tensorflow / tensor2tensor

Library of deep learning models and datasets designed to make deep learning more accessible and accelerate ML research.
Apache License 2.0
15.57k stars 3.51k forks source link

ASR t2t-decoder error! #1138

Open zackkui opened 6 years ago

zackkui commented 6 years ago

Description

I got a unexpected error when I was using t2t-decoder! It worked well before I updated the tensor2tensor code.

Environment information

For bugs: reproduction and error logs

Steps to reproduce:

Here is my decoder script:

!/bin/bash

DATA_DIR=/tensor2tensor/librispeech/data OUT_DIR=/tensor2tensor/librispeech/train mkdir -p $OUT_DIR

CHECK_POINT=head -n 1 $OUT_DIR/checkpoint | tr -cd 0-9 DECODE_FILE_PREFIX="step-$CHECK_POINT"

echo $DECODE_FILE_PREFIX MODEL=transformer HPRAMS=transformer_librispeech_v1 PROBLEM=librispeech_train_full_test_clean

./tensor2tensor/bin/t2t-decoder \ --data_dir=$DATA_DIR \ --problem=$PROBLEM \ --model=$MODEL \ --hparams_set=$HPRAMS \ --output_dir=$OUT_DIR \ --decode_to_file=$DECODE_FILE_PREFIX \ --eval_use_test_set=True \

Error logs:

Traceback (most recent call last): File "./tensor2tensor/bin/t2t-decoder", line 17, in tf.app.run() File "/usr/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 125, in run _sys.exit(main(argv)) File "./tensor2tensor/bin/t2t-decoder", line 12, in main t2t_decoder.main(argv) File "./tensor2tensor/bin/t2t_decoder.py", line 194, in main decode(estimator, hp, decode_hp) File "./tensor2tensor/bin/t2t_decoder.py", line 105, in decode dataset_split="test" if FLAGS.eval_use_test_set else None) File "./tensor2tensor/utils/decoding.py", line 196, in decode_from_dataset checkpoint_path=checkpoint_path) File "./tensor2tensor/utils/decoding.py", line 314, in decode_once if re.match("^({})+$".format(text_encoder.PAD), d_input): File "/usr/lib64/python2.7/re.py", line 137, in match return _compile(pattern, flags).match(string) TypeError: expected string or buffer

WindQAQ commented 6 years ago

I also ran into this problem.

The speech_recognition problem sets inputs' feature_encoders to None, which causes log_decode_results reture None for decoded_inputs, and thus, this line will break due to None value for d_input.

The simplest workaround is modify this line to

if d_input is not None and re.match("^({})+$".format(text_encoder.PAD), d_input):
  continue
zackkui commented 6 years ago

@WindQAQ What is your best WER with transformer? I got WER:18% on librispeech test-clean with 1000h trainning data.

WindQAQ commented 6 years ago

Sorry but I have never run transformer on LibriSpeech. I ran into this problem because I use SpeechRecognitionModality and SpeechRecognitionProblem in my code.

vergilus commented 5 years ago

@zackkui what's your training batch_size? I trained on 1080 with batch_size=2.3M3 and get WER almost your level. it seems that batch=8M8 would be better, but I don't have that kind of devices.