MPL-pytorch

MPL-pytorch copied to clipboard

Hi Thank you for your valuable repository. How to change the loss function for binary classification?

omiderfanmanesh

When I used my dataset for training, the model's accuracy dropped from 96% to 51%. can you help me, thank you

smallzhu

dot_product = s_loss_old - s_loss_new but s_loss_new - s_loss_old?

27

Hello, thanking for your pytorch implement of MPL. I think the dot_prodoct should be s_loss_old - s_loss_new but s_loss_new - s_loss_old for the reason here just flip a coin Am...

zwd973

Typo

1

Thanks for the great code. One minor typo: https://github.com/kekmodel/MPL-pytorch/blob/main/data.py#L147

milanlx

CE(pseudo hard label, teacher_us)

1

I have a small question about the above code. The aim of the cross entropy is to make the two different distributions be similar with each other. I understood the...

miramirakim227

In the forward of class WideResNet，there is a BN normalizing the output of the last res-block and if I comment this line, the loss in the model will become NaN...

MushroomCJ

Teacher Gradients

13

Hi, I understand in the reference implementation, the MPL loss on the teacher does nothing essentially. To fix this we use hard labels rather than soft ones. For this purpose,...

monney

Models

1

what is the model you used in this MPL finetune method?

dandiasd

TypeError: object.__new__() takes exactly one argument (the type to instantiate)

1

I have already successfully run the original code. To try to use new datasets, I inherit torch.utils.data.Dataset and create subclass Webface:   I want to use webface as...

FreshmanLin

Errors

1

i encountered this error while running my code with GPU ```WARNING:__main__:Process rank: 0, device: cuda:0, distributed training: True, 16-bits training: True 2021-03-30 16:11:36.205545: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0...

Dandiasdd