Weird time behaviour for face parsing

[adarc8]

I believe there is something about the JIT (Just-In-Time) load causing some unusual behavior. The first batch takes around 2 seconds, the second batch takes around 20 seconds, but the third batch and subsequent batches only take 0.1 seconds.

Is there any information available about this issue?

[icech]

did you solve this problem?

[adarc8]

@icech Yeah it was some torch command that optimize the performance
I think it was
torch.backends.cudnn.benchmark
but i cant find it in the repo so im not sure..
debug the code when loading the model, there is some torch command that should optimize the model or soemthing like that. but it actually cause a large delay at the begining (this command is probably good when there are many epochs of training, and not just a single inference run)