zzw922cn / tf2_soft_dtw Goto Github PK

Thank you very much for your share! Could you help me figure out how to implement the soft dtw as a custom loss in the deep learning training framework (in tf)? What I hope is to replace the RMSE/MAE.. metrics by the dtw. Thank you in advance.

Running your code causes two problems.

In the line, soft_dtw_distance = batch_soft_dtw(a, b, gamma=0.01, metric="L1"), the code is not executed because no warp is given.

I put a warp value and ran the code, but I had the following error message.
'('custom_gradient function expected to return', 2, 'gradients but returned', 1, 'instead.')'

I tested the code based tf 2.3 and tf_nightly 2.5, and got the same error.

@zzw922cn Hi, I'm a creator of TensorFlowTTS (https://github.com/TensorSpeech/TensorFlowTTS). I have a plan to reproduce Parallel Tacotron2 and will use this implementation :D. Thank you so much, it's not EZ to implement soft_dtw by TensorFlow since tf.Tensor still an immutable object and we need to use tf.TensorArray as an alternative solution.