def main():
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
args.max_len = 100
args.batch_size = 64
args.max_dec_len = 100
args.display_info_step = 1000
args.isPointer = False
args.vocab_limit = 35000
train_data_len = 209940
args.diff_input = True
print(args)
dataloader = IWSLT(batch_size=args.batch_size, vocab_limit=args.vocab_limit, max_input_len=args.max_len, max_output_len=args.max_dec_len)
params = {
'vocab_size_encoder': len(dataloader.idx2token),
'vocab_size': len(dataloader.word2idx),
'word2idx': dataloader.word2idx,
'idx2word': dataloader.idx2word,
'idx2token': dataloader.idx2token}
print('Vocab Size:', params['vocab_size'])
model = VRAE(params)
saver = tf.train.Saver()
exp_path = "./saved/iwslt_seq2seq/"
model_name = "seq2seq.ckpt"
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth=True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
EPOCH_STEPS = (train_data_len-1)//args.batch_size+1
summary_writer = tf.summary.FileWriter(exp_path, sess.graph)
restore_path = tf.train.latest_checkpoint(exp_path)
if restore_path:
saver.restore(sess, restore_path)
last_train_step = int(restore_path.split("-")[-1]) % EPOCH_STEPS
print("Model restore from file: %s, last train step: %d" % (restore_path, last_train_step))
# summary_writer = tf.summary.FileWriter(exp_path)
# saver.restore(sess, exp_path+model_name)
for epoch in range(args.num_epoch):
# dataloader.update_word_dropout()
# print("\nWord Dropout")
# dataloader.shuffle()
# print("Data Shuffled", end='\n\n')
batcher = dataloader.load_data()
step = -1
while True:
try:
# enc_inp, dec_inp_full, dec_out = next(dataloader.data_loader)
(x_enc_inp, x_dec_inp_full, x_dec_out, y_enc_inp, y_dec_inp_full, y_dec_out), x_enc_inp_oovs, data_oovs, _ = next(batcher)
# enc_inp, dec_inp_full, dec_out = dataloader.next_batch()
enc_inp, dec_inp_full, dec_out = x_enc_inp, y_dec_inp_full, y_dec_out
dec_inp = dataloader.update_word_dropout(dec_inp_full)
step += 1
except StopIteration:
print("there are no more examples")
break
# print(step, "enc_inp.shape:", enc_inp.shape)
# print(step, "dec_inp_full.shape:", dec_inp_full.shape)
# print(step, "dec_out.shape:", dec_out.shape)
log = model.train_session(sess, enc_inp, dec_inp, dec_out)
# get the summaries and iteration number so we can write summaries to tensorboard
summaries, train_step = log['summaries'], log['step']
summary_writer.add_summary(summaries, train_step) # write the summaries
if train_step % 100 == 0: # flush the summary writer every so often
summary_writer.flush()
if step % args.display_loss_step == 0:
print("Step %d | [%d/%d] | [%d/%d]" % (log['step'], epoch+1, args.num_epoch, step, train_data_len//args.batch_size), end='')
print(" | loss:%.3f" % (log['loss']))
if step % args.display_info_step == 0:
model.reconstruct(sess, enc_inp[-1], dec_out[-1])
save_path = saver.save(sess, exp_path+model_name, global_step=train_step)
print("Model saved in file: %s" % save_path)
model.reconstruct(sess, enc_inp[-1], dec_out[-1])
save_path = saver.save(sess, exp_path+model_name, global_step=train_step)
print("Model saved in file: %s" % save_path)
def main():
## CUDA
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.cuda)
## Parameters
if args.exp == "NONE":
args.exp = args.graph_type
args.enc_max_len = 100
args.dec_max_len = 100
args.vocab_limit = 35000
exp_path = "./saved/"+args.exp+"/"
if not os.path.exists(exp_path):
os.makedirs(exp_path)
model_name = args.model_name
train_data_len = 209940
args.data_len = train_data_len
train_data_path = args.train_data
EPOCH_STEPS = (train_data_len-1)//args.batch_size+1
args.diff_input = True
print(args)
## DataLoader
dataloader = IWSLT(batch_size=args.batch_size, vocab_limit=args.vocab_limit, max_input_len=args.enc_max_len, max_output_len=args.dec_max_len)
params = {
'vocab_size': len(dataloader.word2idx),
'vocab_size_encoder': len(dataloader.token2idx),
'word2idx': dataloader.word2idx,
'idx2word': dataloader.idx2word,
