if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Sample from a trained SeqGAN model.')
parser.add_argument('sample_len', metavar='N', type=int,
help='length of sample to generate')
parser.add_argument('-t', '--dictionary', default='dictionary.pkl',
type=str, help='path to dictionary file')
parser.add_argument('-d', '--logdir', default='model/', type=str,
help='directory of the trained model')
parser.add_argument('-c', '--only_cpu', default=True, action='store_true',
help='if set, only build weights on cpu')
args = parser.parse_args()
if not os.path.exists(args.dictionary):
raise ValueError('No dictionary file found: "%s". To build it, '
'run train.py' % args.dictionary)
_, rev_dict = utils.get_dictionary(None, dfile=args.dictionary)
num_classes = len(rev_dict)
sess = tf.Session()
model = SeqGAN(sess,
num_classes,
logdir=args.logdir,
only_cpu=args.only_cpu)
model.build()
model.load(ignore_missing=True)
g = model.generate(args.sample_len)
print('Generated text:', utils.detokenize(g, rev_dict))
root.setLevel(logging.DEBUG)
dictionary, rev_dict = utils.get_dictionary(args.text)
num_classes = len(dictionary)
iterator = utils.tokenize(args.text,
dictionary,
batch_size=args.batch_size,
seq_len=args.seq_len)
sess = tf.Session()
model = SeqGAN(sess,
num_classes,
logdir=args.logdir,
learn_phase=args.learn_phase,
only_cpu=args.only_cpu)
model.build()
model.load(ignore_missing=True)
for epoch in xrange(1, args.num_epochs + 1):
for step in xrange(1, args.num_steps + 1):
logging.info('epoch %d, step %d', epoch, step)
model.train_batch(iterator.next())
# Generates a sample from the model.
g = model.generate(1000)
print(utils.detokenize(g, rev_dict))
# Saves the model to the logdir.
model.save()
generator = SeqGAN(vocab_size=3000, emb_dim=128, hidden_dim=128,
sequence_length=40, start_token=0, lstm_layer=2
).to_gpu()
batch_size = 10000
def progress_report(count, start_time, batch_size):
duration = time.time() - start_time
throughput = count * batch_size / duration
sys.stderr.write(
'\rtrain {} updates ({} samples) time: {} ({:.2f} samples/sec)'
.format(count, count * batch_size,
str(datetime.timedelta(seconds=duration)).split('.')[0], throughput))
negative = []
# pool=None
st = time.time()
for x in range(30000 // batch_size):
negative.append(generator.generate(batch_size))
progress_report(x, st, batch_size)
t = time.time()
print()
print(t - st)
for x in range(30000 // batch_size):
negative.append(generator.generate(batch_size, pool))
progress_report(x, t, batch_size)
print(time.time() - t)