vandit15 / class-balanced-loss-pytorch Goto Github PK

It seems that you forgot to include weights for sigmoid loss

Instead calculating a weight for each batch, applying to class using pos_weight argument in torch.nn.BCELoss(pos_weights=weights)

Simply, https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/class_balanced_loss.py#L71-L82

Are those line of codes same with

effective_num = 1.0 - np.power(beta, samples_per_cls)
weights = (1.0 - beta) / np.array(effective_num)
weights = weights / np.sum(weights) * no_of_classes
loss = torch.nn.BCELoss(reduction='mean', pos_weight=weights)

this code?

why add this sum make all class weight the same

does someone has implement focal_loss with softmax correctly？could you recommend its link for me？thank you very much

shouldn't it be F.cross_entropy instead?

I should compute samples_per_cls of whole dataset or each batch? If there is 0 in samples_per_cls of each batch, the loss will be nan.

Hi, thanks for your code sharing!

I am now trying to understand how do you implement and I have something want to discuss.

In your main function you have provide us with a toy example, I noticed that you have set samples_per_cls equals [2,3,1,2,2] which indicates label 0 has 2 samples , am i right?
But you have random labels,so should I count samples of each class every times ?

I would be appreciated if you can answer my question as quick as possible. Have a good day!

Hey can this be applied to multi label problems?

Thanks

should be alpha: A float tensor of size [num_classes]

Thanks for sharing your code.
I have 4D tensor..

I just understand your code. Thanks!

Thanks for sharing code.I have modified the code. I want to calculate the multivariate classification. The code is as follows. I can always report errors. I wonder if you have tried it?

pred = logits.log_softmax(dim=1)
cb_loss = F.cross_etropy(input=pred, target=labels, weights=weights)

Thanks.

@vandit15 thanks for open-sourcing the code , is it possible to use Cb loss in the object detection or segmentation architecture ?? did you experiment it with any std architecture as yolo retina, deeplab ?? since i am planning on using to our custom object detection architecture so

Hi, i'm interested in your work! Now i have problem, why your implement code of focal loss use "modulator = torch.exp(-gamma * labels * logits - gamma * torch.log(1 + torch.exp(-1.0 * logits)))" ?

thank you for sharing Pytorch implementation,
I was wondering if we need to specify "reduction="none"" for the "binary_cross_entropy_with_logits" function here since we are going to do weighting and reduction on our own later?

Hi, I wonder if this implementation is correct?

When I run the code with softmax, I get weights like this:

tensor([[0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [1.7646, 1.7646, 1.7646, 1.7646, 1.7646],
        [1.7646, 1.7646, 1.7646, 1.7646, 1.7646],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824],
        [0.8824, 0.8824, 0.8824, 0.8824, 0.8824]])

where each sample enjoin the same weights.

This is weird, right? According to the original paper, we should have different weight for each class (column), right?

Hi, Inside the CBloss you are using binary cross-entropy, so why not using cross-entropy ? Could you explain? Thank you

original:
cb_loss = F.binary_cross_entropy_with_logits(input = logits,target = labels_one_hot, weights = weights)

update:
cb_loss = F.binary_cross_entropy_with_logits(input = logits,target = labels_one_hot, weight = weights)

• pytorch optimizer takes keyword 'weight', not 'weights'

Thanks for the resource, it worked great!!

Hi @vandit15,

Thanks for sharing your code.

In line 73, weights = weights / np.sum(weights) * no_of_classes, why is no_of_classes included here to normalise the weights? Any help would be thankful.

Thanks