DL之Attention-ED：基於TF NMT利用帶有Attention的 ED模型訓練(中英文平行語料庫)實現將英文翻譯為中文的LSTM翻譯模型過程全記錄

測試輸出結果

模型監控

1、SCALARS

clipped_gradient
grad_norm
train_loss
dev_bleu		dev_ppl
lr_1		test_bleu
test_ppl		train_ppl

2、IMAGES
attention_images_1/image/0   step 6,000

3、GRAPHS

設計思路

訓練過程全記錄

開始訓練
# Job id 0
# Loading hparams from tmp/nmt_model_zh\hparams
  saving hparams to tmp/nmt_model_zh\hparams
  saving hparams to tmp/nmt_model_zh\best_bleu\hparams
  attention=scaled_luong
  attention_architecture=standard
  batch_size=128
  beam_width=0
  best_bleu=0
  best_bleu_dir=tmp/nmt_model_zh\best_bleu
  bpe_delimiter=None
  colocate_gradients_with_ops=True
  decay_factor=0.98
  decay_steps=10000
  dev_prefix=tmp/nmt_zh/dev
  dropout=0.2
  encoder_type=uni
  eos=</s>
  epoch_step=0
  forget_bias=1.0
  infer_batch_size=32
  init_op=uniform
  init_weight=0.1
  learning_rate=1.0
  length_penalty_weight=0.0
  log_device_placement=False
  max_gradient_norm=5.0
  max_train=0
  metrics=['bleu']
  num_buckets=5
  num_embeddings_partitions=0
  num_gpus=1
  num_layers=3
  num_residual_layers=0
  num_train_steps=200000
  num_units=256
  optimizer=sgd
  out_dir=tmp/nmt_model_zh
  pass_hidden_state=True
  random_seed=None
  residual=False
  share_vocab=False
  sos=<s>
  source_reverse=False
  src=en
  src_max_len=50
  src_max_len_infer=None
  src_vocab_file=tmp/nmt_zh/vocab.en
  src_vocab_size=35028
  start_decay_step=0
  steps_per_external_eval=None
  steps_per_stats=100
  test_prefix=tmp/nmt_zh/test
  tgt=zh
  tgt_max_len=50
  tgt_max_len_infer=None
  tgt_vocab_file=tmp/nmt_zh/vocab.zh
  tgt_vocab_size=53712
  time_major=True
  train_prefix=tmp/nmt_zh/train
  unit_type=lstm
  vocab_prefix=tmp/nmt_zh/vocab
# creating train graph ...
  num_layers = 3, num_residual_layers=0
  cell 0  LSTM, forget_bias=1WARNING:tensorflow:From 
Instructions for updating:
This class is deprecated, please use tf.nn.rnn_cell.LSTMCell, which supports all the feature this cell currently has. Please replace the existing code with tf.nn.rnn_cell.LSTMCell(name='basic_lstm_cell').
  DropoutWrapper, dropout=0.2   DeviceWrapper, device=/gpu:0
  cell 1  LSTM, forget_bias=1  DropoutWrapper, dropout=0.2   DeviceWrapper, device=/gpu:0
  cell 2  LSTM, forget_bias=1  DropoutWrapper, dropout=0.2   DeviceWrapper, device=/gpu:0
  cell 0  LSTM, forget_bias=1  DropoutWrapper, dropout=0.2   DeviceWrapper, device=/gpu:0
  cell 1  LSTM, forget_bias=1  DropoutWrapper, dropout=0.2   DeviceWrapper, device=/gpu:0
  cell 2  LSTM, forget_bias=1  DropoutWrapper, dropout=0.2   DeviceWrapper, device=/gpu:0
  start_decay_step=0, learning_rate=1, decay_steps 10000,decay_factor 0.98
# Trainable variables
  embeddings/encoder/embedding_encoder:0, (35028, 256),
  embeddings/decoder/embedding_decoder:0, (53712, 256),
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_2/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_2/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/memory_layer/kernel:0, (256, 256),
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_0/basic_lstm_cell/kernel:0, (768, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_0/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_1/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_1/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_2/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_2/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/luong_attention/attention_g:0, (), /device:GPU:0
  dynamic_seq2seq/decoder/attention/attention_layer/kernel:0, (512, 256), /device:GPU:0
  dynamic_seq2seq/decoder/output_projection/kernel:0, (256, 53712), /device:GPU:0
# creating eval graph ...
  num_layers = 3, num_residual_layers=0
  cell 0  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 1  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 2  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 0  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 1  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 2  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  start_decay_step=0, learning_rate=1, decay_steps 10000,decay_factor 0.98
# Trainable variables
  embeddings/encoder/embedding_encoder:0, (35028, 256),
  embeddings/decoder/embedding_decoder:0, (53712, 256),
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_2/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_2/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/memory_layer/kernel:0, (256, 256),
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_0/basic_lstm_cell/kernel:0, (768, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_0/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_1/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_1/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_2/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_2/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/luong_attention/attention_g:0, (), /device:GPU:0
  dynamic_seq2seq/decoder/attention/attention_layer/kernel:0, (512, 256), /device:GPU:0
