Hugging Face's logo

Datasets:
iejMac
/
CLIP-Kinetics700

Tasks:
Feature Extraction
Zero-Shot Classification
Languages: English
Multilinguality: monolingual
Size Categories: 100K<n<1M
Language Creators: found
Annotations Creators: found
License:
Dataset card Viewer Files Community
1
Dataset Preview
Viewer
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
An error occurred while generating the dataset All the data files must have the same columns, but at some point there are 7 new columns (108.005900, __index_level_1__, 0031sXqr2dM, __index_level_0__, __index_level_3__, __index_level_4__, __index_level_2__) and 8 missing columns (80, 0.811, 0.077, 0902, -5KQ66BBWC4, 0.151, 1, 0.283). This happened while the csv dataset builder was generating data using <datasets.download.streaming_download_manager.FilesIterable object at 0x7f7e6d564280> Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Error code:   UnexpectedError

Need help to make the dataset viewer work? Open a discussion for direct support.

-5KQ66BBWC4
string
0902
int64
0.077
float64
0.151
float64
0.283
float64
0.811
float64
80
float64
1
float64
-5KQ66BBWC4
902
0.077
0.151
0.283
0.811
9
1
-5KQ66BBWC4
902
0.226
0.032
0.366
0.497
12
0
-5KQ66BBWC4
902
0.226
0.032
0.366
0.497
17
0
-5KQ66BBWC4
902
0.226
0.032
0.366
0.497
80
0
-5KQ66BBWC4
902
0.332
0.194
0.481
0.891
80
2
-5KQ66BBWC4
902
0.332
0.194
0.481
0.891
9
2
-5KQ66BBWC4
902
0.505
0.105
0.653
0.78
9
3
-5KQ66BBWC4
902
0.626
0.146
0.805
0.818
9
5
-5KQ66BBWC4
902
0.805
0.222
0.997
1
80
4
-5KQ66BBWC4
902
0.805
0.222
0.997
1
9
4
-5KQ66BBWC4
903
0
0.162
0.177
0.804
80
6
-5KQ66BBWC4
903
0
0.162
0.177
0.804
9
6
-5KQ66BBWC4
903
0.141
0.158
0.298
0.825
12
8
-5KQ66BBWC4
903
0.141
0.158
0.298
0.825
80
8
-5KQ66BBWC4
903
0.226
0.026
0.363
0.512
12
0
-5KQ66BBWC4
903
0.226
0.026
0.363
0.512
80
0
-5KQ66BBWC4
903
0.328
0.182
0.484
0.895
80
2
-5KQ66BBWC4
903
0.328
0.182
0.484
0.895
9
2
-5KQ66BBWC4
903
0.507
0.147
0.666
0.789
80
7
-5KQ66BBWC4
903
0.507
0.147
0.666
0.789
9
7
-5KQ66BBWC4
903
0.642
0.158
0.791
0.859
9
10
-5KQ66BBWC4
903
0.785
0.15
0.886
0.703
12
11
-5KQ66BBWC4
903
0.785
0.15
0.886
0.703
80
11
-5KQ66BBWC4
903
0.802
0.267
0.994
0.971
80
4
-5KQ66BBWC4
903
0.802
0.267
0.994
0.971
9
4
-5KQ66BBWC4
903
0.865
0.158
0.991
0.436
80
9
-5KQ66BBWC4
903
0.865
0.158
0.991
0.436
9
9
-5KQ66BBWC4
904
0.217
0.008
0.982
0.966
12
4
-5KQ66BBWC4
904
0.217
0.008
0.982
0.966
80
4
-5KQ66BBWC4
905
0.044
0.056
0.236
0.891
12
13
-5KQ66BBWC4
905
0.192
0.072
0.411
0.97
12
15
-5KQ66BBWC4
905
0.192
0.072
0.411
0.97
80
15
-5KQ66BBWC4
905
0.392
0.033
0.556
0.618
14
12
-5KQ66BBWC4
905
0.392
0.033
0.556
0.618
17
12
-5KQ66BBWC4
905
0.614
0.078
0.826
0.975
12
14
-5KQ66BBWC4
905
0.614
0.078
0.826
0.975
80
14
-5KQ66BBWC4
906
0.03
0.078
0.225
0.876
