1. Load from disk and upload to the Hub¶

Setup¶

In [ ]:

%pip install detection_datasets

%pip install detection_datasets

In [2]:

import os
import urllib
import zipfile

from detection_datasets import DetectionDataset

import os import urllib import zipfile from detection_datasets import DetectionDataset

Download the files¶

The files (images and annotations) are stored in S3, and the links for downloading them are provided in a GitHub repository.

The dataset is formatted in the COCO format.

Link for the train images:
https://s3.amazonaws.com/ifashionist-dataset/images/train2020.zip
Link for the validation images:
https://s3.amazonaws.com/ifashionist-dataset/images/val_test2020.zip

Link for the train annotations:
https://s3.amazonaws.com/ifashionist-dataset/annotations/instances_attributes_train2020.json
Link for the validation annotations:
https://s3.amazonaws.com/ifashionist-dataset/annotations/instances_attributes_val2020.json

You may notice the the test split is absent: this is because the dataset was part of a Kaggle competition, where the submission are evaluate on a holdout test data that is not public. See notebook 2. to see how to create your custom splits nevertheless.

Let's first define some constants:

In [3]:

# Download from S3
RAW_TRAIN_IMAGES = 'https://s3.amazonaws.com/ifashionist-dataset/images/train2020.zip'
RAW_VAL_IMAGES = 'https://s3.amazonaws.com/ifashionist-dataset/images/val_test2020.zip'
RAW_TRAIN_ANNOTATIONS = 'https://s3.amazonaws.com/ifashionist-dataset/annotations/instances_attributes_train2020.json'
RAW_VAL_ANNOTATIONS = 'https://s3.amazonaws.com/ifashionist-dataset/annotations/instances_attributes_val2020.json'

# to local disk
DATA_DIR = os.path.join(os.getcwd(), 'data')
TRAIN_ANNOTATIONS = 'train.json'
VAL_ANNOTATIONS = 'val.json'

# Download from S3 RAW_TRAIN_IMAGES = 'https://s3.amazonaws.com/ifashionist-dataset/images/train2020.zip' RAW_VAL_IMAGES = 'https://s3.amazonaws.com/ifashionist-dataset/images/val_test2020.zip' RAW_TRAIN_ANNOTATIONS = 'https://s3.amazonaws.com/ifashionist-dataset/annotations/instances_attributes_train2020.json' RAW_VAL_ANNOTATIONS = 'https://s3.amazonaws.com/ifashionist-dataset/annotations/instances_attributes_val2020.json' # to local disk DATA_DIR = os.path.join(os.getcwd(), 'data') TRAIN_ANNOTATIONS = 'train.json' VAL_ANNOTATIONS = 'val.json'

And now download the images and annotations:

In [4]:

def download(url, target):
    """Download image and annotations."""
    
    # Images
    if url.split('.')[-1] == 'zip':
        path, _ = urllib.request.urlretrieve(url=url)
        with zipfile.ZipFile(path, "r") as f:
            f.extractall(target)
            
        os.remove(path)
    
    # Annotations
    else:
        urllib.request.urlretrieve(url=url, filename=target)


os.makedirs(DATA_DIR, exist_ok=True)

download(url=RAW_TRAIN_ANNOTATIONS, target=os.path.join(DATA_DIR, TRAIN_ANNOTATIONS))
download(url=RAW_VAL_ANNOTATIONS, target=os.path.join(DATA_DIR, VAL_ANNOTATIONS))

download(url=RAW_TRAIN_IMAGES, target=DATA_DIR)
download(url=RAW_VAL_IMAGES, target=DATA_DIR)

def download(url, target): """Download image and annotations.""" # Images if url.split('.')[-1] == 'zip': path, _ = urllib.request.urlretrieve(url=url) with zipfile.ZipFile(path, "r") as f: f.extractall(target) os.remove(path) # Annotations else: urllib.request.urlretrieve(url=url, filename=target) os.makedirs(DATA_DIR, exist_ok=True) download(url=RAW_TRAIN_ANNOTATIONS, target=os.path.join(DATA_DIR, TRAIN_ANNOTATIONS)) download(url=RAW_VAL_ANNOTATIONS, target=os.path.join(DATA_DIR, VAL_ANNOTATIONS)) download(url=RAW_TRAIN_IMAGES, target=DATA_DIR) download(url=RAW_VAL_IMAGES, target=DATA_DIR)

Here are the files and directories we have just downloaded:

In [5]:

os.listdir(DATA_DIR)

os.listdir(DATA_DIR)

Out[5]:

['test', 'val.json', 'train.json', 'train']

Note that the the validation images are in the 'test' folder.

