Training is slow after switching to your own data set

[XTian-W]

In the registry script, loading its own dataset using the Imagefolder method (224*224) is slow to train, confirming that it is running on the GPU. Same data set, same hardware device, another script runs at a normal speed。

[Reaidu]

I also encountered this problem, I use 8 workers to pack the data, and during training every 8 batchsize will stop (I suspect in loading etc. threads to pack the data), on the contrary not using pruning doesn't trigger this problem