Chengxu Zhuang

The training will first train an IR model for 50000 steps. This should be done in the run_training.sh as the first step. I guess that step failed for some reason....

From the error message, it seems that you are training the model on CPU and therefore not supporting some operations. Are you training on CPUs? I would not suggest so,...

sorry for being late here, I don't quite know why this is happening, especially as other people are able to run the validation and training without any problem. And based...

thanks for your finding. I will add this requirement into the document.

This nearest neighbor validation is done by getting 5 clips from one validation video, finding 10 nearest neighbor for each clip separately in the memory bank of the training videos,...

Just added this explanation into the document.

ResNet-18 has 9 "layers": the first convolution-pooling layer, and the remaining eight residual blocks. top1_5 means reading out from 5th layer (CONV3 in the paper), top1_7 means reading out from...

See the download dataset link in this website: https://deepmind.com/research/open-source/kinetics.

Hi, TwoPathway models are done by first training each pathway and then combining the checkpoints, as described in the paper. See https://github.com/neuroailab/VIE/blob/master/misc/combine_ckpts.py for how to combine the checkpoint.

The function `main` in this script combines VIE-Single and VIE-Slow to get TwoPathway-S (the slow_single_model in the script), the function `main_sf` combines VIE-Single and VIE-Slowfast to get TwoPathway-SF (the slowfast_single_model...