EVA-CLIP-18B

<div align="center"> <h2><a href="https://arxiv.org/abs/2402.04252">EVA-CLIP-18B: Scaling CLIP to 18 Billion Parameters</a></h2>

Quan Sun^1*, Jinsheng Wang^1*, Qiying Yu^1,2*, Yufeng Cui¹, Fan Zhang¹, Xiaosong Zhang¹, Xinlong Wang¹

¹ BAAI, ² THU <br>^* equal contribution

Paper | Github

</div>

Scaling up contrastive language-image pretraining (CLIP) is critical for empowering both vision and multimodal models. We present EVA-CLIP-18B, the largest and most powerful open-source CLIP model to date, with 18-billion parameters. With only 6-billion training samples seen, EVA-CLIP-18B achieves an exceptional 80.7% zero-shot top-1 accuracy averaged across 27 widely recognized image classification benchmarks, outperforming its forerunner EVA-CLIP (5-billion parameters) and other open-source CLIP models by a large margin. Remarkably, we observe a consistent performance improvement with the model size scaling of EVA-CLIP, despite maintaining a constant training dataset of 2-billion image-text pairs from LAION-2B and COYO-700M. This dataset is openly available and much smaller than the in-house datasets (e.g., DFN-5B, WebLI-10B) employed in other state-of-the-art CLIP models. EVA-CLIP-18B demonstrates the potential of EVA-style weak-to-strong visual model scaling. With our model weights made publicly available, we hope to facilitate future research in vision and multimodal foundation models.

Table of Contents

• Summary of EVA-CLIP performance
• Model Card
  • EVA-CLIP-8B
  • EVA-CLIP-18B
• Usage
• BibTeX & Citation

Summary of EVA-CLIP performance

Scaling behavior of EVA-CLIP with zero-shot classification performance averaged across 27 image classification benchmarks, compared with the current state-of-the-art and largest CLIP models (224px). The diameter of each circle demonstrates the forward GFLOPs × the number of training samples seen. The performance of EVA-CLIP consistently improves as scaling up.

Model Card

EVA-8B and EVA-18B

<div align="center"> </div>

EVA-CLIP-8B

Image encoder MIM teacher: EVA02_CLIP_E_psz14_plus_s9B.

<div align="center"> </div>

EVA-CLIP-18B

Image encoder MIM teacher: EVA02_CLIP_E_psz14_plus_s9B.

<div align="center"> </div>

• To construct Merged-2B, we merged 1.6 billion samples from LAION-2B dataset with 0.4 billion samples from COYO-700M.
• The Merged-2B+ consists of all samples from Merged-2B, along with 20 millions samples from LAION-COCO and 23 millions samples from Merged-video including VideoCC, InternVid and WebVid-10M. Merged-video was added at the end of the training process.

It's important to note that all results presented in the paper are evaluated using PyTorch weights. There may be differences in performance when using Hugging Face (hf) models.

Zero-Shot Evaluation

We use CLIP-Benchmark to evaluate the zero-shot performance of EVA-CLIP models. Following vissl, we evauate the zero-shot video classification using 1 middle frame. Further details regarding the evaluation datasets can be found in our paper, particularly in Table 11.

Usage

Huggingface Version

from PIL import Image
from transformers import AutoModel, AutoConfig
from transformers import CLIPImageProcessor, pipeline, CLIPTokenizer
import torch
import torchvision.transforms as T
from torchvision.transforms import InterpolationMode

image_path = "CLIP.png"
model_name_or_path = "BAAI/EVA-CLIP-18B" # or /path/to/local/EVA-CLIP-18B
image_size = 224

processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-large-patch14")

# use image processor with conig
# processor = CLIPImageProcessor(size={"shortest_edge":image_size}, do_center_crop=True, crop_size=image_size)

## you can also directly use the image processor by torchvision
## squash
# processor = T.Compose(
#     [
#         T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
#         T.Resize((image_size, image_size), interpolation=InterpolationMode.BICUBIC),
#         T.ToTensor(),
#         T.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
#     ]
# )
## shortest
## processor = T.Compose(
#     [
#         T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
#         T.Resize(image_size, interpolation=InterpolationMode.BICUBIC),
#         T.CenterCrop(image_size),
#         T.ToTensor(),
#         T.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
#     ]
# )

model = AutoModel.from_pretrained(
    model_name_or_path, 
    torch_dtype=torch.float16,
    trust_remote_code=True).to('cuda').eval()

image = Image.open(image_path)
captions = ["a diagram", "a dog", "a cat"]
tokenizer = CLIPTokenizer.from_pretrained(model_name_or_path)
input_ids = tokenizer(captions,  return_tensors="pt", padding=True).input_ids.to('cuda')
input_pixels = processor(images=image, return_tensors="pt", padding=True).pixel_values.to('cuda')

with torch.no_grad(), torch.cuda.amp.autocast():
    image_features = model.encode_image(input_pixels)
    text_features = model.encode_text(input_ids)
    image_features /= image_features.norm(dim=-1, keepdim=True)
    text_features /= text_features.norm(dim=-1, keepdim=True)

label_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)
print(f"Label probs: {label_probs}")

Pytorch version

Go to GitHub

import torch
from eva_clip import create_model_and_transforms, get_tokenizer
from PIL import Image

model_name = "EVA-CLIP-18B" 
pretrained = "eva_clip" # or "/path/to/EVA_CLIP_18B_psz14_s6B.fp16.pt"

image_path = "CLIP.png"
caption = ["a diagram", "a dog", "a cat"]

device = "cuda" if torch.cuda.is_available() else "cpu"
model, _, processor = create_model_and_transforms(model_name, pretrained, force_custom_clip=True)
tokenizer = get_tokenizer(model_name)
model = model.to(device)

image = processor(Image.open(image_path)).unsqueeze(0).to(device)
text = tokenizer(["a diagram", "a dog", "a cat"]).to(device)

with torch.no_grad(), torch.cuda.amp.autocast():
    image_features = model.encode_image(image)
    text_features = model.encode_text(text)
    image_features /= image_features.norm(dim=-1, keepdim=True)
    text_features /= text_features.norm(dim=-1, keepdim=True)

    text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)

print("Label probs:", text_probs)

You can leverage deepspeed.zero.Init() with deepspeed zero stage 3 if you have limited CPU memory. For loading a pretrained checkpoint in the context of using deepspeed.zero.Init(), it's advised to use the load_zero_partitions() function in eva_clip/factory.py.

BibTeX & Citation

@article{EVA-CLIP-18B,
  title={EVA-CLIP-18B: Scaling CLIP to 18 Billion Parameters}, 
  author={Quan Sun and Jinsheng Wang and Qiying Yu and Yufeng Cui and Fan Zhang and Xiaosong Zhang and Xinlong Wang},
  journal={arXiv preprint arXiv:2402.04252},
  year={2023}
}