yunxiaos

I guess it makes sense to use ImageNet statistics here since AVA and ImageNet doesn't differ much in terms of domain. Did observe some improvement

From your snapshot I don't have a very good idea as to what triggered this. What is your ```untitled.py```?

I randomly set 0.005 for dense block and 0.0005 for conv base which got me around 0.072 on val set. But I suspect it could better

See ```test.py```. Cvs file is not required you can pass in path to your images with some minor code change

I actually missed that long time ago, by adding something like ``` fc1 = list(model.classifier.children())[0] ``` to ```model.py``` should do the trick

0.005 for the dense block and 0.0005 for the conv base gives you a reasonable result. But I suspect it is not optimal, maybe do a grid/random search

Try add ```python torch.backends.cudnn.deterministic = True ``` and ```python torch.cuda.manual_seed_all(seed) ``` to see if you can get consistent results. If not then according to https://pytorch.org/docs/stable/notes/randomness.html > Completely reproducible results are...

I accidentally wiped my google drive. Also a bit busy lately so going to take a while to regenerate these

Hi @amiltonwong, yes it is quite easy to manually generate them, you just create a simple dictionary that maps the categorical variables to integer encoding after you download the dataset,...

Fixed, it should be ```PointNetSeg``` as in the imports