Error while applying your model

[utkarsh-tyagi]

I have applied your pretrained model on some audio files but it give me error :

RuntimeError: Calculated padded input size per channel: (3 x 752). Kernel size: (9 x 3). Kernel size can't be greater than actual input size

Do you know what is the reason?

[yzyouzhang]

Thank you for your question. Could you please let me know the shape of the input feature you feed into the model? I am not sure if you preprocess the raw audio properly.

[yzyouzhang]

In this repo, we provided the MATLAB code for preprocessing. If you prefer to use Python, please refer to preprocess.py in our new repo for the recent work https://github.com/yzyouzhang/Empirical-Channel-CM. Thanks.

[utkarsh-tyagi]

torch.Size([60, 750])

Please find my notebook file what i am doing.

How can i connect with you? email id? notebook_air_2.zip

[utkarsh-tyagi]

@yzyouzhang Also i am not able to train the model from scratch , it give me error: Batch size = 2 GPU : RTX 2060 6 GB

RuntimeError: CUDA out of memory. Tried to allocate 60.00 MiB (GPU 0; 6.00 GiB total capacity; 4.49 GiB already allocated; 0 bytes free; 4.52 GiB reserved in total by PyTorch)

[yzyouzhang]

torch.Size([60, 750])

Please find my notebook file what i am doing.

How can i connect with you? email id? notebook_air_2.zip

The feature size should be [B, 1, 60, 750]. I checked your code. You have ResNet(1, 16) when you set up the model. It should be ResNet(3, 256) if you want to use my pretrained model.

[yzyouzhang]

@yzyouzhang Also i am not able to train the model from scratch , it give me error: Batch size = 2 GPU : RTX 2060 6 GB

RuntimeError: CUDA out of memory. Tried to allocate 60.00 MiB (GPU 0; 6.00 GiB total capacity; 4.49 GiB already allocated; 0 bytes free; 4.52 GiB reserved in total by PyTorch)

In my case, I use RTX 1080 Ti 11GB. I can load with batch size 64. So I think batch size 2 should be totally OK for your device. Have you made sure there are no other processes occupying the GPU memory?

[utkarsh-tyagi]

@yzyouzhang how can i convert my feature size (60,750) to [B, 1, 60, 750].

what is B and 1 means?

[utkarsh-tyagi]

@yzyouzhang i checked it again , no other process occupying my gpu memory.

could you please share your id or zoom id for further contact to solve this issue?

[utkarsh-tyagi]

@yzyouzhang what threshold score you have set for real and fake samples? score = F.softmax(lfcc_outputs)[:, 0] score

[yzyouzhang]

@yzyouzhang how can i convert my feature size (60,750) to [B, 1, 60, 750].

what is B and 1 means?

B is the batch size, 1 is the number of channels for CNN.

[yzyouzhang]

@yzyouzhang what threshold score you have set for real and fake samples? score = F.softmax(lfcc_outputs)[:, 0] score

We do not need a threshold to calculate EER. If you want to classify samples into two classes, you can choose a value between the r1 and r2 of the OCSoftmax.

[yzyouzhang]

@yzyouzhang i checked it again , no other process occupying my gpu memory.

could you please share your id or zoom id for further contact to solve this issue?

Please contact yzyouzhang@gmail.com for further zoom discussions. Thanks.

[utkarsh-tyagi]

ok thanks :)