Why your model has such impressive performance?

[mlii0117]

I downloaded your trained model (C3D on THMOUS14) and tested, I got mAP 37 for threshold 0.5. It is higher than SOTA. But I noticed that the Len(dataloader) is 8355, which should be 13000+. Could you share more training details? Really appreciate for that

[mlii0117]

Well, I understood why it was 6855, but still be confused to the THUMOS14 performance.

[jafar-pivotchain]

Hi Draven, I am trying to run this code for the CPU by customizing it a little. I am however facing some issues in the roi_temporal_pooling module. While running this block my code is unable to find the file roi_temporal_pool in the R-C3D.pytorch/lib/model/roi_temporal_pooling/_ext/roi_temporal_pooling folder. there just has a init.py file which imports this from ._roi_temporal_pooling import lib as _lib, ffi as _ffi. I am stuck on this error
ModuleNotFoundError: No module named 'model.roi_temporal_pooling._ext.roi_temporal_pooling._roi_temporal_pooling'. I see you have successfully run the code. It would be great if you share the same with me.

[mlii0117]

Hi, I run the code on GPU, I never try it on CPU. Its a important step to make the layer, and I thought the make.sh is related to your GPU. So if you use GPU and follow the instructions, the training will be easier.

[qinguoyi]

hi, the reason was mentioned in author's master's thesis.
The author thought it was mainly due to the adoption of better hyperparameters and the difference between framework execution, Pytorch vs Caffe.