Model accuracy with Jester dataset is poor

[joyalbin]

Hi I have tried to validate the pretrained model with Jester dataset.

Preconditions:

1. Retrained model used jester_resnext_101_RGB_32.pth
2. Dataset Jester
3. Configurations opts.zip
4. Source modifications diff.zip
5. PyTorch version 1.1.0
6. Python version 3.7.3

Test:

1. python utils/jester_json.py 'annotation_Jester' to prepare the dataset
2. python offline_test.py to start the execution

But the output precision is very poor

[11/3721] Time 1.07421 (1.13381) prec@1 0.03409 prec@5 0.20455 precision 0.00000 (0.03213) recall 0.00000 (0.01278)
[12/3721] Time 1.09013 (1.13017) prec@1 0.03646 prec@5 0.20312 precision 0.03030 (0.03198) recall 0.03030 (0.01424)
[13/3721] Time 1.07996 (1.12631) prec@1 0.03365 prec@5 0.20192 precision 0.00000 (0.02952) recall 0.00000 (0.01315)
[14/3721] Time 1.08615 (1.12344) prec@1 0.03125 prec@5 0.20089 precision 0.00000 (0.02741) recall 0.00000 (0.01221)

Could you please help me to find what am missing to get the proper output?
Regards,
Albin

[ahmetgunduz]

You need to provide the model path as resume_path. Since you are providing as a pretrain_path the first convolution kernel is randomly generated regardless of the model size or modality of the inputs.