data_dir = Path('../data/tiny_data/anime_heads')

Items

train_items, valid_items = get_items(data_dir)
test_eq(len(train_items), 240)
test_eq(len(valid_items), 60)
train_items[:5]

	id	cap
293	293	aqua hair black eyes
40	40	aqua hair aqua eyes
84	84	aqua hair aqua eyes
47	47	aqua hair aqua eyes
288	288	aqua hair black eyes

Datasets

class Tokenizer[source]

Tokenizer()

tokenizer = Tokenizer()
ori_cap = 'aqua hair aqua eyes'
tags, tag_len = tokenizer.encode(ori_cap)
test_eq(tags, [3, 18])
test_eq(tag_len, 2)
out_cap = tokenizer.decode(tags)
test_eq(out_cap, ori_cap)

ds = AnimeHeadsDataset(train_items, data_dir)
tag, tag_len, img64 = ds[0]
test_eq(tag.shape, (2,))
test_eq(tag_len.shape, ())
test_eq(img64.shape, (64, 64, 3))

print(tag, tag_len)
plt.imshow(img64)

tensor([ 3, 13]) tensor(2)

<matplotlib.image.AxesImage at 0x7f9287e94748>

class Datasets[source]

Datasets(data_dir, pct=1, valid_pct=0.2)

dsets = Datasets(data_dir)
test_eq(len(dsets.train), 240)
test_eq(len(dsets.valid), 60)

DataLoaders

class DataLoaders[source]

DataLoaders(dsets, bs=64)

dls = DataLoaders(dsets, bs=16)

for tag, tag_len, img in dls.train:
    test_eq(tag.shape, (16, 2))
    test_eq(tag_len.shape, (16,))
    test_eq(img.shape, (16, 64, 64, 3))
    break