About validation and test about wlasl

[CHI-YU-SUNG]

I have the same problem for the first two questions

[dxli94]

(1) As we reported in the paper, we use the validation set to select the hyperparameter. Then we merge the validation set and training set for training and report the result on the testing set with the selected hyperparameters. This is a common practice when the training set is far from being sufficient for training a good model.

(2) At this stage I cannot comment on such a minor difference. Please refer to the original paper for comparisons.

(3) During training, only a fixed number of video frames are sampled for the ease of batching. However, during testing, all of the video frames are fed to the recognition model. This leads to a small gap in the performance difference.

[ENDLICHERE]

Thank you very much for your prompt reply.