args.epochs
positive_data_path = positive_data_path[:300]
negative_data_path = negative_data_path[:300]
batch_size = 32
T = 32
data_gen = DataForGenerator(batch_size=batch_size, T=T)
for d in positive_data_path:
data_gen.load_sentenceid_data(d, label=True)
for d in negative_data_path:
data_gen.load_sentenceid_data(d, label=False)
logger.info("Building vocablary...")
data_gen.load_vocab('./vocab.csv', vocab_size=50000)
data_gen.make_train_and_test_data()
test_data = data_gen.get_test_data()
vocab_size = len(data_gen.vocab.word2id)
logger.info("Vocab size: ", vocab_size)
gen = data_gen.generate_training_data()
model = RNNModel(batch_size=batch_size,
vocab_size=vocab_size,
emb_size=emb_size,
hidden_size=hidden_size,
T=T,
dropout=dropout,
lr=lr,
model_path=result_dir)
model.build_model(optimizer=Adam(lr),
loss='binary_crossentropy',
metrics=['accuracy'])
steps_per_epoch = len(data_gen)
display_params(logger, \
{"batch_size": batch_size, "padding_size":T, "emb_size":emb_size, "hidden_size":128, "dropout":dropout, "learning_rate":lr})
model.fit_generator(generator=gen,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
validation_data=test_data)
