Pre-trained ResNet Models in TensorFlow
ResNet was the winner of ILSVRC 2015. It's currently (2/2016) the most accurate image classification model.
The authors of ResNet have published pre-trained models for Caffe. This is a script to convert those exact models for use in TensorFlow.
Running the conversion script, of course, depends on both Caffe and TensorFlow both being installed. You will also need to downlaod the Caffe pretrained models from here. Running the converted model only depends on TensorFlow. Download it with BitTorrent:
resnet-50-20160305.tfmodel.torrent (138M)
resnet-101-20160305.tfmodel.torrent (210M)
resnet-152-20160305.tfmodel.torrent (270M)
The forward.py script shows how to use it.
Notes
-
The convert.py script checks that activations are similiar to the caffe version but it's not exactly the same. This is because TensorFlow handles padding slightly differentl.
-
Resnet is full convolutional. You can resize the network input down to 65x65 images.
Error Rates
Unfortunately, it runs less accurately than the published error rates for
Caffe - I suspect it's due to the padding algorithms being slightly different.
Maybe fine-tuning will help. Use test_error_rate.py to check:
| model | top-1 | top-5 |
|---|---|---|
| Caffe VGG-16 | 28.5% | 9.9% |
| TF ResNet-50 | 27.4% | 9.0% |
| TF ResNet-101 | 26.1% | 8.1% |
| Caffe ResNet-50 | 24.7% | 7.8% |
| TF ResNet-152 | 25.2% | 7.2% |
| Caffe ResNet-101 | 23.6% | 7.1% |
| Caffe ResNet-152 | 23.0% | 6.7% |