some difference between pytorch and caffe version

[ntuyt]

Hi, Can Thanks so much for your contribution for this pytorch version. I have tested on something dataset, I found the performance is around 3-4 worse than the caffe version. I obtain 38.6 accuracy using pytorch code ECO-lite with 16 segements whereas the caffe achieved 42.4 accuracy.

I have some questions about the difference between this pytorch version and original caffe version.

1. The mean is [104 117 123] in caffe whereas the pytorch version uses [104 117 128].
2. The caffe model resizes the image to [320 240] followed by cropping whereas the pytorch version resize the image to [256 256] followed by cropping.

Do you think these differences will result into performance difference? Meanwhile, I wonder the pretrained model on Kinetics you provided is obtained by training your pytorch model or just convert from the pretrained caffe model?

Looking forward to your reply. Best, Tan

[zhang-can]

Hi @ntuyt,

Thanks for your interests in this repo. I also found this problem these days.

The pretrained model is obtained by training pytorch model on Kinetics dataset.

• Have you tried to modifiy these differences and get any improvement?
• Did you use the latest Kinetics pretrained model (eco_lite_rgb_16F_kinetics_v3.pth.tar)?

I try to figure out the problem recently, if you have any ideas, please feel free to tell me.

Thanks!

Best wishes, Can

[ntuyt]

Hi, Can Thanks so much for your reply. I used the eco_lite_rgb_16F_kinetics_v3.pth.tar. I finetuned the eco_lite_rgb_16F_kinetics_v3.pth.tar on something dataset. Using the default settings in your pytorch code, I achieved 38.6 accuracy. But after I change the resize codes, I first resize it into [240 320] instead of [256,256], I get around 40.1 accuracy. It is still worse than reported 42.4 in the paper. Since your pretrained kinetics model is based on 256x256, I guess we could further improve the performance by training the model in kinetics with 240x360 resizing. Do you have time to retrain your model on kinetics using 240x360 rescaling?

[jslslee]

Hi @zhang-can I used the eco_lite_rgb_16F_kinetics_v3.pth.tar. I saw that the best_prec1 was 61.77%. However, when i retrained on the kinetics, the score was a little worse than this. So what is your training strategy? Thanks!

[chen-ming2019]

@ntuyt Hi , I resize the training image into [240,320],but get an error named "RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 7 and 8 in dimension 3 at c:\a\w\1\s\windows\pytorch\aten\src\thc\generic/THCTensorMath.cu:83" .why occur this phenomenon?can you share your changed code about this ?