search

HobbitLong / SupContrast

PyTorch implementation of "Supervised Contrastive Learning" (and SimCLR incidentally)
BSD 2-Clause "Simplified" License
3.08k stars 528 forks source link

Supervised contrastive loss saturation from the beginning on custom data #42

Open zaid478 opened 3 years ago

zaid478 commented 3 years ago

Hi, Thank you for the great work.

I am trying to run supervised contrastive loss on my custom dataset of images. It seems that loss is stuck at 6.920 from the beginning and 30 epochs are already done. The batch size is 64 due to GPU memory constraints. Can you tell me why it might be the case? The data has around 17000 samples and 6 classes (somewhat imbalanced) and I am just on the training stage 1 for now but loss has begun and stuck with 6.920. Any help woule be appreciated.

zou-longkun commented 3 years ago

Hi，I encountered the same problem, did you solve this problem? @zaid478

zaid478 commented 3 years ago

No, I am still facing this. Infact, I removed minority classes and tried to train on 3 classes (~3000 samples per each class) but the loss just reduced to 5.6 after 800 epochs. The embeddings learned are also very bad.

Any help is appreciated.

zou-longkun commented 3 years ago

When training, please turn on the argument"--warm" or set a larger learning rate（--learning_rate 0.5）, it seems to solve this problem, at least it works for me

ozanpkr commented 3 years ago

When training, please turn on the argument"--warm" or set a larger learning rate（--learning_rate 0.5）, it seems to solve this problem, at least it works for me

Hi.I use this repo for my project.I need accuracy for each epoch and validation step.Have you ever tried code for get this parameters?