'idx2token': dataloader.idx2token,
'token2id': dataloader.token2idx,
'loss_type': args.loss_type,
'graph_type': args.graph_type}
print('Vocab Size:', params['vocab_size'])
## ModelInit
model = VAESEQ(params)
log_path = exp_path+"log.txt"
LOGGER = open(log_path, "a")
## Session
# load some parameters
variables = tf.contrib.framework.get_variables_to_restore()
# print(len(variables), end=",")
# variables = [v for v in variables if not v.name.startswith("optimizer/transformer/trans_mlp/")]
# for v in variables_to_resotre:
# print(type(v.name), v.name)
print(len(variables))
# end load
saver = tf.train.Saver(variables)
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth=True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(exp_path, sess.graph)
# tf.train.write_graph(sess.graph, './saved/vaeseq/', 'train.pbtxt')
keep_on_train_flag = False
restore_path = tf.train.latest_checkpoint(exp_path)
if restore_path:
keep_on_train_flag = True
saver.restore(sess, restore_path)
saver = tf.train.Saver() # new saver
last_train_step = int(restore_path.split("-")[-1]) % EPOCH_STEPS
print("Model restore from file: %s, last train step: %d" % (restore_path, last_train_step))
LOGGER.write("Model restore from file: %s, last train step: %d\n" % (restore_path, last_train_step))
# Train Mode
x_log, y_log, t_log, log = None, None, None, None
for epoch in range(args.num_epoch):
# batcher = dataloader.load_data(fpath=train_data_path)
batcher = dataloader.load_data()
for step in tqdm(range(EPOCH_STEPS)):
if keep_on_train_flag and step < last_train_step: continue
if keep_on_train_flag and step == (last_train_step): keep_on_train_flag=False
# get batch data
try:
(x_enc_inp, x_dec_inp_full, x_dec_out, y_enc_inp, y_dec_inp_full, y_dec_out), x_enc_inp_oovs, data_oovs, _ = next(batcher)
x_dec_inp = dataloader.update_word_dropout(x_dec_inp_full)
y_dec_inp = dataloader.update_word_dropout(y_dec_inp_full)
max_oovs_len = max([len(oov) for oov in data_oovs]) if len(data_oovs) > 0 else 0
except StopIteration:
print("there are no more examples")
break
# for _ in range(2):
# x_log = model.train_encoder(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
# y_log = model.train_decoder(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
# t_log = model.train_transformer(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
# t_log = model.merged_transformer_train(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
# log = model.merged_train(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
x_log = model.train_encoder(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
y_log = model.train_decoder(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out)
if step % 3 != 0:
log = model.merged_seq_train(sess, x_enc_inp, x_dec_inp, x_dec_out, y_enc_inp, y_dec_inp, y_dec_out, x_enc_inp_oovs, max_oovs_len)
# model.show_parameters(sess)
# get the summaries and iteration number so we can write summaries to tensorboard
train_step = summary_flush(x_log, y_log, t_log, log, summary_writer)
if step % args.display_loss_step == 0:
print("Step %d | [%d/%d] | [%d/%d]" % (train_step, epoch+1, args.num_epoch, step, train_data_len//args.batch_size), end='')
LOGGER.write("Step %d | [%d/%d] | [%d/%d]" % (train_step, epoch+1, args.num_epoch, step, train_data_len//args.batch_size))
show_loss(x_log, y_log, t_log, log, LOGGER)
if step % args.display_info_step == 0 and step != 0:
args.training = False
save_path = saver.save(sess, exp_path+model_name, global_step=train_step)
print("Model saved in file: %s" % save_path)
# print("============= Show Encoder ===============")
# model.show_encoder(sess, x_enc_inp[-1], x_dec_inp[-1], LOGGER)
# print("============= Show Decoder ===============")
# model.show_decoder(sess, y_enc_inp[-1], y_dec_inp[-1], LOGGER)
# print("============= Show Sample ===============")
for i in range(3):
model.show_sample(sess, x_enc_inp[i], y_dec_out[i], LOGGER)
LOGGER.flush()
args.training = True