A caffe version of official PyTorch-ResNeSt.
Caffemodels are avaliable here
| Model Name | Crop Size | PyTorch Top1 | Caffe Top1 | Caffe Speed |
|---|---|---|---|---|
| ResNeSt-50 | 224x224 | 81.03 | 81.11 | 12.9ms |
| ResNeSt-101 | 256x256 | 82.83 | 83.06 | 20.9ms |
| ResNeSt-200 | 320x320 | 83.84 | 84.22 | 58.0ms |
| ResNeSt-269 | 416x416 | 84.54 | 84.67 | 105.2ms |
We convert the official PyTorch-ResNeSt to Caffe by pipeline: PyTorch-ONNX-Caffe.
For exported ONNX model, we first merge Exp-ReduceSum-Div into one Softmax node. Then we convert to caffe by our onnx2caffe tools written from scratch.
Caffe models are tested on single GTX-1080Ti. PyTorch results come from official PyTorch-ResNeSt.
We first test accuracy on ImageNet2012 val with large batch.
Then we test forward time with batch=1 for 10k iterations by evaluation.py tools.
It seems caffe models are slower than that in ResNeSt-paper
Some ops may be more friendly for PyTorch, while less for Caffe.
We test on GTX-1080Ti while the latency in paper tested on Tesla-V100.
Need bvlc-caffe and Permute layer from ssd-caffe.
python evaluation.py -imgs /data/ImageNet2012/val -label /data/ImageNet2012/labels/val.txt -proto resnest50.prototxt -model resnest50.caffemodel -size 224 -batch 20