X3D-Multigrid icon indicating copy to clipboard operation
X3D-Multigrid copied to clipboard

Training from scratch

Open giovi1 opened this issue 1 year ago • 4 comments

Is it possible to train it from scratch? Eventually which is the dataset format I have to provide?

giovi1 avatar Jul 07 '23 09:07 giovi1

Yes, Kinetics experiments are trained from scratch. Input clips are provided as RGB frames. You can refer to the dataset file here: https://github.com/kkahatapitiya/X3D-Multigrid/blob/d63d8fe6210d2b38aa26d71b0062b569687d6be2/kinetics.py#L161

kkahatapitiya avatar Jul 07 '23 15:07 kkahatapitiya

Is it possible to use my own dataset to train the network? which dataset file I have to refer to? Thank you

giovi1 avatar Jul 07 '23 17:07 giovi1

Yes, as long as the the dataset is large-enough, you can train form scratch on your data. Otherwise, I would suggest to finetune the K400 pretrained model on your data. Unfortunately, for your own data, you will have to edit the above dataset file yourself. It's straightforward, you can follow how charades.py is adopted from kinetics.py.

kkahatapitiya avatar Jul 07 '23 19:07 kkahatapitiya

I would try to test the code as it is on the Kinetics dataset, essentially by running train_x3d_kinetics_multigrid.py. I have a maximum of 70 GB of storage at my disposal to save the data. Is this sufficient, or is there another way to test the code? Thank you

giovi1 avatar Aug 19 '23 15:08 giovi1