Faster-rcnn.pytorch: difference between ResNet implementation with TorchVision

[EMCP]

Noticed that in PyTorch 1.3 and TorchVision 0.4.x , we get a ResNet that is slightly different from the implementation here.. should it be safe to port to the new version ?

torchvision.models.resnet on the left model/faster_rcnn/resnet.py on the right

[loolzaaa]

May be too late, but i implement this net with torchvision backbones of VGG16 and ResNet networks.

Results (VOC 2007) are the same:

• VGG16 70,9%
• ResNet50 73,1%
• ResNet101 74,5%