mleimeister / ctc_tensorflow_voxforge

Simple example how to use tensorflow's CTC loss with Voxforge speech data
18 stars 3 forks source link

Increasing validation cost #1

Closed TsainGra closed 5 years ago

TsainGra commented 7 years ago

For the results that you got in your decoded_single_layer_blstm file, were they achieved with the same hyper-parameters? My validation cost starts increasing after a few epochs. Training cost is just fine.

Thanks.

mleimeister commented 7 years ago

Hi,

the decodings that are displayed in decoded_single_layer_blstm.txt are from the first batch of the training data, and the LER reported is also the training error. This was basically meant to explore which kind of errors the algorithm does on the training set and how they improve during training. For the validation set, the error kept oscillating for me around 0.7-0.8. Given the little training data I used this is kind of expected I'd say. For me the project was just a trial how to use the ctc_loss rather than build a functional model. One way to go would be to download much more data from voxforge and do an exhaustive hyperparameter search. Let me know if you have any more questions.

Best regards, Matthias

TsainGra commented 7 years ago

To add more data, if I just append the wav_files as wav_files.append(list_files_for_speaker('<name>', voxforge_data_dir)) will it work? P.S. - I will take care of the download module, just want to know this part.

mleimeister commented 7 years ago

Yes, this should do it. But you could also first download all the speakers you're interested in into one folder and then just grab all the wav files in there.

harishyadav1465 commented 6 years ago

Hi, i am getting below error!

TypeError Traceback (most recent call last)

in () 95 logits = tf.transpose(logits, (1, 0, 2)) 96 ---> 97 loss = ctc.ctc_loss(logits, targets, seq_len) 98 cost = tf.reduce_mean(loss) 99 /root/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/ctc_ops.pyc in ctc_loss(labels, inputs, sequence_length, preprocess_collapse_repeated, ctc_merge_repeated, ignore_longer_outputs_than_inputs, time_major) 137 # _CTCLossGrad() below. 138 if not isinstance(labels, sparse_tensor.SparseTensor): --> 139 raise TypeError("Expected labels (first argument) to be a SparseTensor") 140 141 # For internal calculations, we transpose to [time, batch, num_classes] TypeError: Expected labels (first argument) to be a SparseTensor