Why dihedral excluded from TTA

Part 1 (2019) Advanced (Part 1 v3)
1

Hi All,
Does some one know why are we excluding the dihedral part from TTA.While we add back flip lr ,crop pad and zoom but not dihedral

augm_tfm = [o for o in learn.data.train_ds.tfms if o.tfm not in
               (crop_pad, flip_lr, dihedral, zoom)]
    try:
        pbar = master_bar(range(8))
        for i in pbar:
            row = 1 if i&1 else 0
            col = 1 if i&2 else 0
            flip = i&4
            d = {'row_pct':row, 'col_pct':col, 'is_random':False}
            tfm = [*augm_tfm, zoom(scale=scale, **d), crop_pad(**d)]
            if flip: tfm.append(flip_lr(p=1.))
            ds.tfms = tfm
2

Probably an empirical result by @jeremy

3

Dihedral applies a random flip (horizontal/vertical/transpose). Given TTA is generating separate predictions for flipped and non-flipped version of inputs it doesn’t make much sense to also have a random flip going on in dihedral.

1 Like