  dynamic_seq2seq/decoder/output_projection/kernel:0, (256, 53712), /device:GPU:0
# creating infer graph ...
  num_layers = 3, num_residual_layers=0
  cell 0  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 1  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 2  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 0  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 1  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  cell 2  LSTM, forget_bias=1  DeviceWrapper, device=/gpu:0
  start_decay_step=0, learning_rate=1, decay_steps 10000,decay_factor 0.98
# Trainable variables
  embeddings/encoder/embedding_encoder:0, (35028, 256),
  embeddings/decoder/embedding_decoder:0, (53712, 256),
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_2/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_2/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/memory_layer/kernel:0, (256, 256),
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_0/basic_lstm_cell/kernel:0, (768, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_0/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_1/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_1/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_2/basic_lstm_cell/kernel:0, (512, 1024), /device:GPU:0
  dynamic_seq2seq/decoder/attention/multi_rnn_cell/cell_2/basic_lstm_cell/bias:0, (1024,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/luong_attention/attention_g:0, (), /device:GPU:0
  dynamic_seq2seq/decoder/attention/attention_layer/kernel:0, (512, 256), /device:GPU:0
  dynamic_seq2seq/decoder/output_projection/kernel:0, (256, 53712),
# log_file=tmp/nmt_model_zh\log_1539923931
2018-10-19 12:38:51.109178: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
  created train model with fresh parameters, time 0.94s
  created infer model with fresh parameters, time 0.50s
  # 215
    src: second , the human capital , knowledge reserves , and other high - grade economic elements have formed the comparative advantages of the province in terms of technical innovation and the development of new high - tech industries .
    ref: 其二 , 人力 資本 , 知識 儲備 等 高階 經濟 要素 , 形成 了 吉林省 技術 創新 和 發展 高新技術 產業 的 比較 優勢 .
    nmt: b'\xe9\xaa\x87 \xe6\x9d\x83\xe8\xa1\xa1 \xe6\x9d\x83\xe8\xa1\xa1 \xe6\x9d\x83\xe8\xa1\xa1 \xe6\x9d\x83\xe8\xa1\xa1 \xe6\x9d\x83\xe8\xa1\xa1 \xe5\x8d\x8a\xe8\xbe\xb9\xe5\xa4\xa9 \xe5\x8d\x8a\xe8\xbe\xb9\xe5\xa4\xa9 \xe5\x8d\x8a\xe8\xbe\xb9\xe5\xa4\xa9 \xe8\xb4\xa7\xe6\xac\xbe \xe8\xb4\xa7\xe6\xac\xbe \xe8\xb4\xa7\xe6\xac\xbe \xe8\xb4\xa7\xe6\xac\xbe \xe8\xb4\xa7\xe6\xac\xbe \xe6\xb1\x9f\xe5\x8d\xab\xe5\x9b\xbd \xe6\xb1\x9f\xe5\x8d\xab\xe5\x9b\xbd \xe6\xb1\x9f\xe5\x8d\xab\xe5\x9b\xbd \xe6\xb1\x9f\xe5\x8d\xab\xe5\x9b\xbd \xe6\xa1\x91\xe4\xb9\x85\xe7\xbe\x8e \xe6\xa1\x91\xe4\xb9\x85\xe7\xbe\x8e \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe9\x93\xb8\xe6\x88\x90\xe5\xa4\xa7\xe9\x94\x99 \xe4\xbe\x9b\xe9\x94\x80 \xe6\x9d\x8e\xe4\xb8\xbd\xe8\xbe\x89 \xe6\x9d\x8e\xe4\xb8\xbd\xe8\xbe\x89 \xe6\x9d\x8e\xe4\xb8\xbd\xe8\xbe\x89 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x97\xa5\xe6\x9c\xac\xe6\xb5\xb7 \xe6\x9d\x8e\xe6\xb0\xb8\xe5\x88\x9d \xe6\x9d\x8e\xe6\xb0\xb8\xe5\x88\x9d \xe6\x9d\x8e\xe6\xb0\xb8\xe5\x88\x9d \xe6\x9d\x8e\xe6\xb0\xb8\xe5\x88\x9d \xe6\x9d\x8e\xe6\xb0\xb8\xe5\x88\x9d \xe5\x9b\xb4\xe6\xad\xbc \xe5\x9b\xb4\xe6\xad\xbc \xe5\x9b\xb4\xe6\xad\xbc \xe5\x9b\xb4\xe6\xad\xbc \xe5\x9b\xb4\xe6\xad\xbc \xe6\x9d\x8e\xe5\xae\x89\xe5\xb9\xb3 \xe6\x9d\x8e\xe5\xae\x89\xe5\xb9\xb3 \xe5\xae\x8b\xe8\xb1\xab \xe5\xae\x8b\xe8\xb1\xab \xe5\xae\x8b\xe8\xb1\xab \xe5\xae\x8b\xe8\xb1\xab \xe5\xae\x8b\xe8\xb1\xab \xe5\xae\x8b\xe8\xb1\xab \xe5\xae\x8b\xe8\xb1\xab \xe6\xaf\x8f\xe5\x86\xb5\xe6\x84\x88\xe4\xb8\x8b \xe5\xbd\xb1\xe5\x93\x8d\xe5\x88\xb0 \xe5\xbd\xb1\xe5\x93\x8d\xe5\x88\xb0 \xe5\xbd\xb1\xe5\x93\x8d\xe5\x88\xb0 \xe5\xbd\xb1\xe5\x93\x8d\xe5\x88\xb0 \xe5\xbd\xb1\xe5\x93\x8d\xe5\x88\xb0 \xe7\xbb\xa7\xe4\xbd\x8d \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbe\x9d\xe9\x82\xa3\xe5\x90\x90\xe6\x8b\x89 \xe4\xbb\xa5\xe9\x82\xbb\xe4\xb8\xba\xe5\xa3\x91'
  created eval model with fresh parameters, time 0.61s
2018-10-19 12:38:57.783584: W tensorflow/core/framework/allocator.cc:113] Allocation of 641750976 exceeds 10% of system memory.
