Skip to content

Latest commit

 

History

History
74 lines (67 loc) · 4.98 KB

README.md

File metadata and controls

74 lines (67 loc) · 4.98 KB
Raw

Pretrained models

The semantic alignment model trained on PF-PASCAL dataset

The semantic alignment model of Rocco et al. contains the weights of both feature extractor and the transformtion network. The model can be downloaded as follows:

cd $OS2D_ROOT/models
wget http://www.di.ens.fr/willow/research/weakalign/trained_models/weakalign_resnet101_affine_tps.pth.tar

Models trained on ImageNet for classification in PyTorch

The standard PyTorch models can be downloaded as follows (links from torchvision):

cd $OS2D_ROOT/models
wget https://download.pytorch.org/models/resnet50-19c8e357.pth
wget https://download.pytorch.org/models/resnet101-5d3b4d8f.pth

Models trained on ImageNet for classification in Caffe2

Some projects have reported that specific weights os ResNets originally trained in Caffe2 work better, e.g., in image retrieval. We use these weights ported to PyTorch by Radenović F. et al. the models can be downloaded as follows (links from here):

cd $OS2D_ROOT/models
wget http://cmp.felk.cvut.cz/cnnimageretrieval/data/networks/imagenet/imagenet-caffe-resnet50-features-ac468af.pth
wget http://cmp.felk.cvut.cz/cnnimageretrieval/data/networks/imagenet/imagenet-caffe-resnet101-features-10a101d.pth

These models are in a slightly different format from standard PyTorch, so we need to do some weight surgery to convert them.

conda activate os2d
python convert_resnet_caffe2_cirtorch_to_pytorch.py imagenet-caffe-resnet50-features-ac468af.pth
python convert_resnet_caffe2_cirtorch_to_pytorch.py imagenet-caffe-resnet101-features-10a101d.pth

This should produce files imagenet-caffe-resnet50-features-ac468af-converted.pth and imagenet-caffe-resnet101-features-10a101d-converted.pth.

Model with GroupNorm instead of BatchNorm trained on ImageNet for classification in Caffe2

Group normalization has been reported to work better than BatchNorm when the batch size is small. We have tried using ResNet-50 with GroupNorm. Download the model with 32 groups:

cd $OS2D_ROOT/models
wget https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl

Converting the model:

conda activate os2d
python convert_resnet_caffe2_groupnorm_to_pytorch.py R-50-GN.pkl --num_layers 50

This should produce file resnet50_caffe2_groupnorm.pth

Models trained on COCO for object detection

We have tried to initialize our models from the weights of detection models trained in the maskrcnn-benchmark framework. Download the models and their configs:

cd $OS2D_ROOT/models
wget -P maskrcnn-benchmark https://download.pytorch.org/models/maskrcnn/e2e_mask_rcnn_R_50_FPN_1x.pth
wget -P maskrcnn-benchmark https://raw.githubusercontent.com/facebookresearch/maskrcnn-benchmark/master/configs/e2e_mask_rcnn_R_50_FPN_1x.yaml
wget -P maskrcnn-benchmark https://download.pytorch.org/models/maskrcnn/e2e_mask_rcnn_R_101_FPN_1x.pth
wget -P maskrcnn-benchmark https://raw.githubusercontent.com/facebookresearch/maskrcnn-benchmark/master/configs/e2e_mask_rcnn_R_101_FPN_1x.yaml

To convert the models one needs to have maskrcnn-benchmark installed (we used v0.1) or just have their default config. Scripts for convertion:

conda activate os2d
python convert_resnet_maskrcnnbenchmark_to_pytorch.py maskrcnn-benchmark/e2e_mask_rcnn_R_50_FPN_1x.pth maskrcnn-benchmark/e2e_mask_rcnn_R_50_FPN_1x.yaml
python convert_resnet_maskrcnnbenchmark_to_pytorch.py maskrcnn-benchmark/e2e_mask_rcnn_R_101_FPN_1x.pth maskrcnn-benchmark/e2e_mask_rcnn_R_101_FPN_1x.yaml

This should produce files maskrcnn-benchmark/e2e_mask_rcnn_R_50_FPN_1x_converted.pth and maskrcnn-benchmark/e2e_mask_rcnn_R_101_FPN_1x_converted.pth

Model trained on google-landmarks-2018 for image retrieval

We have tried to initialize from a model trained for large scale image retrieval in the cnnimageretrieval-pytorch project. Download the model:

cd $OS2D_ROOT/models
wget http://cmp.felk.cvut.cz/cnnimageretrieval/data/networks/gl18/gl18-tl-resnet101-gem-w-a4d43db.pth

Scripts for convertion:

conda activate os2d
python convert_resnet_cirtorch_to_pytorch.py gl18-tl-resnet101-gem-w-a4d43db.pth