Datasets:
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
license_name: custom-apple-license
|
4 |
+
license_link: https://github.com/apple/ml-mobileclip/blob/main/LICENSE
|
5 |
+
task_categories:
|
6 |
+
- text-to-image
|
7 |
+
- image-to-text
|
8 |
+
language:
|
9 |
+
- en
|
10 |
+
---
|
11 |
+
|
12 |
+
# Dataset Card for DataCompDR-12M
|
13 |
+
|
14 |
+
<!-- Provide a quick summary of the dataset. -->
|
15 |
+
|
16 |
+
This dataset contains UIDs of DataComp-12M that is a 12M subset of [DataComp-1B-BestPool](https://huggingface.co/datasets/mlfoundations/datacomp_1b).
|
17 |
+
Image-text models trained on DataComp-12M are significantly better than on CC-12M/YFCC-15M as well as DataComp-Small/Medium.
|
18 |
+
For details on this dataset and the improved [DataCompDR-12M](https://huggingface.co/datasets/apple/DataCompDR-12M),
|
19 |
+
please visit our [MobileCLIP paper](https://arxiv.org/abs/2311.17049).
|
20 |
+
|
21 |
+
## Dataset Details
|
22 |
+
|
23 |
+
### Dataset Description
|
24 |
+
|
25 |
+
<!-- Provide a longer summary of what this dataset is. -->
|
26 |
+
|
27 |
+
DataCompDR is an image-text dataset and an enhancement to the DataComp dataset.
|
28 |
+
We reinforce the DataComp dataset using our multi-modal dataset reinforcement strategy.
|
29 |
+
In particular, we create DataCompDR-1B and DataCompDR-12M by reinforcing the DataComp-1B (BestPool filtering) and a uniform subset of 12.8M samples, DataCompDR-12M.
|
30 |
+
We have a one-time generation process, the cost of which is amortized over multiple architectures and extensive ablations.
|
31 |
+
We generate 5 synthetic captions per image using the `coca_ViT-L-14` model in OpenCLIP, and strong random image augmentations (10 for DataCompDR-1B and 30 for DataCompDR-12M).
|
32 |
+
We compute embeddings of an ensemble of two strong teachers (`ViT-L-14` with pretrained weights `datacomp_xl_s13b_b90k` and openai in OpenCLIP) on augmented images as well as real and synthetic captions.
|
33 |
+
Embeddings are 1536-D concatenations of 2x768-D vectors.
|
34 |
+
One seen sample for DataCompDR is a triplet of one randomly augmented image, one ground-truth caption, and one randomly picked synthetic caption.
|
35 |
+
|
36 |
+
- **Curated by:** Original data by [DataComp](https://www.datacomp.ai/) and metadata by Apple.
|
37 |
+
- **License:** We distribute our metadata under our [license](https://github.com/apple/ml-mobileclip/blob/main/LICENSE). The original image url-text samples and metadata were released by [DataComp](https://www.datacomp.ai/) under Creative Common CC-BY-4.0 license. The individual images are under their own copyrights.
|
38 |
+
- **Repository:** [ml-mobileclip GitHub](https://github.com/apple/ml-mobileclip)
|
39 |
+
- **Paper:** [MobileCLIP paper](https://arxiv.org/abs/2311.17049)
|
40 |
+
- **Demo:** Coming Soon
|
41 |
+
|
42 |
+
## Uses
|
43 |
+
|
44 |
+
<!-- Address questions around how the dataset is intended to be used. -->
|
45 |
+
|
46 |
+
Training with DataCompDR shows significant learning efficiency improvement compared to the standard CLIP training.
|
47 |
+
For example, with a single node of 8×A100 GPUs, we achieve 61.7% zero-shot classification on ImageNet-val in approximately one day when training a ViT-B/16 based CLIP from scratch on DataCompDR-12M.
|
48 |
+
Training with DataCompDR-1B sets new state-of-the-art performance on several metrics (Fig. 2) while still using a fraction of the training compute budget compared to previous works.
|
49 |
+
Using DataCompDR, we demonstrate 10x-1000x learning efficiency in comparison to DataComp.
|
50 |
+
|
51 |
+
## Dataset Structure
|
52 |
+
|
53 |
+
<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
|
54 |
+
|
55 |
+
```
|
56 |
+
- uids.txt: List of 12779520 (65536*195) UIDs, one UID per line.
|
57 |
+
- uids.npy: List of 12779520 (65536*195) UIDs as a NumPy array of type `numpy.dtype("u8,u8")`.
|
58 |
+
```
|
59 |
+
|
60 |
+
|
61 |
+
## Citation
|
62 |
+
|
63 |
+
**[MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training](https://arxiv.org/pdf/2311.17049.pdf). (CVPR 2024)**
|
64 |
+
*Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, Oncel Tuzel.*
|
65 |
+
|
66 |
+
```bibtex
|
67 |
+
@InProceedings{mobileclip2024,
|
68 |
+
author = {Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, Oncel Tuzel},
|
69 |
+
title = {MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training},
|
70 |
+
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
|
71 |
+
month = {June},
|
72 |
+
year = {2024},
|
73 |
+
}
|
74 |
+
```
|