Open AltoriaDD opened 11 months ago
Hi, we carry out experiments on FERPlus with 7 basic expression classes, same as RAF-DB. As for AffectNet, we conduct pre-processing on the data and use a balanced sampler during the training. We refer you to the following repository for the data pre-processing. https://github.com/HSE-asavchenko/face-emotion-recognition
Are Affectnet and ferplus both using the same learning rate and lr_scheduler as rafdb? And, I tried the performance of "torchsampler.ImbalancedDatasetSampler" on rafdb, but I found that the balanced sampler doesn't seem to work, and accuracy even decreased somewhat.
train_loader = torch.utils.data.DataLoader(train_dataset,
sampler=ImbalancedDatasetSampler(train_dataset),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
Which balanced sampler are you using?
thanks for your great work! I can't reproduce the accuracy on both the AffectNet and FERplus datasets as mentioned in the paper. Can you provide some details of the experiment, such as hyperparameters and random number seed settings? Thank you very much.