metadata
license: other
license_name: apple-sample-code-license
license_link: LICENSE
A CLIP (Contrastive Language-Image Pre-training) model trained on DFN-2B. Data Filtering Networks (DFNs) are small networks used to automatically filter large pools of uncurated data. This model was trained on 2B images that were filtered from a pool of 12.8B uncurated image-text pairs (12.8B image-text pairs from CommonPool-12.8B).
This model has been converted to PyTorch from the original JAX checkpoints from Axlearn (https://github.com/apple/axlearn). These weights are directly usable in OpenCLIP (image + text).
Model Details
- Model Type: Contrastive Image-Text, Zero-Shot Image Classification.
- Dataset: DFN-2b
- Papers:
- Data Filtering Networks: https://arxiv.org/abs/2309.17425
- Examples Seen: 39B
Model Metrics
| Eval Dataset | Metric |
|---|---|
| ImageNet 1k | 0.8219 |
| Caltech-101 | 0.9500 |
| CIFAR-10 | 0.9864 |
| CIFAR-100 | 0.8934 |
| CLEVR Counts | 0.3403 |
| CLEVR Distance | 0.2321 |
| Country211 | 0.3198 |
| Describable Textures | 0.6681 |
| EuroSAT | 0.6819 |
| FGVC Aircraft | 0.4829 |
| Food-101 | 0.9498 |
| GTSRB | 0.6329 |
| ImageNet Sketch | 0.7043 |
| ImageNet v2 | 0.7570 |
| ImageNet-A | 0.6745 |
| ImageNet-O | 0.3605 |
| ImageNet-R | 0.9184 |
| KITTI Vehicle Distance | 0.2391 |
| MNIST | 0.8745 |
| ObjectNet | 0.7477 |
| Oxford Flowers-102 | 0.8784 |
| Oxford-IIIT Pet | 0.9611 |
| Pascal VOC 2007 | 0.8472 |
| PatchCamelyon | 0.6418 |
| Rendered SST2 | 0.5815 |
| RESISC45 | 0.7300 |
| Stanford Cars | 0.9465 |
| STL-10 | 0.9889 |
| SUN397 | 0.7594 |
| SVHN | 0.6573 |
| Flickr | 0.8467 |
| MSCOCO | 0.5957 |
| WinoGAViL | 0.5551 |
| iWildCam | 0.1857 |
| Camelyon17 | 0.6540 |
| FMoW | 0.1824 |
| Dollar Street | 0.6822 |
| GeoDE | 0.9253 |
| Average | 0.68039 |
Model Usage
With OpenCLIP
import torch
import torch.nn.functional as F
from urllib.request import urlopen
from PIL import Image
from open_clip import create_model_from_pretrained, get_tokenizer
model, preprocess = create_model_from_pretrained('hf-hub:apple/DFN2B-CLIP-ViT-L-14')
tokenizer = get_tokenizer('ViT-L-14')
image = Image.open(urlopen(
'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))
image = preprocess(image).unsqueeze(0)
labels_list = ["a dog", "a cat", "a donut", "a beignet"]
text = tokenizer(labels_list, context_length=model.context_length)
with torch.no_grad(), torch.cuda.amp.autocast():
image_features = model.encode_image(image)
text_features = model.encode_text(text)
image_features = F.normalize(image_features, dim=-1)
text_features = F.normalize(text_features, dim=-1)
text_probs = torch.sigmoid(image_features @ text_features.T * model.logit_scale.exp() + model.logit_bias)
zipped_list = list(zip(labels_list, [round(p.item(), 3) for p in text_probs[0]]))
print("Label probabilities: ", zipped_list)
Citation
@article{fang2023data,
title={Data Filtering Networks},
author={Fang, Alex and Jose, Albin Madappally and Jain, Amit and Schmidt, Ludwig and Toshev, Alexander and Shankar, Vaishaal},
journal={arXiv preprint arXiv:2309.17425},
year={2023}
}