2018-10-19 12:39:01.312543: W tensorflow/core/framework/allocator.cc:113] Allocation of 812769984 exceeds 10% of system memory.
2018-10-19 12:39:05.244676: W tensorflow/core/framework/allocator.cc:113] Allocation of 660013056 exceeds 10% of system memory.
2018-10-19 12:39:08.661067: W tensorflow/core/framework/allocator.cc:113] Allocation of 800523648 exceeds 10% of system memory.
  eval dev: perplexity 53779.34, time 16s, Fri Oct 19 12:39:11 2018.
2018-10-19 12:39:14.135415: W tensorflow/core/framework/allocator.cc:113] Allocation of 641750976 exceeds 10% of system memory.
  eval test: perplexity 53779.35, time 17s, Fri Oct 19 12:39:29 2018.
2018-10-19 12:39:29.544562: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 12:39:29.544589: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 12:39:29.544778: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  created infer model with fresh parameters, time 0.50s
# Start step 0, lr 1, Fri Oct 19 12:39:29 2018
# Init train iterator, skipping 0 elements
  global step 100 lr 1 step-time 12.18s wps 0.61K ppl 370383.80 bleu 0.00
  global step 200 lr 1 step-time 11.44s wps 0.64K ppl 69384.04 bleu 0.00
  global step 300 lr 1 step-time 11.49s wps 0.65K ppl 22598.76 bleu 0.00
  global step 400 lr 1 step-time 11.55s wps 0.64K ppl 14178.53 bleu 0.00
  global step 500 lr 1 step-time 11.50s wps 0.65K ppl 10184.07 bleu 0.00
  global step 600 lr 1 step-time 12.13s wps 0.61K ppl 6656.94 bleu 0.00
  global step 700 lr 1 step-time 13.51s wps 0.55K ppl 2673.24 bleu 0.00
# Finished an epoch, step 785. Perform external evaluation
2018-10-19 15:16:44.846427: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 15:16:44.846427: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 15:16:44.867191: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
  created infer model with fresh parameters, time 1.07s
  # 64
    src: under the leadership of the state civil air defense office , the center will assume overall work related to china 's civil air defense engineering as well as nuclear and chemical protection .
    ref: 這個 中心 將 在 國家 人民 防空 辦公室 領導 下 , 擔負 我國 人防 工程 核 化 防護 的 全面 工作 .
    nmt: b'\xe6\x96\xb9\xe5\x85\xb3\xe6\x96\xbc 2020\xe5\xb9\xb4 \xe9\x9f\xa6\xe5\xa4\x9a \xe9\x9f\xa6\xe5\xa4\x9a 19-02-2000 19-02-2000 19-02-2000 19-02-2000 \xe7\xb3\xbb\xe7\xbb\x9f\xe5\x9c\xb0 \xe7\xb3\xbb\xe7\xbb\x9f\xe5\x9c\xb0 \xe5\x97\x85\xe5\x87\xba \xe5\x97\x85\xe5\x87\xba \xe5\x97\x85\xe5\x87\xba \xe6\xac\xa1\xe4\xbe\x8d \xe6\xac\xa1\xe4\xbe\x8d \xe6\xac\xa1\xe4\xbe\x8d \xe6\xac\xa1\xe4\xbe\x8d \xe6\xac\xa1\xe4\xbe\x8d \xe5\x8f\x8d\xe8\xb4\xaa\xe5\xb1\x80 \xe5\x8f\x8d\xe8\xb4\xaa\xe5\xb1\x80 \xe5\x8f\x8d\xe8\xb4\xaa\xe5\xb1\x80 \xe8\x96\x84\xe2\x85\xb0\xe5\x82\xb2 \xe8\x96\x84\xe2\x85\xb0\xe5\x82\xb2 \xe8\x96\x84\xe2\x85\xb0\xe5\x82\xb2 \xe8\x96\x84\xe2\x85\xb0\xe5\x82\xb2 \xe5\x8f\x8c\xe5\xba\xa7 \xe5\x8f\x8c\xe5\xba\xa7 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe6\x96\xbd\xe5\x90\x9b\xe7\x8e\x89 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe9\x99\xa2 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\xa4\xa7\xe4\xbd\xbf\xe7\xba\xa7 \xe5\x98\x89 \xe5\x98\x89 \xe5\x98\x89 \xe5\x98\x89 \xe5\x98\x89 \xe5\x98\x89 \xe5\x98\x89 \xe6\xb1\x82\xe6\x95\x99 \xe6\xb1\x82\xe6\x95\x99 \xe6\xb1\x82\xe6\x95\x99'
2018-10-19 15:16:46.110426: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 15:16:46.110445: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 15:16:46.110444: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  created infer model with fresh parameters, time 0.88s
  global step 800 lr 1 step-time 12.50s wps 0.57K ppl 2048.36 bleu 0.00
  global step 900 lr 1 step-time 13.74s wps 0.54K ppl 1883.18 bleu 0.00
  global step 1000 lr 1 step-time 14.76s wps 0.50K ppl 1583.37 bleu 0.00
# Save eval, global step 1000
2018-10-19 16:07:39.926588: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 16:07:39.926612: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 16:07:39.926588: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-1000, time 1.62s
  # 125
    src: the united states suffered heavy casualties in the korean war in the 1950 's and the vietnam war of the 1960 's , especially the latter , and this directly caused strong antiwar sentiment within the united states and played a very great driving role in the ultimate and undoubted us defeat in the war .
