Seonghun Noh

@zhangzhenyu-zzy I bulit docker image for my environment, cuda11.3 with GTX3080. use this image if you are interested https://hub.docker.com/repository/docker/jzi040941/nsh_nhmt

I trained the model about 50epoch on dress1_training set with default setting. Click the link that I've uploaded on my onedrive. take a look [render_final](https://sogang365-my.sharepoint.com/:u:/g/personal/jzi040941_o365_sogang_ac_kr/ES9N6uEamO1BtiueVTBhsAcBrfDz2eiPtR1bA80OzQ_6AA?e=SmBa8B) [segmentation_final](https://sogang365-my.sharepoint.com/:u:/g/personal/jzi040941_o365_sogang_ac_kr/EdOXQuCFpCJGrwpYodNs4FoBm925KfMc68RrKn5O4ybm5A?e=BF6u7q) [structure_final](https://sogang365-my.sharepoint.com/:u:/g/personal/jzi040941_o365_sogang_ac_kr/EZ9FeDOTAsVFqN87usQ-e4AB2PTctg2-sYtk48eD8I57aQ?e=3ZjnVu)

Hi, @dtx525942103 now I'm trying to training with small dev dataset and verifying this enhancement process is work or not before upload pretrained model to here Stay tune Thanks

for my case, nan loss did not appear with DNS-Challenge dataset. I think someone who has Nan loss problem could try this solution.

there is kaldi style `utils/run.sh` file that can pass arguments to rnn_train.py it depends on your parameter in run.sh I would like to recommend to do this way instead of...

Yes, you are right, I've removed those feature normalization both inference and train in other branch, feel free to remove it

Frame Lookahead is needed for internal reason have to care about overflow in pitch filter buffer I draw scratch about pitch buffer in case FRAME LOOKAHEAD is 3(current FRAME_LOOKAHEAD is...

Can you tell me which dataset you use for training?

Hi @YangangCao yes I was dumb, I only check for iter=1, epoch=1 that's why I didn't notice this printed loss increasing error on iteration. I fixed on commit 9de28e0 for...

> Code has error in 'rnn_train.py' so that loss is nan. > > ``` Thanks I've fix in #24