WLASL experiments, training log

[chevalierNoir]

Hi,

Regarding the experiments of transfer learning on WLASL, using a baseline Kinetics pre-trained model seems to converge very slow. After 10 epochs, the training loss is still > 7, which should be around 3.6 according to the provided training log. I downloaded WLASL data from https://www.robots.ox.ac.uk/~vgg/research/bsl1k/data/info/wlasl.tar, the pre-trained model using misc/pretrained_models/download.sh and also used the same running script:
python main.py
--checkpoint checkpoint/wlasl_i3d_pkinetics
--datasetname wlasl
--num-classes 2000
--num_gpus 2
--num_in_frames 64
--pretrained misc/pretrained_models/kinetics.pth
--ram_data 1
--workers 16

Thanks.

[gulvarol]

Hi,

Have you checked whether the videos in the figs folder look okay, as in there was no problem with the preprocessing of the frames? Maybe try debugging with --ram_data 0 to eliminate the problems with preprocessing.

[chevalierNoir]

Thanks. It was an issue with the data. It works after redownloading.