fbuchert/fixmatch-pytorch

PyTorch implementation of FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence

fixmatch-pytorch's Issues

Model Input

Hello,

First off, great work. I have been looking through your code to get ideas on how to implement something similar for a project of mine. I have a question about the last layer of the model and training though.

Do I understand correctly, that the inputs (weak, u_weak, u_strong) are concatenated in the batch dimension, meaning the model input is now [B x 3, C, H, W], and the output also [B x 3, ...] ? There are no other "special" modification to the model itself to produce the correct output?

Best Regards,
L

`    logits = model(inputs)
    logits_x = logits[:len(x_weak)]
    logits_u_weak, logits_u_strong = logits[len(x_weak):].chunk(2)
    del inputs

labeled_loss = F.cross_entropy(logits_x, labels, reduction="mean")

  with torch.no_grad():
      pseudo_labels = torch.softmax(logits_u_weak, dim=1)
      max_probs, targets_u = torch.max(pseudo_labels, dim=1)
      mask = max_probs.ge(args.threshold).float()`