About reimplementing the model

[sheldontsui]

About reimplementing the model

[SheldonTsui]

@avis-ma Hello!
Under the binary mask mode, you can add an "assert" to ensure the values of gt masks are in {0,1} and the predicted masks are in the range [0,1]. This trick is very useful for your debug.

[avis-ma]

Thank you very much. I will retry this trick. 👍

[813665269]

HI,
When I use the dataset you provide, the training of the model will terminate at the time of torch.utils.data.dataloader.You added the sys.exit (- 1) command to the Musicmixdata function.When I deleted this command,I got this error:Sizes of tensors must match except in dimension 3.
I know this problem caused by the different dimension,but I don't know how to solve this problem.

[SheldonTsui]

@813665269 Please check whether you can load the frames and audio correctly. Debug in the "try" code part, and you will find the problems.

[813665269]

hi,
I print the error data and try to delect all of them,but they are too many.They can run in the Sound-of-Pixels-master,but can not run in your project.I want to know how to Process the datasets to suit this project.
Show the error:

[SheldonTsui]

@813665269 Hello, don't delete the "maybe" error data. Please GO INTO the code block to debug. Maybe there are some little differences in the data format. My data loader is largely based on "Sound-of-Pixels-master", but slightly different.
For example, you can manually load training data here:

Minus-Plus-Network/main.py

Line 419 in 7aa281b

Add visualization code "for data in dataset_train: print(data.keys())", and you can go into such code block and find the exact bug.

[813665269]

Thanks for your answer,
I print the datasets and get the construction of the wrong data and correct data.In your project ,if the function cant read the data,it will return a tensor with all zeros.

[813665269]

Hi,
I reshape the error data's tensor from [1,751,255] to [1,751,256] and then drop the command 'sys.exit(-2)' in MusicmixData .Now,this project can run.But I think the wrong tensor with all zeros must effect the model quality.
Can you give me some advice to treat it?
Thanks a lot.

[SheldonTsui]

@813665269 You should check why the error data happens. In other words, the normal shape of that tensor should be [1, 751, 256]. Why will such wrong tensors present in your case? To ensure all the training data is well-prepared is very important.