Read the downloaded files¶

In [6]:

config = {
    'dataset_format': 'coco',                        # the format of the dataset on disk
    'path': DATA_DIR,                                # where the dataset is located
    'splits': {                                      # how to read the files
        'train': (TRAIN_ANNOTATIONS, 'train'),       # name of the split (annotation file, images directory)
        'val': (VAL_ANNOTATIONS, 'test'),               # the val directory get unziped in 'test'
    },
}

dd = DetectionDataset().from_disk(**config)

config = { 'dataset_format': 'coco', # the format of the dataset on disk 'path': DATA_DIR, # where the dataset is located 'splits': { # how to read the files 'train': (TRAIN_ANNOTATIONS, 'train'), # name of the split (annotation file, images directory) 'val': (VAL_ANNOTATIONS, 'test'), # the val directory get unziped in 'test' }, } dd = DetectionDataset().from_disk(**config)

Analyse the data¶

DataFrame¶

Internally the data is stored in a Pandas DataFrame.

It can viewed grouped by image (the default):

In [7]:

dd.data      # This is the same as calling dd.get_data(index='image')

dd.data # This is the same as calling dd.get_data(index='image')

Out[7]:

	image_path	width	height	split	bbox_id	bbox	category_id	category	area
image_id
23	/content/data/train/3ce385855f07c77fdeb911ed15...	682	1024	train	[150311, 150312, 150313, 150314]	[Bbox id 150311 [445.0, 910.0, 505.0, 983.0], ...	[23, 23, 33, 10]	[shoe, shoe, neckline, dress]	[1422, 843, 373, 56375]
25	/content/data/train/97e45101f7235a9e56fa95c5e4...	683	1024	train	[158953, 158954, 158955, 158956, 158957, 15895...	[Bbox id 158953 [182.0, 220.0, 472.0, 647.0], ...	[2, 33, 31, 31, 13, 7, 22, 22, 23, 23]	[sweater, neckline, sleeve, sleeve, glasses, s...	[87267, 1220, 16895, 18541, 1468, 9360, 8629, ...
26	/content/data/train/47cbe3ead1617a9971dccc438a...	1024	683	train	[169196, 169197, 169198, 169199, 169200, 16920...	[Bbox id 169196 [441.0, 132.0, 499.0, 150.0], ...	[13, 29, 28, 32, 32, 31, 31, 0, 31, 31, 18, 4,...	[glasses, lapel, collar, pocket, pocket, sleev...	[587, 2922, 931, 262, 111, 1171, 540, 3981, 44...
27	/content/data/train/361cc7654672860b1b7c85fe8e...	682	1024	train	[167967, 167968, 167969, 167970, 167971, 16797...	[Bbox id 167967 [300.0, 421.0, 460.0, 846.0], ...	[6, 23, 23, 31, 31, 4, 1, 35, 32, 35, 35, 35, ...	[pants, shoe, shoe, sleeve, sleeve, jacket, to...	[44062, 2140, 2633, 9206, 5905, 44791, 12948, ...
28	/content/data/train/8a20effd8b6ebcaf2b74caa7d3...	853	1024	train	[168041, 168042, 168043, 168044, 168045, 16804...	[Bbox id 168041 [238.0, 309.0, 471.0, 1022.0],...	[10, 32, 35, 31, 4, 29, 33]	[dress, pocket, zipper, sleeve, jacket, lapel,...	[12132, 1548, 755, 43926, 178328, 9316, 136]
...	...	...	...	...	...	...	...	...	...
50525	/content/data/train/1a38cf39ee6e64481e4a22293d...	660	1024	train	[82929, 82930, 82931, 82932, 82933, 82934, 82935]	[Bbox id 82929 [419.0, 214.0, 576.0, 637.0], B...	[31, 31, 33, 2, 15, 8, 32]	[sleeve, sleeve, neckline, sweater, headband, ...	[39360, 26851, 2431, 262021, 1422, 34694, 294]
50526	/content/data/train/ef7574efb1b15c5d28523dc8da...	682	1024	train	[82936, 82937, 82938, 82939, 82940]	[Bbox id 82936 [168.0, 846.0, 504.0, 1022.0], ...	[6, 1, 33, 32, 32]	[pants, top, t-shirt, sweatshirt, neckline, po...	[49375, 151458, 4122, 993, 371]
50528	/content/data/train/0142a4ba3023f646c7b1efbebd...	682	1024	train	[177641, 177642, 177643, 177644, 177645, 177646]	[Bbox id 177641 [195.0, 256.0, 408.0, 725.0], ...	[10, 33, 31, 31, 23, 23]	[dress, neckline, sleeve, sleeve, shoe, shoe]	[47419, 267, 12965, 2090, 2914, 1171]
50530	/content/data/train/32d71f1d77543d2f040f233b7f...	682	1024	train	[82941, 82942, 82943, 82944, 82945]	[Bbox id 82941 [157.0, 236.0, 366.0, 591.0], B...	[11, 33, 31, 23, 23]	[jumpsuit, neckline, sleeve, shoe, shoe]	[46325, 684, 2113, 1714, 2927]
50531	/content/data/train/217a8c4122165839c6967ab743...	682	1024	train	[82946, 82947, 82948, 82949, 82950, 82951, 829...	[Bbox id 82946 [308.0, 900.0, 337.0, 976.0], B...	[23, 23, 10, 19, 34, 31, 31, 39, 33]	[shoe, shoe, dress, belt, buckle, sleeve, slee...	[919, 2813, 86771, 4029, 384, 2542, 1102, 2923...

