model = TextCNN(config)
model.build()
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=sess_config)
init = tf.global_variables_initializer()
sess.run(init)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and tf.train.get_checkpoint_state(checkpoint_dir):
model.restore(sess, ckpt.model_checkpoint_path)
else:
print("no checkpoint saved")
exit()
test_data = list(zip(test_docs, seq_len, test_labels))
batches = data_util.batch_loader(test_data, 32)
total_preds = []
for batch in batches:
batch_docs, batch_seq_len, batch_labels = zip(*batch)
preds = model.predict(sess, batch_docs, batch_seq_len)
total_preds = np.concatenate([total_preds, preds], axis=0)
test_acc = np.mean(total_preds == test_labels)
rmse = np.mean(np.square(total_preds - test_labels), axis=0)
rmse = np.sqrt(rmse)
print("test accuracy : {}".format(test_acc))
print(rmse)
reviews = open(test_data_file, "r", encoding="utf-8").readlines()
reviews = list(map(lambda x: x.strip().replace("</s>", ""), reviews))
result = np.column_stack((total_preds, test_labels, reviews))
with open(result_file, "w", encoding="utf-8") as f:
fw = csv.writer(f)
model.embedding_matrix.assign(pretrained_embedding)
del pretrained_embedding
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print("Reload model paramters")
model.restore(sess, ckpt.model_checkpoint_path)
else:
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
print("create new model parameter")
sess.run(tf.global_variables_initializer())
step = None
best_loss = 1e10
no_improv = 0
for epoch in range(num_epochs):
batches = batch_loader(list(zip(docs, sums)), batch_size)
dev_batches = batch_loader(list(zip(dev_docs, dev_sums)), batch_size)
for batch in batches:
batch_source, batch_target = zip(*batch)
batch_encoder_input, batch_encoder_len, \
batch_decoder_input, batch_decoder_len = \
make_array_format(batch_source, batch_target)
batch_loss, summary_op = model.train(
sess,
encoder_inputs=batch_encoder_input,
encoder_length=batch_encoder_len,
decoder_inputs=batch_decoder_input,
decoder_length=batch_decoder_len,
dropout=dropout)
step = model.global_step.eval(sess)
print("epoch : {} step : {}, loss : {}".format(
sess.run(init)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print("reload model")
model.restore(sess, ckpt.model_checkpoint_path)
else:
print("start from scratch")
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
print("start training")
best_acc = 0
for i in range(config.num_epochs):
epoch = i + 1
train_data = list(zip(train_docs, train_seq_len, train_labels))
batches = data_util.batch_loader(train_data, batch_size, shuffle=True)
for batch in batches:
batch_docs, batch_seq_len, batch_labels = zip(*batch)
step, loss, acc, rmse = model.train(sess, batch_docs, batch_seq_len,
batch_labels, config.dropout)
print("epoch: {}, loss: {}, acc: {}".format(epoch, loss, acc))
del train_data
dev_data = list(zip(dev_docs, dev_seq_len, dev_labels))
dev_batches = data_util.batch_loader(dev_data, batch_size)
total_preds = []
for dev_batch in dev_batches:
batch_docs, batch_seq_len, batch_labels = zip(*dev_batch)
preds = model.predict(sess, batch_docs, batch_seq_len, dropout=1.0)
total_preds = np.concatenate([total_preds, preds], axis=0)
dev_acc = np.mean(total_preds == dev_labels)
print("dev accuracy : {}".format(dev_acc))
state_size=state_size,
decoder_vocab_size=max_vocab_size,
filter_width=filter_width,
zoneout=zoneout,
attention_hidden_size=attention_hidden_size,
mode=mode,
beam_depth=beam_depth,
learning_rate=learning_rate,
max_iter=summary_len)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print("Reload model parameters")
model.restore(sess, ckpt.model_checkpoint_path)
else:
exit()
batches = data_util.batch_loader(docs, batch_size)
fw = open(summary_file, "w", encoding="utf-8")
for idx, batch in enumerate(batches):
batch_encoder_input, batch_encoder_len = data_util.make_array_format(
batch)
# if BeamSearchDecoder : [batch, max_time_step, beam_Depth]
batch_summaries = model.predict(sess, batch_encoder_input,
batch_encoder_len)
# iterate every example, get best beam search result
for summary in batch_summaries:
best_summary = summary[:, 0]
# convert indices to corresponding tokens
summary_tokens = data_util.map_idx2tok(best_summary, vocab)
# remove EOS token
try:
stop_index = summary_tokens.index(data_util.ID_EOS)
def main():
config = Config()
vocab = Vocab(config.dict_file)
tq, tc, q, c, a_s, a_t = load_data(config.train_query_file,
config.train_context_file,
config.train_tables_file,
config.train_answer_file, vocab,
config.debug)
#dev_q, dev_c, dev_s, dev_a_s, dev_a_t = load_data(config.dev_query_file, config.dev_context_file, config.dev_tables_file, config.dev_answer_file, vocab, config.debug)
max_q_len = max_len(q)
max_c_len = max_len(c)
max_a_len = max_len(a_s)
train_data = list(zip(tq, tc, q, c, a_s, a_t))
#dev_data = list(zip(dev_q, dev_c, dev_s, dev_a_s, dev_a_t))
wiki = Wiki(config)
wiki.build_model()
best_score = 0
print("start")
for i in range(config.num_epochs):
epoch = i + 1
print(epoch)
batches = batch_loader(train_data, config.batch_size, shuffle=False)
for batch in batches:
batch_tq, batch_tc, batch_q, batch_c, batch_a_s, batch_a_t = zip(
*batch)
question_lengths, padded_q = zero_padding(batch_q, max_q_len)
context_lengths, padded_c = zero_padding(batch_c, max_c_len)
answer_start_lengths, padded_a_s = zero_padding(
batch_a_s, max_a_len)
# loss, acc, pred, step = wiki.train(padded_q, question_lengths, padded_c, context_lengths,
# padded_a_s, answer_start_lengths, config.dropout)
loss, step = wiki.train(padded_q, question_lengths, padded_c,
context_lengths, padded_a_s, batch_a_t,
answer_start_lengths, max_q_len, max_c_len,
max_a_len, config.dropout)
# train_batch_acc, train_batch_em, train_batch_loss = wiki.eval(padded_q, question_lengths, padded_c,
# context_lengths,
# padded_s, sequence_lengths,
# sentence_lengths, batch_s_idx,
# batch_ans, batch_spans)
train_batch_loss = wiki.eval(padded_q, question_lengths, padded_c,
context_lengths, padded_a_s,
batch_a_t, answer_start_lengths,
max_q_len, max_c_len, max_a_len,
config.dropout)
print(loss)
print(train_batch_loss)
print("done")
# dev_em = np.mean(total_em)
# dev_acc = np.mean(total_acc)
# dev_loss = np.mean(total_loss)
# wiki.write_summary(dev_acc, dev_em, dev_loss, mode="dev")
# wiki.write_summary(train_batch_acc, train_batch_em, train_batch_loss, mode="train")
# print("after %d step, dev_em:%.2f" % (step, dev_em))
# if dev_em > best_score:
# best_score = dev_em
# print("new score! em: %.2f, acc:%.2f" % (dev_em, dev_acc))
# wiki.save_session(config.dir_model)
print("end")