    ref: 50年 代 的 朝鮮 戰爭 , 60年 代 的 越南 戰爭 , 尤其是 後者 , 美軍 大量 傷亡 , 直接 造成 了 美國 國內 強烈 的 反戰 情緒 , 對 美國 最終 戰敗 無疑 起到 了 巨大 的 推動 作用 .
    nmt: b'\xe4\xbb\x96 \xe8\xaf\xb4 , \xe4\xbb\x96 \xe5\x9c\xa8 \xe4\xbb\x96 , \xe4\xbb\x96 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 . </s>'
2018-10-19 16:07:41.873322: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 16:07:41.873376: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded eval model parameters from tmp/nmt_model_zh\translate.ckpt-1000, time 1.54s
  eval dev: perplexity 1055.18, time 19s, Fri Oct 19 16:08:01 2018.
  eval test: perplexity 1055.18, time 22s, Fri Oct 19 16:08:24 2018.
  global step 1100 lr 1 step-time 14.13s wps 0.53K ppl 1435.07 bleu 0.00
  global step 1200 lr 1 step-time 13.73s wps 0.54K ppl 1276.74 bleu 0.00
  global step 1300 lr 1 step-time 12.87s wps 0.58K ppl 1170.78 bleu 0.00
  global step 1400 lr 1 step-time 14.89s wps 0.50K ppl 1066.19 bleu 0.00
  global step 1500 lr 1 step-time 15.34s wps 0.48K ppl 1046.93 bleu 0.00
# Finished an epoch, step 1570. Perform external evaluation
2018-10-19 18:26:54.761219: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 18:26:54.784158: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 18:26:54.784232: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-1000, time 3.16s
  # 21
    src: this saying figuratively shows that this is indeed an absurd cycle .
    ref: 這個 說法 以 擬人化 手法 生動 地 表明 , 這 的確 是 個 怪圈 .
    nmt: b'\xe4\xbb\x96 \xe8\xaf\xb4 , \xe4\xbb\x96 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 \xe7\x9a\x84 . </s>'
2018-10-19 18:26:55.501112: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 18:26:55.501112: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 18:26:55.501131: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-1000, time 0.59s
# External evaluation, global step 1000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 34s, Fri Oct 19 18:27:29 2018.
  bleu dev: 0.0
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 1000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 33s, Fri Oct 19 18:28:04 2018.
  bleu test: 0.0
  saving hparams to tmp/nmt_model_zh\hparams
  global step 1600 lr 1 step-time 17.33s wps 0.41K ppl 935.22 bleu 0.00
  global step 1700 lr 1 step-time 17.57s wps 0.43K ppl 914.06 bleu 0.00
  global step 1800 lr 1 step-time 15.53s wps 0.48K ppl 833.33 bleu 0.00
  global step 1900 lr 1 step-time 14.46s wps 0.51K ppl 784.47 bleu 0.00
  global step 2000 lr 1 step-time 15.27s wps 0.49K ppl 730.18 bleu 0.00
# Save eval, global step 2000
2018-10-19 20:21:36.887237: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 20:21:36.887234: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 20:21:36.887234: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-2000, time 2.08s
  # 171
    src: the sino - dprk traditional friendship created and nurtured personally by chairman mao zedong , premier zhou enlai , comrade deng xiaoping , president kim il - song and other leaders of the older generation have withstood historical tests and have taken deep roots in the hearts of peoples of the two countries .
    ref: 毛澤東 主席 , 周恩來 總理 , 鄧小平 同志 和 金日成 主席 等 老一輩 領導人 親手 締造 和 培育 的 中朝 傳統 友誼 經受 了 歷史 的 考驗 , 已 深深 紮根 於 兩國 人民 心中 .
