apple/DFN2B-CLIP-ViT-L-14

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: 12.8B

Model Metrics

Eval Dataset	Metric
ImageNet 1k	0.81396
Caltech-101	0.953141
CIFAR-10	0.9836
CIFAR-100	0.8835
CLEVR Counts	0.3338
CLEVR Distance	0.248733
Country211	0.28237
Describable Textures	0.66117
EuroSAT	0.646296
FGVC Aircraft	0.395945
Food-101	0.945861
GTSRB	0.616152
ImageNet Sketch	0.683311
ImageNet v2	0.7453
ImageNet-A	0.6676
ImageNet-O	0.3915
ImageNet-R	0.900033
KITTI Vehicle Distance	0.201125
MNIST	0.8468
ObjectNet	0.739367
Oxford Flowers-102	0.865822
Oxford-IIIT Pet	0.954941
Pascal VOC 2007	0.81644
PatchCamelyon	0.63028
Rendered SST2	0.551345
RESISC45	0.733175
Stanford Cars	0.947146
STL-10	0.976625
SUN397	0.754565
SVHN	0.653503
Flickr	0.8244
MSCOCO	0.570363
WinoGAViL	0.551645
iWildCam	0.18877
Camelyon17	0.626179
FMoW	0.222137
Dollar Street	0.688084
GeoDE	0.91023
Average	0.668558

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}
}