12
13
-5KQ66BBWC4
906
0.191
0.073
0.399
0.971
12
15
-5KQ66BBWC4
906
0.408
0.008
0.586
0.639
14
12
-5KQ66BBWC4
906
0.408
0.008
0.586
0.639
17
12
-5KQ66BBWC4
906
0.614
0.075
0.821
0.999
12
14
-5KQ66BBWC4
906
0.859
0.104
0.996
0.885
12
16
-5KQ66BBWC4
907
0.062
0.074
0.234
0.926
12
13
-5KQ66BBWC4
907
0.19
0.109
0.396
0.995
12
15
-5KQ66BBWC4
907
0.42
0.115
0.616
0.883
14
12
-5KQ66BBWC4
907
0.42
0.115
0.616
0.883
17
12
-5KQ66BBWC4
907
0.615
0.082
0.823
1
12
14
-5KQ66BBWC4
908
0.046
0.072
0.34
1
12
15
-5KQ66BBWC4
908
0.046
0.072
0.34
1
80
15
-5KQ66BBWC4
908
0.191
0.07
0.32
0.348
12
17
-5KQ66BBWC4
908
0.38
0.086
0.693
0.991
14
12
-5KQ66BBWC4
908
0.38
0.086
0.693
0.991
17
12
-5KQ66BBWC4
908
0.634
0.06
0.919
0.762
12
14
-5KQ66BBWC4
909
0.118
0.01
0.889
0.984
12
12
-5KQ66BBWC4
909
0.118
0.01
0.889
0.984
17
12
-5KQ66BBWC4
909
0.118
0.01
0.889
0.984
79
12
-5KQ66BBWC4
909
0.778
0.035
1
0.966
12
14
-5KQ66BBWC4
910
0.053
0.045
0.824
0.983
12
12
-5KQ66BBWC4
910
0.053
0.045
0.824
0.983
17
12
-5KQ66BBWC4
910
0.053
0.045
0.824
0.983
79
12
-5KQ66BBWC4
911
0.037
0.025
0.826
0.978
12
12
-5KQ66BBWC4
911
0.037
0.025
0.826
0.978
17
12
-5KQ66BBWC4
911
0.037
0.025
0.826
0.978
79
12
-5KQ66BBWC4
912
0
0.114
0.602
1
12
12
-5KQ66BBWC4
912
0
0.114
0.602
1
74
12
-5KQ66BBWC4
912
0
0.114
0.602
1
80
12
-5KQ66BBWC4
912
0.374
0.151
0.874
0.982
12
18
-5KQ66BBWC4
912
0.374
0.151
0.874
0.982
79
18
-5KQ66BBWC4
912
0.374
0.151
0.874
0.982
80
18
-5KQ66BBWC4
913
0
0
0.282
0.638
12
19
-5KQ66BBWC4
913
0
0
0.282
0.638
74
19
-5KQ66BBWC4
913
0
0.147
0.564
0.992
12
12
-5KQ66BBWC4
913
0
0.147
0.564
0.992
74
12
-5KQ66BBWC4
913
0
0.147
0.564
0.992
80
12
-5KQ66BBWC4
913
0.385
0.15
0.883
0.987
12
18
-5KQ66BBWC4
913
0.385
0.15
0.883
0.987
79
18
-5KQ66BBWC4
913
0.385
0.15
0.883
0.987
80
18
-5KQ66BBWC4
914
0
0.137
0.6
1
12
12
-5KQ66BBWC4
914
0
0.137
0.6
1
74
12
-5KQ66BBWC4
914
0
0.137
0.6
1
79
12
-5KQ66BBWC4
914
0
0.137
0.6
1
80
12
-5KQ66BBWC4
914
0.384
0.161
0.882
0.995
12
18
-5KQ66BBWC4
914
0.384
0.161
0.882
0.995
79
18
-5KQ66BBWC4
914
0.384
0.161
0.882
0.995
80
18
-5KQ66BBWC4
915
0.29
0.217
0.68
0.975
12
20
-5KQ66BBWC4
915
0.29
0.217
0.68
0.975
17
20
-5KQ66BBWC4
915
0.29
0.217
0.68
0.975
79
20
-5KQ66BBWC4
915
0.29
0.217
0.68
0.975
80
20
-5KQ66BBWC4
915
0.541
0.211
0.657
0.447
12
23
-5KQ66BBWC4
915
0.541
0.211
0.657
0.447
74
23
-5KQ66BBWC4
915
0.541
0.211
0.657
0.447
80
23
-5KQ66BBWC4
915
0.612
0.246
0.811
0.969
12
21
-5KQ66BBWC4
915
0.612
0.246
0.811
0.969
74
21
-5KQ66BBWC4
915
0.612
0.246
0.811
0.969
80
21
-5KQ66BBWC4
915
0.7
0.077
1
0.991
12
22
-5KQ66BBWC4
915
0.7
0.077
1
0.991
74
22
-5KQ66BBWC4
916
0.09
0.304
0.296
1
12
25
-5KQ66BBWC4
916
0.09
0.304
0.296
1
74
25
-5KQ66BBWC4
916
0.09
0.304
0.296
1
80
25
-5KQ66BBWC4
916
0.297
0.211
0.666
0.978
12
20
End of preview.