    nmt: b'\xe4\xb8\xad\xe5\x9b\xbd \xe6\x98\xaf \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe5\x92\x8c \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe5\x92\x8c \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe5\x92\x8c \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe5\x92\x8c \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe5\x92\x8c \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 . </s>'
2018-10-19 20:21:37.898203: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 20:21:37.898825: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded eval model parameters from tmp/nmt_model_zh\translate.ckpt-2000, time 0.49s
  eval dev: perplexity 694.89, time 26s, Fri Oct 19 20:22:04 2018.
  eval test: perplexity 694.89, time 28s, Fri Oct 19 20:22:32 2018.
  global step 2100 lr 1 step-time 14.39s wps 0.51K ppl 689.72 bleu 0.00
  global step 2200 lr 1 step-time 13.70s wps 0.54K ppl 649.35 bleu 0.00
  global step 2300 lr 1 step-time 13.78s wps 0.54K ppl 609.52 bleu 0.00
# Finished an epoch, step 2355. Perform external evaluation
2018-10-19 21:44:15.494641: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 21:44:15.494641: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 21:44:15.494654: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-2000, time 0.75s
  # 277
    src: the complete reunification of the motherland is the trend of the times .
    ref: 祖國 的 完全 統一 是 大勢所趨 .
    nmt: b'\xe4\xb8\xad\xe5\x9b\xbd \xe6\x98\xaf \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 . </s>'
2018-10-19 21:44:16.136067: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-19 21:44:16.136100: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-19 21:44:16.136067: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-2000, time 0.57s
# External evaluation, global step 2000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 44s, Fri Oct 19 21:45:00 2018.
  bleu dev: 0.0
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 2000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 42s, Fri Oct 19 21:45:43 2018.
  bleu test: 0.0
  saving hparams to tmp/nmt_model_zh\hparams
  global step 2400 lr 1 step-time 13.68s wps 0.52K ppl 578.41 bleu 0.00
  global step 2500 lr 1 step-time 14.21s wps 0.52K ppl 532.20 bleu 0.00
  global step 2600 lr 1 step-time 14.41s wps 0.52K ppl 522.94 bleu 0.00
  global step 2700 lr 1 step-time 13.97s wps 0.53K ppl 477.39 bleu 0.00
  global step 2800 lr 1 step-time 11.82s wps 0.63K ppl 458.58 bleu 0.00
  global step 2900 lr 1 step-time 11.21s wps 0.66K ppl 426.22 bleu 0.00
  global step 3000 lr 1 step-time 11.19s wps 0.66K ppl 408.34 bleu 0.00
# Save eval, global step 3000
2018-10-20 00:04:43.141117: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 00:04:43.141117: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 00:04:43.162239: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-3000, time 1.25s
  # 22
    src: we will seize the historic opportunity of the development of the western region and earnestly study new ideas , mechanism , and measures suited to tibet 's characteristics .
    ref: 我們 要 抓住 西部 大 開發 的 歷史 機遇 , 認真 研究 適應 西藏 特點 的 新 思路 , 新 機制 , 新 措施 .
    nmt: b'\xe5\x9c\xa8 \xe8\xbf\x99 \xe4\xb8\x80 \xe9\x97\xae\xe9\xa2\x98 \xe4\xb8\x8a , \xe5\x9c\xa8 \xe7\xbb\x8f\xe6\xb5\x8e \xe4\xb8\x8a , \xe7\xbb\x8f\xe6\xb5\x8e \xe5\x8f\x91\xe5\xb1\x95 , \xe5\x8f\x91\xe5\xb1\x95 \xe5\x8f\x91\xe5\xb1\x95 , \xe5\x8f\x91\xe5\xb1\x95 \xe5\x8f\x91\xe5\xb1\x95 , \xe5\x8f\x91\xe5\xb1\x95 \xe5\x8f\x91\xe5\xb1\x95 . </s>'
2018-10-20 00:04:43.760353: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 00:04:43.760378: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded eval model parameters from tmp/nmt_model_zh\translate.ckpt-3000, time 0.45s
  eval dev: perplexity 387.54, time 15s, Sat Oct 20 00:04:59 2018.
  eval test: perplexity 387.54, time 17s, Sat Oct 20 00:05:17 2018.
  global step 3100 lr 1 step-time 11.16s wps 0.66K ppl 385.81 bleu 0.00
# Finished an epoch, step 3140. Perform external evaluation
2018-10-20 00:30:34.679555: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 00:30:34.680522: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 00:30:34.680586: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-3000, time 0.56s
  # 232
    src: because taiwan 's new leader deliberately obscured the one - china principle and denied the " 1992 consensus " of the association for relations across the taiwan strait and the strait exchange foundation , the two sides have been unable to ease their tense relations and achieve a breakthrough in their political impasse .
    ref: 過去 一 年 , 臺灣 政局 發生 了 重大 變化 , 由於 臺灣 新 領導人 刻意 模糊 一個 中國 的 原則 , 否認 兩岸 兩會 " 九二 共識 " , 致使 兩岸 關係 不能 緩和 , 政治 僵局 難以 打破 , 同時 造成 島內 社會 混亂 , 經濟 動盪 , 百姓 不安 .