46781 rows × 9 columns

Or it can be viewed with one row for each annotation:

In [8]:

dd.get_data(index='bbox')

dd.get_data(index='bbox')

Out[8]:

		image_path	width	height	split	bbox	category_id	category	area
image_id	bbox_id
23	150311	/content/data/train/3ce385855f07c77fdeb911ed15...	682	1024	train	Bbox id 150311 [445.0, 910.0, 505.0, 983.0]	23	shoe	1422
	150312	/content/data/train/3ce385855f07c77fdeb911ed15...	682	1024	train	Bbox id 150312 [239.0, 940.0, 284.0, 994.0]	23	shoe	843
	150313	/content/data/train/3ce385855f07c77fdeb911ed15...	682	1024	train	Bbox id 150313 [298.0, 282.0, 386.0, 352.0]	33	neckline	373
	150314	/content/data/train/3ce385855f07c77fdeb911ed15...	682	1024	train	Bbox id 150314 [210.0, 282.0, 448.0, 665.0]	10	dress	56375
25	158953	/content/data/train/97e45101f7235a9e56fa95c5e4...	683	1024	train	Bbox id 158953 [182.0, 220.0, 472.0, 647.0]	2	sweater	87267
...	...	...	...	...	...	...	...	...	...
50531	82950	/content/data/train/217a8c4122165839c6967ab743...	682	1024	train	Bbox id 82950 [285.0, 383.0, 311.0, 429.0]	34	buckle	384
	82951	/content/data/train/217a8c4122165839c6967ab743...	682	1024	train	Bbox id 82951 [429.0, 261.0, 492.0, 354.0]	31	sleeve	2542
	82952	/content/data/train/217a8c4122165839c6967ab743...	682	1024	train	Bbox id 82952 [259.0, 278.0, 294.0, 362.0]	31	sleeve	1102
	82953	/content/data/train/217a8c4122165839c6967ab743...	682	1024	train	Bbox id 82953 [289.0, 274.0, 353.0, 338.0]	39	flower	2923
	82954	/content/data/train/217a8c4122165839c6967ab743...	682	1024	train	Bbox id 82954 [329.0, 237.0, 393.0, 296.0]	33	neckline	288

