search

lbaermann / qaego4d

Code and Dataset for the CVPRW Paper "Where did I leave my keys? — Episodic-Memory-Based Question Answering on Egocentric Videos"
20 stars 0 forks source link

missing file: importance_teacher_weights.pt #2

Closed chongjg closed 1 year ago

chongjg commented 1 year ago

Hi there, It seems that the "../22_04_b5aa4428_0/importance_teacher_weights.pt" file cannot be found. Could you please provide it or tell me how to get it? Thanks in advance!

lbaermann commented 1 year ago

Hi chongjg,

you only need that file in case you actually want to train the rehearsal memory model (model: rehearsal_mem in config/base.yaml). To create it, you first need to train with model: importance_teacher, then run python -m model.importance_teacher --output_file importance_teacher_weights.pt --checkpoint_path path/to/checkpoint/of/previous/training.pt, and finally adapt the config/model/rehearsal_mem.yaml to point to the correct importance teacher weights file. With this process, the RM model can pick meaningful segments to attend to, as the attention scores of the unconstrained importance teacher model are used for supervision.

If you configure any other model, such two-step process is not needed.