Unit with subsampling in a block

[TropComplique]

Is the position of the subsampling unit in each resnet block correct?

In your network you subsample at the beginning of each block: https://github.com/zengarden/light_head_rcnn/blob/master/experiments/lizeming/light_head_rcnn.ori_res101.coco.ps_roialign/network_desp.py#L93

But in the original tensorflow implementation they subsample at the end of each block:

in the current implementation we subsample the output activations in the last residual unit of each block, instead of subsampling the input activations in the first residual unit of each block.

See: https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_utils.py#L30 https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_v1.py#L271

Is it alright to use their pretrained resnet-101 model in this case?

[HiKapok]

@TropComplique same question, the weights in slim resnet101 has several differences with the original resnet. hope answer from authors

[zengarden]

@TropComplique @HiKapok There are some differences between slim resnet101 and original caffe resnet101. I have tried to modified slim resnet101 to have the similar output compared with caffe resnet101~(same input, same output feature map). And the resnet101 in this repo should be same as caffe version.