    nmt: b'\xe5\x9c\xa8 \xe8\xbf\x99 \xe4\xb8\x80 \xe9\x97\xae\xe9\xa2\x98 \xe4\xb8\x8a , \xe5\x9c\xa8 \xe7\xbb\x8f\xe6\xb5\x8e \xe4\xb8\x8a , \xe4\xb8\xad\xe5\x9b\xbd \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 \xe7\xad\x89 \xe7\xbb\x8f\xe6\xb5\x8e \xe5\x90\x88\xe4\xbd\x9c , \xe4\xbf\x83\xe8\xbf\x9b \xe7\xbb\x8f\xe6\xb5\x8e \xe5\x8f\x91\xe5\xb1\x95 , \xe4\xbf\x83\xe8\xbf\x9b \xe7\xbb\x8f\xe6\xb5\x8e \xe5\x8f\x91\xe5\xb1\x95 . </s>'
2018-10-20 00:30:35.313256: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 00:30:35.313274: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 00:30:35.313289: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-3000, time 0.38s
# External evaluation, global step 3000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 30s, Sat Oct 20 00:31:05 2018.
  bleu dev: 0.2
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 3000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 28s, Sat Oct 20 00:31:37 2018.
  bleu test: 0.2
  saving hparams to tmp/nmt_model_zh\hparams
  global step 3200 lr 1 step-time 10.83s wps 0.66K ppl 357.57 bleu 0.17
  global step 3300 lr 1 step-time 11.15s wps 0.66K ppl 342.12 bleu 0.17
  global step 3400 lr 1 step-time 11.37s wps 0.66K ppl 325.57 bleu 0.17
  global step 3500 lr 1 step-time 11.14s wps 0.66K ppl 312.20 bleu 0.17
  global step 3600 lr 1 step-time 11.14s wps 0.66K ppl 307.63 bleu 0.17
  global step 3700 lr 1 step-time 11.28s wps 0.66K ppl 296.23 bleu 0.17
  global step 3800 lr 1 step-time 11.14s wps 0.66K ppl 278.77 bleu 0.17
  global step 3900 lr 1 step-time 11.11s wps 0.67K ppl 275.24 bleu 0.17
# Finished an epoch, step 3925. Perform external evaluation
2018-10-20 02:57:30.646778: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 02:57:30.646780: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 02:57:30.646874: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-3000, time 0.54s
  # 120
    src: in china , socialism has succeeded in greatly emancipating and developing productive forces in society .
    ref: 社會主義 中國 的 社會 生產力 得到 極大 解放 和 發展 .
    nmt: b'\xe5\x9c\xa8 \xe8\xbf\x99 \xe4\xb8\x80 \xe9\x97\xae\xe9\xa2\x98 \xe4\xb8\x8a , \xe6\x88\x91\xe4\xbb\xac \xe8\xa6\x81 \xe5\x8a\xa0\xe5\xbc\xba \xe5\x8f\x91\xe5\xb1\x95 \xe5\x92\x8c \xe5\x8f\x91\xe5\xb1\x95 . </s>'
2018-10-20 02:57:31.074731: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 02:57:31.074738: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 02:57:31.074738: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-3000, time 0.35s
# External evaluation, global step 3000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 29s, Sat Oct 20 02:58:01 2018.
  bleu dev: 0.2
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 3000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 30s, Sat Oct 20 02:58:31 2018.
  bleu test: 0.2
  saving hparams to tmp/nmt_model_zh\hparams
  global step 4000 lr 1 step-time 10.69s wps 0.66K ppl 252.99 bleu 0.17
# Save eval, global step 4000
2018-10-20 03:12:28.024198: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 03:12:28.024218: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 03:12:28.024223: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-4000, time 0.36s
  # 224
    src: the ninth guangdong provincial people 's congress concluded its meeting this morning . pursuant to the proposal of guangdong provincial higher people 's court president lu botao , ling qiman and li yifeng were appointed vice presidents of the provincial higher people 's court .
    ref: 今天 上午 閉幕 的 省九屆 人大 常委會 , 還 根據 廣東省 高階 人民 法院 院長 呂伯濤 的 提請 , 任命 凌祁漫 李毅峰 為 廣東省 高院 副 院長 .
    nmt: b'\xe5\x9c\xa8 \xe8\xb0\x88\xe5\x88\xb0 \xe4\xb8\xad\xe5\x9b\xbd \xe5\x85\xb1\xe4\xba\xa7\xe5\x85\x9a , \xe5\x9b\xbd\xe5\x8a\xa1\xe9\x99\xa2 \xe6\x80\xbb\xe7\x90\x86 \xe6\x9c\xb1\xe9\x8e\x94\xe5\x9f\xba \xe4\xbb\x8a\xe5\xa4\xa9 \xe4\xb8\x8a\xe5\x8d\x88 , \xe5\x9b\xbd\xe5\x8a\xa1\xe9\x99\xa2 \xe5\x89\xaf\xe6\x80\xbb\xe7\x90\x86 \xe9\x92\xb1\xe5\x85\xb6\xe7\x90\x9b , \xe5\x9b\xbd\xe5\x8a\xa1\xe9\x99\xa2 , \xe4\xb9\xa6\xe8\xae\xb0\xe5\xa4\x84 , \xe4\xb9\xa6\xe8\xae\xb0\xe5\xa4\x84 , \xe4\xb9\xa6\xe8\xae\xb0\xe5\xa4\x84 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 , \xe6\x9b\xbe\xe5\xba\x86\xe7\xba\xa2 . </s>'
2018-10-20 03:12:28.632761: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 03:12:28.632820: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
  loaded eval model parameters from tmp/nmt_model_zh\translate.ckpt-4000, time 0.36s
  eval dev: perplexity 252.28, time 15s, Sat Oct 20 03:12:44 2018.
