Antoine Yang

Shouldn't it be y = self.bn3(self.conv3(y)) at line 46 ?

After finetuning, the code loads after the final epoch the checkpoint with best validation results and evaluates it on the test set.

Hi, I also observed overfitting during VideoQA finetuning. I kept logs for the model pretrained on HowToVQA69M + WebVidVQA3M which achieves 47.47% test accuracy on MSVD-QA after finetuning. When training...

There it is! [stdout.log](https://github.com/antoyang/just-ask/files/8765942/stdout.log)

From your message, it is unclear whether 1) you want to generate VideoQA annotations for your own narrated videos as in our method 2) you want to train a VideoQA...

Hi, I do not have this problem when I change the seed for the search in DARTS. Could you provide the exact config you are running ?

I haven't. You may verify that you use the suggested FFMPEG version, that you set up the path to the videos correctly in the config files and that the videos...

The folder "vidstg_vid_path" (in the config files) should contain a folder "video" in which there are the different folders containing VidSTG videos ("0000" etc).

As I haven't met this error, I would suggest looking at which videos are problematic and what's the difference between len(images_list) and len(frame_ids) for these videos.

I did all experiments on 16 GPUs (with batch size 1 per GPU) that's why I used batch size 16 on all datasets. For HC-STVG2.0, in initial experiments I tried...