The provided ImageNet pre-trained model has some unexpected keys

[XuYunqiu]

Hi, thanks for sharing your implementation.
The shared ImageNet pre-trained model resnet_101_MSRA_C4.pth might be incorrect.

I find it contains som unexpected weights and biases, i.e., roi_heads.{visual_attention_head.meta_box_head, box_head, weak_box_head}.res5.0.{shortcut, conv1, conv2, conv3}, roi_heads.box_predictor.{cls_score, bbox_pred} and proposal_generator.rpn_head.{conv, objectness_logits, anchor_deltas}.

In my understanding, such a pre-trained model usually only contain the a backbone and a classifier for ImageNet 1K classification.
Would you mind checking that the link provided is correct?