search

alibabasglab / D2Former

This repository contains the audio samples for "D2Former: A Fully Complex Dual-Path Dual-Decoder Conformer Network using Joint Complex Masking and Complex Spectral Mapping for Monaural Speech Enhancement" which is submitted to ICASSP 2023.
MIT License
31 stars 4 forks source link

About the model size #1

Open PrShi113 opened 9 months ago

PrShi113 commented 9 months ago

Hi,

Thanks for releasing code of D2Former. It is a very interesting work!

I tried to calculate the number of the parameters of TSCNet in your "generator" file. But I found the model size was 3.28M, rather than 0.87M in your paper.

I wonder if I forgot something...Look forward to hearing from you.

alibabasglab commented 9 months ago

Hi,

Thanks for your interests. Regarding your query, we used the following code in the 'train.py' file for computing the number of parameters:

self.model = TSCNet(num_channel=32, num_features=self.n_fft // 2 + 1).cuda() summary(self.model, [(1, 2, args.cut_len//self.hop+1, int(self.n_fft/2)+1)])

You may follow the same procedure to obtain the number of model parameters.

Thank you.

PrShi113 commented 9 months ago

Thank you for your rapid response!

One more question, I am using a RTX-2080 8G GPU for training, but find CUDA out of memory. Could you inform me the GPUs you used to train your D2Former recipe? Many thanks!