victory

Training process occurs nan at the first ten batch.

2

envs: Mxnet1.8 fp16 resnet50, byteps 0.2.5.post15. Training process occurs the nan, as shown in the following: ( changing lr from 0.1 to 0.001, the nan disappear, But the loss seems...

The problem for not fixing BN

6

What is the Map in pascal voc 07 dataset using this code? As i read your code, i find that you do not fix the BN parameters. The batchsize is...

fix some small problems

Some questions about HungarianMatcher for focal_loss_cost

3

In the [HungarianMatcher](https://github.com/IDEA-Research/detrex/blob/main/detrex/modeling/matcher/matcher.py#:~:text=%5B%3A%2C%20tgt_ids%5D-,elif%20self.cost_class_type%20%3D%3D%20%22focal_loss_cost%22%3A,cost_class%20%3D%20pos_cost_class%5B%3A%2C%20tgt_ids%5D%20%2D%20neg_cost_class%5B%3A%2C%20tgt_ids%5D,-%23%20Compute%20the%20L1) ``` elif self.cost_class_type == "focal_loss_cost": alpha = self.alpha gamma = self.gamma neg_cost_class = (1 - alpha) * (out_prob**gamma) * (-(1 - out_prob + 1e-8).log()) pos_cost_class = alpha...

IdentityBasedConv1x1

1

Can I repalce IdentityBasedConv1x1 with conv1x1.

Where is the combination of FPN and Light_head RCNN.

1

I see this in the paper, but i can not find this in the code.

final_target_logit = torch.where(target_logit > self.threshold, cos_theta_m, target_logit - self.mm)

1

In MV-softmax, the released code use easy_margin like following: ``` final_gt = torch.where(gt > 0, cos_theta_m, gt) ``` but the official arcface implement does not use easy_margin, like following: ```...

We have calculated the loss of the gate, but does this have any effect on training? Where is this Loss used? ``` logits = self.wg(input) #dim: [bxs, num_experts] if self.k...

When will RevColv2 be released ?