search

lvapeab / nmt-keras

Neural Machine Translation with Keras
http://nmt-keras.readthedocs.io
MIT License
533 stars 130 forks source link

Error during decoding #86

Closed VP007-py closed 6 years ago

VP007-py commented 6 years ago

After training the model,I tried decoding it with two iterations and i am getting the following error.

python sample_ensemble.py --models trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_1 trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_2 --dataset datasets/Dataset_EuTrans_esen.pkl --text examples/EuTrans/test.en 
[01/11/2018 12:49:07] Reading parameters from config.py
[01/11/2018 12:49:07] <<< Cupy not available. Using numpy. >>>
Using Theano backend.
[01/11/2018 12:49:08] <<< Cupy not available. Using numpy. >>>
[01/11/2018 12:49:08] Using an ensemble of 2 models
[01/11/2018 12:49:08] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_1_Model_Wrapper.pkl ... >>>
[01/11/2018 12:49:08] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_1.h5 ... >>>
[01/11/2018 12:49:12] <<< Loading optimized model... >>>
[01/11/2018 12:49:15] <<< Optimized model loaded. >>>
[01/11/2018 12:49:15] <<< Model loaded in 7.5578 seconds. >>>
[01/11/2018 12:49:15] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_2_Model_Wrapper.pkl ... >>>
[01/11/2018 12:49:15] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_2.h5 ... >>>
[01/11/2018 12:49:17] <<< Loading optimized model... >>>
[01/11/2018 12:49:21] <<< Optimized model loaded. >>>
[01/11/2018 12:49:21] <<< Model loaded in 5.2314 seconds. >>>
[01/11/2018 12:49:21] <<< Loading Dataset instance from datasets/Dataset_EuTrans_esen.pkl ... >>>
[01/11/2018 12:49:21] <<< Dataset instance loaded >>>
[01/11/2018 12:49:21] Removed "val" set outputs of type "text" with id "target_text.
[01/11/2018 12:49:21]   Applying tokenization function: "tokenize_none".
[01/11/2018 12:49:21] Loaded "val" set inputs of data_type "text" with data_id "source_text" and length 2996.
[01/11/2018 12:49:21] Loaded "val" set inputs of data_type "ghost" with data_id "state_below" and length 2996.
[01/11/2018 12:49:21] Loaded "val" set inputs of type "file-name" with id "raw_source_text".
[01/11/2018 12:49:21] 
 <<< Predicting outputs of val set >>>
[01/11/2018 12:49:28] We did not find a dynamic library in the library_dir of the library we use for blas. If you use ATLAS, make sure to compile it with dynamics library.
[01/11/2018 12:49:29] We did not find a dynamic library in the library_dir of the library we use for blas. If you use ATLAS, make sure to compile it with dynamics library.
Traceback (most recent call last):
  File "sample_ensemble.py", line 61, in <module>
    sample_ensemble(args, params)
  File "/home/Vinay737/nmt-keras/nmt_keras/apply_model.py", line 94, in sample_ensemble
    predictions = beam_searcher.predictBeamSearchNet()[s]
  File "/home/Vinay737/nmt-keras/src/keras-wrapper/keras_wrapper/model_ensemble.py", line 249, in predictBeamSearchNet
    return_alphas=self.return_alphas, model_ensemble=True, n_models=len(self.models))
  File "/home/Vinay737/nmt-keras/src/keras-wrapper/keras_wrapper/search.py", line 135, in beam_search
    new_hyp_samples.append(hyp_samples[ti] + [wi])
IndexError: list index out of range

Any fixes ?

lvapeab commented 6 years ago

Hi @PVinay737 ,

in fact, you were doing nothing wrong. Through your problem I detected a recently introduced bug. I've just fixed it (https://github.com/lvapeab/multimodal_keras_wrapper/commit/f6d87f7607f263cd8517107202e1561aec2ac72a). You should update the keras_wrapper to the last version.

Cheers!

