RuntimeError: expected backend CUDA and dtype Double but got backend CUDA and dtype Float

[chenzhe-cool]

run train.py
RuntimeError: expected backend CUDA and dtype Double but got backend CUDA and dtype Float

[basilevh]

What line in which file does this occur at?

[chenzhe-cool]

What line in which file does this occur at?
Appears when running train.py file
File "D:/1-image_cropping/Code_with_Paper/Dissecting Image Crops/dissecting-image-crops-main/train.py", line 839, in
main(args)
File "D:/1-image_cropping/Code_with_Paper/Dissecting Image Crops/dissecting-image-crops-main/train.py", line 826, in main
start_epoch, args.num_epochs, checkpoint_dir, stop_after_epoch)
File "D:/1-image_cropping/Code_with_Paper/Dissecting Image Crops/dissecting-image-crops-main/train.py", line 621, in _training_loop
mix_patch_batch)
File "D:/1-image_cropping/Code_with_Paper/Dissecting Image Crops/dissecting-image-crops-main/train.py", line 332, in _train_epoch
binary_loss_value = binary_loss(predicted_scores, true_scores)
File "D:\anaconda3\envs\pytorch_py36\lib\site-packages\torch\nn\modules\module.py", line 493, in call
result = self.forward(*input, **kwargs)
File "D:\anaconda3\envs\pytorch_py36\lib\site-packages\torch\nn\modules\loss.py", line 617, in forward
reduction=self.reduction)
File "D:\anaconda3\envs\pytorch_py36\lib\site-packages\torch\nn\functional.py", line 2163, in binary_cross_entropy_with_logits
return torch.binary_cross_entropy_with_logits(input, target, weight, pos_weight, reduction_enum)
RuntimeError: expected backend CUDA and dtype Double but got backend CUDA and dtype Float

[basilevh]

From your logs, it looks like at some point, a float64 was introduced, but I'm not sure where since I've never had this error. It may be happening due to some unusual image format, or perhaps some modification was made elsewhere where arrays / tensors are instantiated without passing dtype=float32. As a possible (but not perfect) quick fix, could you try adding these lines after true_scores = mbatch['cropped'].to(device)?

        patches = patches.type(torch.float32)
        global_images = global_images.type(torch.float32)
        true_rectangles = true_rectangles.type(torch.float32)
        true_scores = true_scores.type(torch.float32)

and if that still doesn't work, just after if not patch_localization_only:, add the line predicted_rectangles_scores = predicted_rectangles_scores.type(torch.float32). Basically, we're just trying to get everything to stay in (or rather, return to) the float32 domain. However, debugging your code to find out where things become float64 will be a cleaner solution.