griffintaur / caffe2pytorch-tsn Goto Github PK

The model dir saves my transferred model. I test them on UCF101 and HMDB51 dataset (split1), and get comparable results with the paper as follows.

• first get .hdf5file from caffemodel by export_to_hdf5.py.
• then use the .hdf5 file to get torch model by googlenet.lua. (Because the kinetics caffe model modified the layer name, there is small change in googlenet_kinetics.lua)
• finally transfer torch model to pytorch model by convert_torch.py.

• You also need to modify test_videos.py because of the problem of state_dict, please see details in test_videos.py