Function error

[jefferyZhan]

Hi, when i used the EQL loss function, it seems there is an error in the threshold function.
Take Lvis v0.5 as example:
The shape of pred_class_logits is 1231, including a background class. But the shape of the return of get_image_count_frequency is 1230, excluding the background class. Then the indexerror occured. Is it because of version or exactly an error?
It can be fixed by expanding the freq_info, setting the backgound frequence.

[tztztztztz]

Hi, the shape of pred_class_logits in EQL should be 1230 instead of 1231 since EQL uses the sigmoid loss function and does not include an objectiveness branch.

Can you give more information about the error you came across? For example, which codebase and config did you use?

[jefferyZhan]

I used the faster_rcnn_r50_fpn_1x_coco.py(config), LVIS v0.5 dataset, and the mmdetection. Also, I modified the rio bbox head with EQL with sigmoid function.
Without modifying the freq_info, the error report is:
weight[self.freq_info < self.lambda_] = 1
IndexError: The shape of the mask [1230] at index 0 does not match the shape of the indexed tensor [1231] at index 0
after modifying the freq_info, all the loss became "nan".
A little bit wired.

[tztztztztz]

To use EQL in your own repo, please keep these in mind.

1. change the shape of fc-cls layer to (num_classes) instead of (num_classes + 1)
   

   eqlv2/mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py

   Line 88 in 4c7ea97

   self.fc_cls = nn.Linear(self.cls_last_dim, self.num_classes)

2. init parameters with prior bias to avoid "nan"
   

   eqlv2/mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py

   Lines 154 to 156 in 4c7ea97

   	if self.use_sigmoid:
   	bias_cls = bias_init_with_prob(0.001)
   	normal_init(self.fc_cls, std=0.01, bias=bias_cls)

3. adopt your activation ways for testing
   

   eqlv2/mmdet/models/roi_heads/bbox_heads/bbox_head.py

   Lines 212 to 214 in 4c7ea97

   	scores = F.sigmoid(cls_score)
   	dummpy_prob = scores.new_zeros((scores.size(0), 1))
   	scores = torch.cat([scores, dummpy_prob], dim=1)

However, I recommend that use this repo first and get familiar with the equalization losses

[jefferyZhan]

Thanks a lot~