342182 rows × 8 columns

Image¶

We can show an image an the annotations:

In [9]:

dd.show()

dd.show()

Out[9]:

Numbers¶

In [10]:

dd.n_images

dd.n_images

Out[10]:

46781

In [11]:

dd.n_bbox

dd.n_bbox

Out[11]:

342182

As mentionned earlier, there is no 'test' dataset here:

In [12]:

dd.splits

dd.splits

Out[12]:

['train', 'val']

In [13]:

dd.split_proportions

dd.split_proportions

Out[13]:

	train	val	test
0	0.975246	0.024754	0.0

We also see that >97.5% of the images belong to the training dataset.

Categories¶

There are 46 categories in this dataset, we can get the full list:

In [14]:

dd.n_categories

dd.n_categories

Out[14]:

46

In [15]:

dd.category_names

dd.category_names

Out[15]:

['shirt, blouse',
 'top, t-shirt, sweatshirt',
 'sweater',
 'cardigan',
 'jacket',
 'vest',
 'pants',
 'shorts',
 'skirt',
 'coat',
 'dress',
 'jumpsuit',
 'cape',
 'glasses',
 'hat',
 'headband, head covering, hair accessory',
 'tie',
 'glove',
 'watch',
 'belt',
 'leg warmer',
 'tights, stockings',
 'sock',
 'shoe',
 'bag, wallet',
 'scarf',
 'umbrella',
 'hood',
 'collar',
 'lapel',
 'epaulette',
 'sleeve',
 'pocket',
 'neckline',
 'buckle',
 'zipper',
 'applique',
 'bead',
 'bow',
 'flower',
 'fringe',
 'ribbon',
 'rivet',
 'ruffle',
 'sequin',
 'tassel']

Let's also show the categories with their ids:

In [16]:

dd.categories

dd.categories

Out[16]:

	category
category_id
0	shirt, blouse
1	top, t-shirt, sweatshirt
2	sweater
3	cardigan
4	jacket
5	vest
6	pants
7	shorts
8	skirt
9	coat
10	dress
11	jumpsuit
12	cape
13	glasses
14	hat
15	headband, head covering, hair accessory
16	tie
17	glove
18	watch
19	belt
20	leg warmer
21	tights, stockings
22	sock
23	shoe
24	bag, wallet
25	scarf
26	umbrella
27	hood
28	collar
29	lapel
30	epaulette
31	sleeve
32	pocket
33	neckline
34	buckle
35	zipper
36	applique
37	bead
38	bow
39	flower
40	fringe
41	ribbon
42	rivet
43	ruffle
44	sequin
45	tassel

Upload to the Hub¶

Before uploading to the Hugging Face Hub, we need to authenticate with our access token:

In [ ]:

! huggingface-cli login

! huggingface-cli login

We can now upload the dataset to the Hugging Face Hub:

In [18]:

dd.to_hub(dataset_name='fashionpedia', repo_name='detection-datasets')

dd.to_hub(dataset_name='fashionpedia', repo_name='detection-datasets')

WARNING:datasets.dataset_dict:Pushing split train to the Hub.

  0%|          | 0/7 [00:00<?, ?ba/s]

Pushing dataset shards to the dataset hub:   0%|          | 0/7 [00:00<?, ?it/s]

  0%|          | 0/7 [00:00<?, ?ba/s]

  0%|          | 0/7 [00:00<?, ?ba/s]

  0%|          | 0/7 [00:00<?, ?ba/s]

  0%|          | 0/7 [00:00<?, ?ba/s]

  0%|          | 0/7 [00:00<?, ?ba/s]

  0%|          | 0/7 [00:00<?, ?ba/s]

WARNING:datasets.dataset_dict:Pushing split val to the Hub.

  0%|          | 0/2 [00:00<?, ?ba/s]

Pushing dataset shards to the dataset hub:   0%|          | 0/1 [00:00<?, ?it/s]

Out[18]:

<detection_datasets.detection_dataset.DetectionDataset at 0x7f42731e2b10>