ResNeXt-WSL-Caffe

Introduction

The ResNext models trained with billion scale weakly-supervised data. Please refer to Exploring the Limits of Weakly Supervised Pretraining presented at ECCV 2018 for the details of model training. The Caffe models are converted from PyTorch.

Results on ImageNet

Note: Testing Input Size: 224 x 224

Evaluate Models with a Single Image

Evaluate ResNeXt-101 32x8d:

python2 eval_img.py --proto resnext-101-32x8d-wsl.prototxt --model resnext-101-32x8d-wsl.caffemodel --image imgs/cat.jpg

Expected Outputs:

0.82 - 'n02124075 Egyptian cat'0.09 - 'n02123045 tabby, tabby cat'0.04 - 'n02127052 lynx, catamount'0.03 - 'n02123159 tiger cat'0.00 - 'n02123394 Persian cat'

Evaluate ResNeXt-101 32x16d:

python2 eval_img.py --proto resnext-101-32x16d-wsl.prototxt --model resnext-101-32x16d-wsl.caffemodel --image imgs/cat.jpg

Expected Outputs:

0.85 - 'n02124075 Egyptian cat'0.06 - 'n02127052 lynx, catamount'0.04 - 'n02123159 tiger cat'0.03 - 'n02123045 tabby, tabby cat'0.00 - 'n02123394 Persian cat'

Evaluate ResNeXt-101 32x32d:

python2 eval_img.py --proto resnext-101-32x32d-wsl.prototxt --model resnext-101-32x32d-wsl.caffemodel --image imgs/cat.jpg

Expected Outputs:

0.91 - 'n02124075 Egyptian cat'0.04 - 'n02123045 tabby, tabby cat'0.02 - 'n02127052 lynx, catamount'0.01 - 'n02123159 tiger cat'0.00 - 'n02123394 Persian cat'

How to Use Converter

python3 converter/pt_to_caffe.py --save-root tmp --net resnext-101-32x48d

Requirements:

• PyTorch >= 1.0.1

• Python >= 3.5

• torchvision >= 0.2.0

Support & Contributing

A problem will be encountered when converting the largest model ResNeXt-101 32x48d. If you have enough space to make it work, welcome to contribute.

Transform Completed
[libprotobuf FATAL google/protobuf/wire_format.cc:752] CHECK failed: (output->ByteCount()) == (expected_endpoint): : Protocol message serialized to a size different from what was originally expected.  Perhaps it was modified by another thread during serialization?terminate called after throwing an instance of 'google::protobuf::FatalException'
  what():  CHECK failed: (output->ByteCount()) == (expected_endpoint): : Protocol message serialized to a size different from what was originally expected.  Perhaps it was modified by another thread during serialization?Aborted (core dumped)

References

• https://pytorch.org/hub/facebookresearch_WSL-Images_resnext/

• https://github.com/shicai/MobileNet-Caffe/blob/master/eval_image.py

• https://github.com/xxradon/PytorchToCaffe