I’ve also been looking for while. My specific issue is with the dog breeds problem I’m getting a log_preds.shape of (5, 10357, 120) as opposed to the expected (10357, 120). What I think is happening is that we’re given 5 different (possible) sets of predictions. What I found some other people are doing is just getting the mean among all them, e.g log_preds,y = learn.TTA(); probs = np.mean(np.exp(log_preds), axis=0)
Like you can see in here Wiki: Lesson 2
1 Like