I’m using model.eval(). Thus it is not the problem of dropout
However, I have same input but different output.
Thus I debug the process.
The interesting thing was found in middle.
Sometimes , tensor looks like 4.498e-01 and sometimes look like 4.488 (without ‘e-01’)
As you see, two values are really similar but they are different. Also I don’t know why some tensor has ‘e-01’ or ‘e+00’ etc.
(And I set dtype=torch.float32 to input value. Thus input has always same value and same dtype)
As a result final output shows different result eventhough the input was exactly same.
Can you explain reason?