def run_epoch(session, model, char_data, tag_data, dict_data, len_data, eval_op, batch_size, verbose=False):
"""Runs the model on the given data."""
start_time = time.time()
losses = 0.0
iters = 0.0
char_data, tag_data, dict_data, len_data = reader.shuffle(char_data, tag_data, dict_data, len_data)
xArray, yArray, dArray, lArray = reader.iterator(char_data, tag_data, dict_data, len_data, batch_size)
for x, y, d, l in zip(xArray, yArray, dArray, lArray):
fetches = [model.loss, model.logits, eval_op]
feed_dict = {}
feed_dict[model.input_data] = x
feed_dict[model.targets] = y
feed_dict[model.dicts] = d
feed_dict[model.seq_len] = l
loss, logits, _ = session.run(fetches, feed_dict)
losses += loss
iters += 1
if verbose and iters % 50 == 0:
print("%.3f perplexity: %.3f" %
(iters / float(len(xArray)), np.exp(losses / iters / len(xArray))))
return np.exp(losses / iters)
def load_data():
with open('input/data.txt', encoding='utf-8') as file:
X = []
y = []
seq_len = []
X, y = shuffle(X, y)
X = X.tolist()
y = y.tolist()
for index, lien in enumerate(file.readlines()):
if index % 2 == 0:
X.append(lien.strip())
else:
y.append(lien.strip())
word2idx, _ = load_vocab()
x_index = []
for sentence in X:
sentence = list(sentence.replace(' ', ''))
if len(sentence) < 15:
seq_len.append(len(sentence))
else:
seq_len.append(15)
# max_len = max(seq_len)
sentence_index = [
word2idx[i] if i in list(word2idx.keys()) else 1
for i in sentence
]
x_index.append(sentence_index)
X = pad_sequences(x_index, maxlen=15, value=0)
tag2idx, _ = load_tag()
tag_indexs = []
for tag in y:
tag = tag.split(' ')
tag_index = [
tag2idx[i] if i in list(tag2idx.keys()) else 0 for i in tag
]
tag_indexs.append(tag_index)
y = pad_sequences(tag_indexs, maxlen=15, value=0)
return X, y, seq_len
def main():
parser = argparse.ArgumentParser(
description=
'Split train.csv into train, dev, and test splits. Specify dev and validation set sizes with args, the remainder is used for training.'
)
parser.add_argument(
'--dataset-file',
required=True,
help='path to the train.csv file containing the quora training data')
parser.add_argument('--ndev',
type=int,
default=1e4,
help='size of dev set to create')
parser.add_argument('--nvalid',
type=int,
default=5e4,
help='size of validation set to create')
parser.add_argument(
'--output-dir',
required=True,
help='directory to which to write train.csv, dev.csv, and valid.csv')
parser.add_argument(
'--seed',
type=int,
help=
'optional random seed to have reproducibility between multiple uses of this tool'
)
args = parser.parse_args()
data = du.load_csv(args.dataset_file)
shuffled = du.shuffle(data, args.seed)
ntrain = len(data) - args.ndev - args.nvalid
train, dev, valid = du.split(shuffled, ntrain, args.ndev, args.nvalid)
du.write_csv(train, os.path.join(args.output_dir, 'train.csv'))
du.write_csv(dev, os.path.join(args.output_dir, 'dev.csv'))
du.write_csv(valid, os.path.join(args.output_dir, 'valid.csv'))
##############################################################################################
# TRAINING and results
##############################################################################################
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
FG = []
FD = []
F_GEO = []
for epoch in tqdm(range(n_epoch), total=n_epoch):
BHA_gen = np.array([]);
AHB_gen = np.array([]);
AH_dataset, BH_dataset= shuffle(AH_dataset, BH_dataset)
for x, y in iter_data(AH_dataset, BH_dataset, size=batch_size):
for _ in range(1):
f_d, _ = sess.run([disc_loss, train_disc_op], feed_dict={x_AH: x, y_BH:y})
for _ in range(5):
f_g, _ = sess.run([[adv_loss, cost_BH, cost_AH], train_gen_op], feed_dict={x_AH: x, y_BH:y})
FG.append(f_g)
FD.append(f_d)
for x, y in iter_data(AH_dataset, BH_dataset, size=batch_size):
temp_BHA_gen = sess.run(BHA, feed_dict={x_AH: x, y_BH:y})
BHA_gen = np.vstack([BHA_gen, temp_BHA_gen]) if BHA_gen.size else temp_BHA_gen
#minimize generators loss
train_gen_op = opt.minimize(gen_loss, var_list=qvars + pvars)
#minimize discirimaintors loss
train_disc_op = opt.minimize(disc_loss, var_list=dvars_x + dvars_z)
""" training """
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
FG = []
FD = []
#for each epoch (log the status bar)
for epoch in tqdm(range(n_epoch), total=n_epoch):
#sample from both our datasets
X_dataset, Z_dataset = shuffle(X_dataset, Z_dataset)
#for each x and z in our data
for xmb, zmb in iter_data(X_dataset, Z_dataset, size=batch_size):
#minimize our loss functions
for _ in range(1):
f_d, _ = sess.run([disc_loss, train_disc_op],
feed_dict={
x: xmb,
z: zmb
})
for _ in range(5):
#3 components that make up generator loss
f_g, _ = sess.run([[adv_loss, cost_x, cost_z], train_gen_op],
feed_dict={