TSN model - Caffe2pytorch
This project transfers tsn-caffe model to tan-pytorch model.
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.
| Dataset |
RGB |
Flow |
Fusion |
| UCF101 |
86.01% |
87.70% |
93.82% |
| HMDB51 |
54.90% |
63.53% |
71.18% |
This project has three steps
- 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.
Something to be noticed:
- You also need to modify test_videos.py because of the problem of state_dict, please see details in test_videos.py