Dataset Card for CLIP-Kinetics70

Dataset Description

Dataset Summary

CLIP-Kinetics700 is a compressed version of the Kinetics700 dataset using OpenAI's CLIP model.

The original dataset is ~700 GB making it difficult to use and hold in memory on one machine. By downsampling each video to 1 FPS and encoding the frames using CLIP we we're able to compress the dataset to ~8 GB making it very memory-friendly and easy to use.

Dataset Preprocessing

clip-video-encode is a tool you can use to easily and efficiently compute CLIP embeddings from video frames. We used it to generate the embeddings for this dataset.

Dataset Structure

Data Format

We formatted this as a WebDataset for better data-loading performance when training the models. Each split contains a list of tar files each with 10000 data samples. This format can be read and used easily using the EmbeddingWebDatasetReader from clip-video-encode.

CLIP-Kinetics700
 ├── splits.csv
 ├── ds_00000.tar
 |     ├── vid_00000.npy
 |     ├── vid_00000.txt
 |     ├── vid_00000.json
 |     ├── vid_00001.npy
 |     ├── vid_00001.txt
 |     ├── vid_00001.json
 |     └── ...
 |     ├── vid_10000.npy
 |     ├── vid_10000.txt
 |     ├── vid_10000.json
 ├── ds_00001.tar
 |     ├── vid_10001.npy
 |     ├── vid_10001.txt
 |     ├── vid_10001.json
 │     ...
 ...

Data Fields

  • vid.npy: the numpy array with the per-frame embeddings. Shape -> (n_frames, 512)
  • vid.cap: the "caption" of the video. In this case it is the Kinetics700 label.
  • vid.json: additional metadata - YouTube video ID, start time, end time.

Data Splits

  • Train - 536489 samples | 54 tar's
  • Validation - 33966 samples | 4 tar's
  • Test - 64532 samples | 7 tar's

Dataset Creation

Source Data

Data was sourced from DeepMind's Kinetics700 dataset and downloaded using this convenient repository.

Simple Experiments

Using this repository we evaluate CLIP-Kinetics700 with the following simple methods:

Zero-shot Evaluation

Accuracy
Top-1 0.31
Top-5 0.56
mean(Top1, Top5) 0.44

Linear-probe Evaluation

Accuracy
Top-1 0.41
Top-5 0.65
mean(Top1, Top5) 0.53
Downloads last month
2
Edit dataset card