Thanks firstly for sharing this. Could you/anyone please let know why the bottleneck_IR model is different in body from the basic Bottleneck block of resnet50? Link of resnet Bottleneck here https://github.com/pytorch/vision/blob/a4736ea6e8ff25a1265ca5adf9e6e244d78500e8/torchvision/models/resnet.py#L75 . The Bottleneck layer of official Resnet50 shows three conv layer (two 1x1 and one 3x3s) whereas bottleneck_IR has two 3*3 conv layers.Also the identity-layer of original Resnet is now replaced in insightface by conv 1x1 wrapper around identity.)Was this an experimentally different achievement? If yes why is it called a resnet block. Pardon if I am missing something basic.
Thanks firstly for sharing this. Could you/anyone please let know why the bottleneck_IR model is different in body from the basic Bottleneck block of resnet50? Link of resnet Bottleneck here https://github.com/pytorch/vision/blob/a4736ea6e8ff25a1265ca5adf9e6e244d78500e8/torchvision/models/resnet.py#L75 . The Bottleneck layer of official Resnet50 shows three conv layer (two 1x1 and one 3x3s) whereas bottleneck_IR has two 3*3 conv layers.Also the identity-layer of original Resnet is now replaced in insightface by conv 1x1 wrapper around identity.)Was this an experimentally different achievement? If yes why is it called a resnet block. Pardon if I am missing something basic.
Thanks,