Noticing the following . As I understand, python float is 64 bit. Hence , converting to torch.float32 would be lossy. What is the recommended way to preserve as much accuracy as possible in this ...