  eval test: perplexity 252.28, time 16s, Sat Oct 20 03:13:00 2018.
  global step 4100 lr 1 step-time 11.56s wps 0.64K ppl 246.02 bleu 0.17
  global step 4200 lr 1 step-time 11.29s wps 0.66K ppl 237.87 bleu 0.17
  global step 4300 lr 1 step-time 11.32s wps 0.66K ppl 233.55 bleu 0.17
  global step 4400 lr 1 step-time 11.36s wps 0.66K ppl 221.90 bleu 0.17
  global step 4500 lr 1 step-time 11.06s wps 0.66K ppl 224.83 bleu 0.17
  global step 4600 lr 1 step-time 11.32s wps 0.66K ppl 211.41 bleu 0.17
  global step 4700 lr 1 step-time 11.21s wps 0.66K ppl 213.31 bleu 0.17
# Finished an epoch, step 4710. Perform external evaluation
2018-10-20 05:26:06.693124: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 05:26:06.693124: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 05:26:06.693124: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-4000, time 0.84s
  # 33
    src: peng peiyun and he luli respectively attended the plenary meetings of the zhejiang and beijing delegations .
    ref: 彭 ( 佩 去 人 旁 加 王旁 ) 雲 參加 了 浙江 代表團 的 全體 會議 . 何魯麗 參加 了 北京 代表團 的 全體 會議 .
    nmt: b'\xe4\xb8\xad\xe5\x9b\xbd \xe6\x94\xbf\xe5\xba\x9c \xe5\x92\x8c \xe4\xba\xba\xe6\xb0\x91 , \xe4\xba\xba\xe6\xb0\x91 , \xe7\x9b\xb4\xe8\xbe\x96\xe5\xb8\x82 , \xe7\x9b\xb4\xe8\xbe\x96\xe5\xb8\x82 , \xe7\x9b\xb4\xe8\xbe\x96\xe5\xb8\x82 . </s>'
2018-10-20 05:26:07.132553: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 05:26:07.132553: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 05:26:07.132618: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-4000, time 0.37s
# External evaluation, global step 4000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 45s, Sat Oct 20 05:26:52 2018.
  bleu dev: 0.2
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 4000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 43s, Sat Oct 20 05:27:39 2018.
  bleu test: 0.2
  saving hparams to tmp/nmt_model_zh\hparams
  global step 4800 lr 1 step-time 10.88s wps 0.66K ppl 199.44 bleu 0.18
  global step 4900 lr 1 step-time 11.33s wps 0.66K ppl 192.33 bleu 0.18
  global step 5000 lr 1 step-time 11.41s wps 0.65K ppl 186.30 bleu 0.18
# Save eval, global step 5000
2018-10-20 06:22:32.135143: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 06:22:32.135150: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 06:22:32.135223: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-5000, time 0.76s
  # 119
    src: as a result they could only avoid the differences of opinion and delay the solutions .
    ref: 結果 只能 是 迴避 分歧 , 推遲 解決 .
    nmt: b'\xe4\xbb\x8e \xe6\xa0\xb9\xe6\x9c\xac\xe4\xb8\x8a \xe7\x9c\x8b , \xe6\x88\x91\xe4\xbb\xac \xe5\xbf\x85\xe9\xa1\xbb \xe5\x9d\x9a\xe6\x8c\x81 \xe4\xb8\x80\xe4\xb8\xaa \xe4\xb8\xad\xe5\x9b\xbd \xe7\x9a\x84 \xe5\x8e\x9f\xe5\x88\x99 . </s>'
2018-10-20 06:22:32.937075: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 06:22:32.937075: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
  loaded eval model parameters from tmp/nmt_model_zh\translate.ckpt-5000, time 0.94s
  eval dev: perplexity 188.49, time 15s, Sat Oct 20 06:22:49 2018.
  eval test: perplexity 188.49, time 17s, Sat Oct 20 06:23:06 2018.
2018-10-20 06:23:09.323110: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 06:23:09.323133: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 06:23:09.323159: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-5000, time 0.36s
  # 208
    src: by the middle of this century , we should build the western region into a new region enjoying economic prosperity , social progress , a good living environment , and nationality unity , and having green mountains and clean rivers .
    ref: 到 21世紀中葉 , 將 西部 地區 建成 一個 經濟 繁榮 , 社會 進步 , 生活 安定 , 民族團結 , 山川 秀美 的 新 西部 .
