Questions about decathlon challenge about mtan

[lorenmt]

mtan/visual_decathlon/model_wrn_eval.py

Line 240 in 4a3406c

if opt.dataset == 'notimagenet':

This line gives the details of 'notimagenet' flag.

[colorjam]

Sorry, I thought it was not training with Imagenet so I can't find the flags in model_srn_mtan.py. In model_wrn_mtan.py, the data_path only include decathlon-data, so during training you not use the Imagenet dataset? And you train the WideResNet from scratch instead of a pre-trained model?

[lorenmt]

So for decathlon challenge, we don't allow to use external data. So the imagenet here is the reduced resolution version included in the decathlon challenge.

mtan/visual_decathlon/model_wrn_mtan.py

Line 298 in 4a3406c

torch.save(WideResNet_MTAN.state_dict(), 'model_weights/imagenet')

So in `model_wrn_mtan.py' we trained all datasets together but save imagenet features first then fine-tune to the other 9 datasets.