With the commit prior to the unsupervised additions, I am able to easily get 70+% top-1 accuracy using the Kinetics/SLOW_8x8_R50.yaml finetuning config. But with the new unsupervised code, I can barely get 40% top-1 accuracy on Kinetics-400 using the same config and training command.
There are a number of bugs I had to fix for the code to run, but I think there may be more. Could the authors consider carefully examining the new codebase for bugs that were introduced with recent code?
I am also unable to reproduce the results of the MoCo unsupervised model trained with p=2 for 200 epochs, but there are larger issues if supervised R50 no longer works on Kinetics either.
With the commit prior to the unsupervised additions, I am able to easily get 70+% top-1 accuracy using the
Kinetics/SLOW_8x8_R50.yamlfinetuning config. But with the new unsupervised code, I can barely get 40% top-1 accuracy on Kinetics-400 using the same config and training command.There are a number of bugs I had to fix for the code to run, but I think there may be more. Could the authors consider carefully examining the new codebase for bugs that were introduced with recent code?
run_net.py. Thus LR scaling doesn't work as nowhere else in the code scales the LR by the number of shards.I am also unable to reproduce the results of the MoCo unsupervised model trained with p=2 for 200 epochs, but there are larger issues if supervised R50 no longer works on Kinetics either.