    nmt: b'\xe4\xbb\x8e \xe4\xbb\x8a\xe5\xb9\xb4 \xe4\xb8\x8b\xe5\x8d\x8a\xe5\xb9\xb4 \xe5\xbc\x80\xe5\xa7\x8b , \xe6\x88\x91\xe5\x9b\xbd \xe5\xb0\x86 \xe7\xbb\xa7\xe7\xbb\xad \xe5\x8a\xa0\xe5\xbc\xba \xe5\xaf\xb9 \xe8\xa5\xbf\xe9\x83\xa8 \xe5\x9c\xb0\xe5\x8c\xba \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 , \xe5\xb9\xb6 \xe8\xbf\x9b\xe4\xb8\x80\xe6\xad\xa5 \xe6\x8e\xa8\xe5\x8a\xa8 \xe4\xb8\xad\xe5\x9b\xbd \xe6\x94\xbf\xe5\xba\x9c \xe5\xaf\xb9 \xe4\xb8\xad\xe5\x9b\xbd \xe4\xba\xba\xe6\xb0\x91 \xe7\x9a\x84 \xe5\x88\xa9\xe7\x9b\x8a . </s>'
2018-10-20 06:23:09.887178: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 06:23:09.887162: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 06:23:09.887366: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-5000, time 0.77s
# External evaluation, global step 5000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 29s, Sat Oct 20 06:23:39 2018.
  bleu dev: 0.4
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 5000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 29s, Sat Oct 20 06:24:12 2018.
  bleu test: 0.4
  saving hparams to tmp/nmt_model_zh\hparams
  global step 5100 lr 1 step-time 11.38s wps 0.65K ppl 185.57 bleu 0.39
  global step 5200 lr 1 step-time 11.30s wps 0.65K ppl 178.95 bleu 0.39
  global step 5300 lr 1 step-time 11.18s wps 0.66K ppl 174.15 bleu 0.39
  global step 5400 lr 1 step-time 11.39s wps 0.66K ppl 173.02 bleu 0.39
# Finished an epoch, step 5495. Perform external evaluation
2018-10-20 07:57:03.000782: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 07:57:03.000804: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 07:57:03.000921: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-5000, time 0.82s
  # 6
    src: we should truly realize " six unifications , " especially the unification of fees to improve efficiency . we should never allow the practice of " ostensibly combining the three types of inspections while failing to do so in private , " and should truly combine the " three types of inspections . "
    ref: 真正 實現 " 六 統一 " , 特別是 統一 收費 和 提高 效率 , 決不能 出現 那種 " 明合暗 不合 , 面 合心 不合 " 的 現 像 , 把 " 三 檢 " 合一 真正 落實 下去 .
    nmt: b'\xe6\x88\x91\xe4\xbb\xac \xe8\xa6\x81 \xe6\x8a\x8a " \xe4\xb8\x89 \xe4\xbb\xa3\xe8\xa1\xa8 " \xe4\xb8\xba \xe6\xa0\xb8\xe5\xbf\x83 \xe7\x9a\x84 \xe6\x80\x9d\xe6\x83\xb3 \xe6\x94\xbf\xe6\xb2\xbb \xe5\xbb\xba\xe8\xae\xbe \xe4\xbd\x9c\xe4\xb8\xba \xe6\x96\xb0 \xe7\x9a\x84 \xe4\xbc\x9f\xe5\xa4\xa7 \xe7\x9a\x84 \xe4\xbc\x9f\xe5\xa4\xa7 \xe4\xbb\xbb\xe5\x8a\xa1 , \xe6\x88\x91\xe4\xbb\xac \xe5\x85\x9a \xe7\x9a\x84 \xe5\xbb\xba\xe8\xae\xbe \xe6\x98\xaf \xe4\xb8\x80\xe4\xb8\xaa \xe6\x96\xb0 \xe7\x9a\x84 \xe5\x8f\x91\xe5\xb1\x95 \xe6\x97\xb6\xe6\x9c\x9f , \xe6\x98\xaf \xe6\x88\x91\xe4\xbb\xac \xe5\x85\x9a \xe7\x9a\x84 \xe5\xbb\xba\xe8\xae\xbe \xe7\x9a\x84 \xe4\xbc\x9f\xe5\xa4\xa7 \xe4\xbb\xbb\xe5\x8a\xa1 . </s>'
2018-10-20 07:57:04.047105: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.en is already initialized.
2018-10-20 07:57:04.047147: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
2018-10-20 07:57:04.047113: I tensorflow/core/kernels/lookup_util.cc:376] Table trying to initialize from file tmp/nmt_zh/vocab.zh is already initialized.
  loaded infer model parameters from tmp/nmt_model_zh\translate.ckpt-5000, time 0.79s
# External evaluation, global step 5000
  decoding to output tmp/nmt_model_zh\output_dev.
  done, num sentences 400, time 30s, Sat Oct 20 07:57:35 2018.
  bleu dev: 0.4
  saving hparams to tmp/nmt_model_zh\hparams
# External evaluation, global step 5000
  decoding to output tmp/nmt_model_zh\output_test.
  done, num sentences 400, time 30s, Sat Oct 20 07:58:06 2018.
  bleu test: 0.4
  saving hparams to tmp/nmt_model_zh\hparams
  global step 5500 lr 1 step-time 10.96s wps 0.64K ppl 170.63 bleu 0.39
  global step 5600 lr 1 step-time 11.52s wps 0.65K ppl 154.09 bleu 0.39

……

  global step 6100