Acc of resnet18_dctcn_video_boundary in my test is wrong

[Chengzhig]

I was preprocessed LRW data with /preprocessing/crop_mouth_from_video.py code . then put .npz data with /lipreading/dataloader into model . But I get the result in the image.
Am I forget some steps?

[mpc001]

Hello, that's a weird result on the test set. When testing a model without word boundaries, have you encountered a similar issue? By the way, when testing a model that was trained using word boundary information, please remember to include <ANNONATION-DIRECTORY>