Datasets:

antofuller
/

mini-VTAB-corruptions

img_idx int64 0 999	label int64 0 101	corruption stringclasses 1 value
0	76	clean
100	101	clean
101	40	clean
102	45	clean
103	27	clean
104	0	clean
105	13	clean
106	57	clean
107	4	clean
108	69	clean
109	7	clean
10	30	clean
110	90	clean
111	63	clean
112	55	clean
113	28	clean
114	7	clean
115	100	clean
116	60	clean
117	42	clean
118	89	clean
119	98	clean
11	94	clean
120	67	clean
121	96	clean
122	32	clean
123	11	clean
124	91	clean
125	28	clean
126	62	clean
127	63	clean
128	78	clean
129	62	clean
12	11	clean
130	39	clean
131	40	clean
132	17	clean
133	39	clean
134	47	clean
135	56	clean
136	61	clean
137	73	clean
138	91	clean
139	48	clean
13	38	clean
140	54	clean
141	76	clean
142	13	clean
143	0	clean
144	98	clean
145	52	clean
146	97	clean
147	90	clean
148	100	clean
149	85	clean
14	57	clean
150	92	clean
151	69	clean
152	44	clean
153	46	clean
154	29	clean
155	71	clean
156	78	clean
157	67	clean
158	74	clean
159	1	clean
15	42	clean
160	79	clean
161	21	clean
162	42	clean
163	88	clean
164	82	clean
165	51	clean
166	12	clean
167	5	clean
168	82	clean
169	26	clean
16	30	clean
170	25	clean
171	90	clean
172	60	clean
173	99	clean
174	89	clean
175	83	clean
176	70	clean
177	17	clean
178	80	clean
179	57	clean
17	101	clean
180	63	clean
181	24	clean
182	12	clean
183	69	clean
184	75	clean
185	24	clean
186	47	clean
187	3	clean
188	8	clean
189	19	clean
18	68	clean

End of preview. Expand in Data Studio

mini-VTAB-C

A collection of VTAB (Visual Task Adaptation Benchmark) datasets. We sampled 1K training samples and 1K testing samples for each task. For each test set, we apply all 15 corruption types from ImageNet-C.

Tasks

datasets = [
    "caltech101",
    "cifar10",
    "cifar100",
    "dtd",
    "flowers",
    "pets",
    "sun397",
    "svhn",

    "pcam",
    "eurosat",
    "resisc45",
    "diabetic_retinopathy",

    "clevr_count_all",
    "clevr_closest_object_distance",
    "dmlab",
    "dsprites_label_x_position",
    "dsprites_label_y_position",
    "dsprites_label_orientation",
    "kitti_distance",
    "smallnorb_label_azimuth",
    "smallnorb_label_elevation",
]

Corruptions

corruptions = [
    "clean",

    "gaussian_noise",
    "shot_noise",
    "impulse_noise",

    "defocus_blur",
    "glass_blur",
    "motion_blur",
    "zoom_blur",

    "snow",
    "frost",
    "fog",
    "brightness",

    "contrast",
    "elastic",
    "pixelate",
    "jpeg",
]

Usage

from datasets import load_dataset

# Load a specific task
ds = load_dataset("antofuller/mini-vtab-corruptions", "cifar10")

# Access splits
train_ds = ds["train"]
test_ds = ds["test"]

# Filter by corruption type
fog_test = test_ds.filter(lambda x: x["corruption"] == "fog")
clean_test = test_ds.filter(lambda x: x["corruption"] == "clean")

# Get all corruption types
corruptions = test_ds.unique("corruption")

Downloads last month: 47

Size of downloaded dataset files:

17 GB

Size of the auto-converted Parquet files:

17 GB

Number of rows:

352,376