Imagenette¶

%reload_ext autoreload
%autoreload 2
%matplotlib inline

from fastai.vision import *

path = Path('/mnt/fe2_disk/imagenet/')
names = {o.name for o in (path/'train').ls()}
classes = pd.read_csv(path/'classids.txt', delimiter=' ', header=None, names=['id','name'])
classes = {k:v for k,v in [o.strip().split() for o in (path/'classids.txt').open().readlines()] if k in names}

#doggies
subids = 'n02096294 n02093754 n02111889 n02088364 n02086240 n02089973 n02087394 n02115641 n02099601 n02105641'.split()

#nette
subids = 'n01440764 n02102040 n02979186 n03000684 n03028079 n03394916 n03417042 n03425413 n03445777 n03888257'.split()

ds = '{' + ','.join(subids) + '}'
ds

', '.join([classes[o] for o in subids])

for o in ['160','320','imagenet']:
    for p in ['val','train']:
        src = Path('/mnt/fe2_disk')/o/p
        dst = Path('/mnt/fe2_disk/imagewoof')/o/p
        dst.mkdir(exist_ok=True, parents=True)
        for s in subids: shutil.copytree(src/s, dst/s)

bs,size = 64,224
path = untar_data(URLs.PETS)/'images'

src = ImageList.from_folder(path).split_by_rand_pct(0.2)
data = src.label_from_re(r'([^/]+)_\d+.jpg$').presize(size)
img = data.train.x[0]
img.px.shape

%timeit flip_lr(img)

%timeit t = tensor(np.ascontiguousarray(np.array(img.px)[...,::-1]))

idxs = torch.arange(224-1,-1,-1)

%timeit t=img.px[...,idxs];

%timeit t=img.px[...,torch.arange(224-1,-1,-1)];