Loss decreases when noise is introduced to the inputs. Loss calculation might be wrong.

[Shah-imran]

When I introduce random noise to the input loss.item() decreases. Part of the code for training is included here.

  for epoch in range(epochs):
        # model.train()
        epoch_loss = 0
        epoch_step = 0

        with tqdm(total=n_train, desc=f'Epoch {epoch + 1}/{epochs}', unit='img', ncols=50) as pbar:
            for i, batch in enumerate(train_loader):
                global_step += 1
                epoch_step += 1
                images = batch[0]
                bboxes = batch[1]
                paths = batch[2]

                images = images.to(device=device, dtype=torch.float32)
                images.requires_grad = True
                bboxes = bboxes.to(device=device)
                bboxes_pred = model(images)

                loss, loss_xy, loss_wh, loss_obj, loss_cls, loss_l2 = criterion(
                    bboxes_pred, bboxes)

                print("Without Noise - ", loss.item(), loss_xy.item(), loss_wh.item(),
                      loss_obj.item(), loss_cls.item(), loss_l2.item())

                grad = torch.autograd.grad(loss, [images], create_graph=True,
                                           retain_graph=True, allow_unused=True)[0]

                # loss.backward()

                noise = torch.randn(images.size())
                imgs = images.detach().cpu() + noise

                imgs = imgs.to(device=device, dtype=torch.float32)

                bboxes_pred = model(imgs)

                loss, loss_xy, loss_wh, loss_obj, loss_cls, loss_l2 = criterion(
                    bboxes_pred, bboxes)

                print("With Noise - ", loss.item(), loss_xy.item(), loss_wh.item(),
                      loss_obj.item(), loss_cls.item(), loss_l2.item())

Output -
Without Noise - 20.36496353149414 0.0 0.0 20.36496353149414 0.0 6.040985584259033 With Noise - 1.169505000114441 0.0 0.0 1.169505000114441 0.0 0.05301150307059288 --------- looping again Without Noise - 25.24418830871582 0.0 0.0 25.24418830871582 0.0 7.514528274536133 With Noise - 1.8068726062774658 0.0 0.0 1.8068726062774658 0.0 0.12245824933052063 --------- looping again Without Noise - 1.1534618139266968 0.0 0.0 1.1534618139266968 0.0 0.2820863425731659 With Noise - 0.8495630025863647 0.0 0.0 0.8495630025863647 0.0 0.034061141312122345 --------- looping again Without Noise - 2754.517333984375 3.337832450866699 2730.53369140625 18.649364471435547 1.9963617324829102 5465.08642578125 With Noise - 2763.16845703125 2.5691161155700684 2741.94189453125 13.955108642578125 4.702309608459473 5486.26513671875