ASL

ASL copied to clipboard

https://github.com/Alibaba-MIIL/ASL/blob/7068ff007ac5f56a0ae1794b61a7571c3e4561ec/src/loss_functions/losses.py#L107

ignitemylife

ResNet101+ASL in MS-COCO

7

 This result is difficult to reproduce. Not only am I having a hard time solving this problem, but others are also having this problem too. Would you share the...

sorrowyn

Hi, thanks for the work. I see in the code that targets are built by object areas as ``` output = torch.zeros((3, 80), dtype=torch.long) for obj in target: if obj['area']...

PT0X0E

In my task,it alwasy return nan,May be it is not as wide use as BCE?

3

Why It always return nan?here is my log (l1_loss): L1Loss() (new_loss): AsymmetricLossOptimized() (bcewithlog_loss): AsymmetricLossOptimized() (iou_loss): IOUloss() ) ) 2022-04-20 13:03:33 | INFO | yolox.core.trainer:202 - ---> start train epoch1 2022-04-20...

MangoFF

The following error occurred in anti_aliasing.py:40 when I used multiple Gpus for training RuntimeError: Assertion `THCTensor_(checkGPU)(state, 3, input, output, weight)' failed. Some of weight/gradient/input tensors are located on different GPUs....

ckx506772099

Maybe a Bug

1

In train.py, between line 54 and 69, https://github.com/Alibaba-MIIL/ASL/blob/7068ff007ac5f56a0ae1794b61a7571c3e4561ec/train.py#L54-L69 It says that >#normalize, # no need, toTensor does normalization I have different thought. To Tensor scales images to (0, 1) while...

QuarTerll

Some questions about openimages v6 datasets?

1

Thanks for your work! But when i use your special OpenImages datasets i found there are some class from mid_to_classes.pth not in the idx_to_class in your pretrained model -- Open_ImagesV6_TRresNet_L_448.pth...

Xiaoyiyong555

您好，我看您在论文里提到了使用自适应的方法来选取超参数，请问有没有具体的方法和代码可以学习？

SweetIncubus

the problem of shifted probability

1

position: https://github.com/Alibaba-MIIL/ASL/blob/main/src/loss_functions/losses.py line: 30, 86 code: xs_neg = (xs_neg + self.clip).clamp(max=1), self.xs_neg.add_(self.clip).clamp_(max=1) problem： In the paper, shifted probability should be " xs_neg = (xs_neg - self.clip).clamp(min=0, max=1) "

xiaochang129

Nice work! I wonder how to pass weights to this custom ASL loss, because our previous task used 'criterion = nn.BCEWithLogitsLoss(weight=self.y_pos_weight)'. Can you give some advice on handling the weights...

RayShark