y should be in torch.int64 dtype without one-hot encoding. And CrossEntropyLoss() will auto encoding it with one-hot (while out is the probability distribution of prediction like one-hot format)