Simplify train.py, evaluate.py, infer.py and tune.py by adding DeepSpeech2Model class for DS2. #183
Conversation
xinghai-sun
commented
Aug 1, 2017
xinghai-sun
commented
- Move functions in model.py into layer.py
- Add a DeepSpeech2Model class in model.py, which has train() and infer_batch() methods.
- Simplify train.py, evaluate.py, infer.py and tune.py。
…infer.py for DS2.
…eech2Model class.
deep_speech_2/tune.py
Outdated
| @@ -62,10 +66,10 @@ | |||
| type=str, | |||
| help="Manifest path for normalizer. (default: %(default)s)") | |||
| parser.add_argument( | |||
| "--decode_manifest_path", | |||
| "--tune_manifest_path", | |||
| default='datasets/manifest.test', | |||
There was a problem hiding this comment.
manifest.dev would be better.
| padding=(5, 10), | ||
| act=paddle.activation.BRelu()) | ||
| output_num_channels = 32 | ||
| output_height = 160 // pow(2, num_stacks) + 1 |
There was a problem hiding this comment.
Can we figure out a way to avoid hardcode here?
There was a problem hiding this comment.
Will be fixed in a later PR (this problem is not coming from current PR).
| @@ -127,100 +129,47 @@ | |||
|
|
|||
There was a problem hiding this comment.
Is it necessary to add output_model_dir into parser as an argument?
deep_speech_2/layer.py
Outdated
|
|
||
| def conv_group(input, num_stacks): | ||
| """ | ||
| Convolution group with several stacking convolution layers. |
There was a problem hiding this comment.
Seems that stacked is often used instead of stacking. The same below.
| for target, result in zip(target_transcripts, result_transcripts): | ||
| wer_sum += wer(target, result) | ||
| num_ins += 1 | ||
| print("WER (%d/?) = %f" % (num_ins, wer_sum / num_ins)) |
There was a problem hiding this comment.
It refers to an unknown size of the validation set. It will become known at the end of the evaluation.
| padding=(5, 10), | ||
| act=paddle.activation.BRelu()) | ||
| output_num_channels = 32 | ||
| output_height = 160 // pow(2, num_stacks) + 1 |
There was a problem hiding this comment.
Since the model is refactored, please make 160 exposed instead of hard-coding.
There was a problem hiding this comment.
Will be fixed in a later PR (this problem is not coming from current PR).
deep_speech_2/layer.py
Outdated
|
|
||
| import paddle.v2 as paddle | ||
|
|
||
| DISABLE_CUDNN_BATCH_NORM = True |
There was a problem hiding this comment.
Why not make DISABLE_CUDNN_BATCH_NORM a parameter of conv_bn_layer? If hard coding here, user has to modify this file when training on cpu mode.
There was a problem hiding this comment.
Removed since the cudnn problem has been fixed.
| reader=dev_batch_reader, feeding=feeding_dict) | ||
| output_model_path = os.path.join( | ||
| output_model_dir, "params.pass-%d.tar.gz" % event.pass_id) | ||
| with gzip.open(output_model_path, 'w') as f: |
There was a problem hiding this comment.
Maybe we should make sure output_model_dir exist, the saving operation may fail if the directory is not exist.
| # of input batch data will be induced during training. | ||
| audio_data = paddle.layer.data( | ||
| name="audio_spectrogram", | ||
| type=paddle.data_type.dense_array(161 * 161)) |
There was a problem hiding this comment.
Just curious about 161 * 161, is this setting proper for other type of feature like mfcc?
There was a problem hiding this comment.
Will fix this in another PR. Lets' just do not bring in any difference to this PR.
deep_speech_2/model.py
Outdated
| def _create_network(self, vocab_size, num_conv_layers, num_rnn_layers, | ||
| rnn_layer_size): | ||
| # paddle.data_type.dense_array is used for variable batch input. | ||
| # The size 161 * 161 is only an placeholder value and the real shape |
There was a problem hiding this comment.
The log information about output dimensions of conv layers is induced from here, which is misleading if this placeholder is not consistent with the real shape. Maybe a better choice is to use the real dimension of feature vector.
There was a problem hiding this comment.
Will fix this in another PR.
