Tacotron2.decoder.infer behaves incorrectly

[QuentinJGMace]

Describe the bug

I tried to train a French Tacotron2 model. Given the samples taken from training, it went perfectly well.
However, when inferring after training using sb's usual pipeline, the mel output was white noise.

After investigating, the problem seems to be coming from the tacotron2 decoder when using decoder.infer. We manually loaded the weights and used the forward function instead and it worked like a charm.

Here are the mel spectrograms obtained by both functions (same weights used):

Expected behaviour

The mel outputs from the decoder.infer and decoder.forward functions should be the same.

To Reproduce

https://colab.research.google.com/drive/1KrOSM-cuiOQESVkPjpKO4_BFNNdyakuy?usp=sharing

Environment Details

No response

Relevant Log Output

No response

Additional Context

No response

[shankarsharma8089]

Hi @QuentinJGMace can i work on this issue

[QuentinJGMace]

Hey @shankarsharma8089 , sure !