search

Xiaobin-Rong / deepvqe

An unofficial implementation of DeepVQE proposed by Microsoft Corp.
62 stars 19 forks source link

Parameters about DeepVQE-S #6

Open jeffery-work opened 6 months ago

jeffery-work commented 6 months ago

Thanks for your Repo. I just want to know if you have reproduced the DeepVQE-S model ？In the original paper, it says only have 0.59M parameters. “. DeepVQE-S is a downscaled version of our best model. The DeepVQE-S microphone branch has 4 blocks with 16, 40, 56, and 24 filters, the far end branch has 8 and 24 filters, and the decoding branch has 4 blocks with 40, 32, 32, and 27 filters. Additionally, the residual block is omitted in all the encoder blocks and in the first and last decoder blocks to save more computing” According to this part, I implemented the DeepVQE-S model, but it still have "933.32k parameters, 328 MMacs" from get_model_complexity_info(). It is far different from the number of parameters in the article. If you try it, can you synchronize the results? I think the difference mainly comes from Bottleneck(). I set the Bottleneck() input_size=24x11, hidden_size=32x11. FrameSize is 20ms, hop length is 10ms.

Xiaobin-Rong commented 6 months ago

Well, I haven't reproduced the DeepVQE-S model. However, based on my replication of the original model, there is indeed a discrepancy between my results and the data provided in the paper in terms of the model complexity. Additionally, according to my vague memory, there seem to be some missing values for key hyperparameters, which poses challenges for exact reproduction.