Sergei Belousov

Results 76 comments of Sergei Belousov

Extremely low mAP for more classes

@alexandoikon13 what is values of per-class AP by mAP = 0.0123? Maybe detector works too bad for some classes ...

Extremely low mAP for more classes

So, library works correct for more than one class at my side. If you have reproducer with some errors related to multiclass mAP, please provide it.

Loss goes to 0

Hi, It's works for me with latest version of PyTorch. Which version of PyTorch do you use? Best regards, BeS

Loss goes to 0

This code works fine for me: ``` class Loss: def __init__(self): self.criterion = nn.NLLLoss(reduce=False) self.mpl = mpl.MaxPoolingLoss(ratio=0.3, p=2.0, reduce=True) def __call__(self, pred, gt): loss = self.criterion(pred, gt) loss = self.mpl(loss)...

Loss goes to 0

You can try to visualize loss before and after mpl and check that it works correct. ``` class Loss: def __init__(self): self.criterion = nn.NLLLoss(reduce=False) self.mpl = mpl.MaxPoolingLoss(ratio=0.3, p=2.0, reduce=False) def...

loss : nan when train custom data set

Hi, Can you provide small reproducer for this bug?

how to embed this loss function in keras or tensorflow?

I think that you should write custom API of compute_weights function for keras/tf.

Sample code for num_classes>1 for ver 0.0.2.1

@ne1114 sorry for the late response. You can use sample code straight forward for more than one classes using class_id fields in preds/gt

Could you please upload the C++ implementation ?

Hi, what do you mean 'C++ implementation'? There are C implementation of compute_weights function, it can be found here: https://github.com/BelBES/mpl.pytorch/blob/master/mpl/src/lib_mpl.cpp

Could you please upload the C++ implementation ?

I'm not sure what do you mean: would you like clean c++ API without pytorch? For example using numpy API for compute_weight function? It's good idea, if you'll implement it...