R-50.pkl does not match the original converted

[takaya-akiyama]

I am trying to learn end to end with detectron.

As pre-training model of the backbone, I download the original ResNet-50 model of MSRA and use the model converted by "pickle_caffe_blobs.py".

But "bbox_coco_2014_minival_results.json" generated by the test after learning is empty "[]".

python3 tools/train_net.py --cfg configs/12_2017_baselines/e2e_faster_rcnn_R-50-FPN_2x.yaml OUTPUT_DIR /tmp/detectron-output NUM_GPUS 2

The model converted by the above tool "pickle_caffe_blobs.py" is different from R-50.pkl (ImageNet Pretrained Models) published by detectron.

The model converted by "pickle_caffe_blobs.py" is 12 bytes smaller than R-50.pkl.

Could you tell me why the model converted by "pickle_caffe_blobs.py" is different from R-50.pkl?

Thank you.

original data

ResNet-50-deploy.prototxt (MSRA's original ResNet-50 model):

https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777#authkey=%21AAFW2%2DFVoxeVRck&cid=4006CBB8476FF777&id=4006CBB8476FF777%2117891&parId=4006CBB8476FF777%2117887&o=OneUp

ResNet-50-model.caffemodel (MSRA's original ResNet-50 model):

https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777#authkey=%21AAFW2%2DFVoxeVRck&cid=4006CBB8476FF777&id=4006CBB8476FF777%2117895&parId=4006CBB8476FF777%2117887&o=OneUp

convert tool in this project

pickle_caffe_blobs.py https://github.com/facebookresearch/Detectron/blob/master/tools/pickle_caffe_blobs.py

published model

R-50.pkl https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/MSRA/R-50.pkl