VP007-py commented 6 years ago

@lvapeab,that doesn't seem to be the fix

python sample_ensemble.py --models trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_1  trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_2 --dataset datasets/Dataset_EuTrans_esen.pkl --text examples/EuTrans/test.en 
[02/11/2018 17:13:05] Reading parameters from config.py
[02/11/2018 17:13:05] <<< Cupy not available. Using numpy. >>>
Using Theano backend.
[02/11/2018 17:13:06] <<< Cupy not available. Using numpy. >>>
[02/11/2018 17:13:06] Using an ensemble of 2 models
[02/11/2018 17:13:06] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_1_Model_Wrapper.pkl ... >>>
[02/11/2018 17:13:06] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_1.h5 ... >>>
[02/11/2018 17:13:10] <<< Loading optimized model... >>>
[02/11/2018 17:13:13] <<< Optimized model loaded. >>>
[02/11/2018 17:13:13] <<< Model loaded in 7.5471 seconds. >>>
[02/11/2018 17:13:13] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_2_Model_Wrapper.pkl ... >>>
[02/11/2018 17:13:13] <<< Loading model from trained_models/EuTrans_esen_AttentionRNNEncoderDecoder_src_emb_250_bidir_True_enc_LSTM_32_dec_ConditionalLSTM_32_deepout_linear_trg_emb_250_Adam_0.001/epoch_2.h5 ... >>>
[02/11/2018 17:13:15] <<< Loading optimized model... >>>
[02/11/2018 17:13:18] <<< Optimized model loaded. >>>
[02/11/2018 17:13:18] <<< Model loaded in 5.2335 seconds. >>>
[02/11/2018 17:13:18] <<< Loading Dataset instance from datasets/Dataset_EuTrans_esen.pkl ... >>>
[02/11/2018 17:13:19] <<< Dataset instance loaded >>>
[02/11/2018 17:13:19] Removed "val" set outputs of type "text" with id "target_text.
[02/11/2018 17:13:19]   Applying tokenization function: "tokenize_none".
[02/11/2018 17:13:19] Loaded "val" set inputs of data_type "text" with data_id "source_text" and length 2996.
[02/11/2018 17:13:19] Loaded "val" set inputs of data_type "ghost" with data_id "state_below" and length 2996.
[02/11/2018 17:13:19] Loaded "val" set inputs of type "file-name" with id "raw_source_text".
[02/11/2018 17:13:19] 
 <<< Predicting outputs of val set >>>
[02/11/2018 17:13:26] We did not find a dynamic library in the library_dir of the library we use for blas. If you use ATLAS, make sure to compile it with dynamics library.
[02/11/2018 17:13:27] We did not find a dynamic library in the library_dir of the library we use for blas. If you use ATLAS, make sure to compile it with dynamics library.
Total cost of the translations: 25239.669452     Average cost of the translations: 8.424456
The sampling took: 104.785528 secs (Speed: 0.034975 sec/sample)
Traceback (most recent call last):
  File "sample_ensemble.py", line 61, in <module>
    sample_ensemble(args, params)
  File "/home/Vinay737/nmt-keras/nmt_keras/apply_model.py", line 113, in sample_ensemble
    verbose=args.verbose)
  File "/home/Vinay737/nmt-keras/src/keras-wrapper/keras_wrapper/utils.py", line 998, in decode_predictions_beam_search
    x_text))
  File "/home/Vinay737/nmt-keras/src/keras-wrapper/keras_wrapper/utils.py", line 997, in <lambda>
    map(lambda alignment, x_sentence: np.argmax(alignment[:, :max(1, len(x_sentence))], axis=1), alphas,
IndexError: too many indices for array
lvapeab commented 6 years ago

Just pushed (another) fix. Please, try again.

VP007-py commented 6 years ago

Working...!!

How about integrating evaluation metrics after this step with bleu or toolkits like this

lvapeab commented 6 years ago

I'd rather to maintain the sampling process independent of the evaluation. After sampling, you can run your desired evaluator. There is a script (utils/evaluate_from_file.py) for computing several metrics according to the coco evaluation package. You can use this (or any other) for computing BLEU.