kaleab-k / videoat Goto Github PK

Hi @kaleab-k ,
The paper mentions that the results are based on ResNext101 architecture for the UCF101 dataset. For benign training of 3D convs on smaller datasets such as UCF101, the general idea is to first train it on a larger dataset such as Kinetics400 and then fine-tune the model on UCF101 by freezing initial conv layers. How the model was adversarially trained is not clear from the paper. Can you please elaborate if one of the following is true?

1. Start with weights from the Kinetics400 dataset, freeze some layers, and then adversarially train ResNext101 for the presented accuracies in the paper
2. Adversarially train ResNext101 from scratch on UCF101 (i didn't achieve the accuracies presented in the paper with this)
3. Any other procedure

Line 173 in traning.py is y_real = D(adv_inputs).squeeze(),but according to the description in the paper it should bey_real = D(inputs).squeeze().
Is there somethong wrong with code?