Pairs Datasets

---

source

Sameness

 Sameness ()

Converts boolean/binary labels into “Not Same”/“Same” labels

---

source

ImagePair

 ImagePair (x=None, *rest)

Adds showing functionality to fastai’s fastuple

l = TfmdLists([False, 0, True, 1], Sameness())
test_eq([l.decode(o) for o in l], ['Not Same', 'Not Same', 'Same', 'Same'])

---

source

Pairs

 Pairs (singles:fastai.data.core.Datasets, factor=10, tuple_type=<class
        '__main__.ImagePair'>)

Fixed dataset of randomly-sampled pairs

	Type	Default	Details
singles	Datasets		Used to construct pairs
factor	int	10	ratio between numer of pairs and number of single items, for each split
tuple_type	type	ImagePair	To add show logic for pairs
Returns	Datasets

from fastai_datasets.mnist import TinyMNIST

singles = TinyMNIST()
factor = .5
pairs = Pairs(singles, factor, ImagePair)

for i, (idx1, idx2) in enumerate(pairs.items):
    pair, target = pairs[i]
    test_eq(pair, ImagePair(singles.tls[0][(idx1, idx2)]))
    test_eq(target, int(singles.i2t[idx1] == singles.i2t[idx2]))

The pairing is done separately for each subset:

for singles_ss, pairs_ss in zip(singles.subsets, pairs.subsets):
    test_close(len(pairs_ss), len(singles_ss)*factor, 2)
    test_eq(set(pairs_ss.by_target.keys()), {'Not Same', 'Same'})
    test_eq(len(pairs.by_target['Same']), len(pairs.by_target['Not Same']))

pairs.dls().show_batch()