1Tongji Universiy, 2Microsoft Corporation *Equal contribution † Corresponding to: yifanyang@microsoft.com
In this paper, we propose LLM2CLIP, a novel approach that embraces the power of LLMs to unlock CLIP’s potential. By fine-tuning the LLM in the caption space with contrastive learning, we extract its textual capabilities into the output embeddings, significantly improving the output layer’s textual discriminability. We then design an efficient training process where the fine-tuned LLM acts as a powerful teacher for CLIP’s visual encoder. Thanks to the LLM’s presence, we can now incorporate longer and more complex captions without being restricted by vanilla CLIP text encoder’s context window and ability limitations. Our experiments demonstrate that this approach brings substantial improvements in cross-modal tasks. Our method directly boosted the performance of the previously SOTA EVA02 model by 16.5% on both long-text and short-text retrieval tasks, transforming a CLIP model trained solely on English data into a state-of-the-art cross-lingual model. Moreover, when integrated into mul- timodal training with models like Llava 1.5, it consistently outperformed CLIP across nearly all benchmarks, demonstrating comprehensive performance improvements.
LLM2CLIP performance
Model Details
- Model Type: vision foundation model, feature backbone
- Pretrain Dataset: CC3M, CC12M, YFCC15M and Recap-DataComp-1B(30M subset)
Usage
Pytorch version
pip install llm2vec pip install transformers==4.46
Go to GitHub
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
from transformers import AutoModel, AutoConfig, AutoTokenizer
from eva_clip import create_model_and_transforms
from llm2vec import LLM2Vec
from PIL import Image
import torch.nn as nn
import torch
class LLM2VecWrapper(LLM2Vec):
def prepare_for_tokenization(self, text):
text = (
"<|start_header_id|>user<|end_header_id|>\n\n"
+ text.strip()
+ "<|eot_id|>"
)
return text
class LlamaVec_1B_FeatureExtractor(nn.Module):
def __init__(self):
super().__init__()
model_path = 'microsoft/LLM2CLIP-Llama-3.2-1B-Instruct-CC-Finetuned'
config = AutoConfig.from_pretrained(model_path)
model = AutoModel.from_pretrained(model_path, config=config, trust_remote_code=True, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_path)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"
self.l2v = LLM2VecWrapper(model, tokenizer, pooling_mode="mean", max_length=512, skip_instruction=True)
def extract_features(self, text):
with torch.amp.autocast('cuda'):
reps_norm = self.l2v.encode(text, convert_to_tensor=True)
reps_norm = torch.nn.functional.normalize(reps_norm, p=2, dim=1)
return reps_norm
model, _, preprocess_val = create_model_and_transforms('Llama3.2-1B-EVA02-L-14-336', force_custom_clip=True)
ckpt = torch.load('LLM2CLIP-Llama3.2-1B-EVA02-L-14-336.pt')
model.load_state_dict(ckpt)
model = model.cuda().eval()
text_model = LlamaVec_1B_FeatureExtractor()
image_path = "CLIP.png"
captions = ["a diagram", "a dog", "a cat"]
image = preprocess_val(Image.open(image_path)).cuda().unsqueeze(dim=0)
text_features = text_model.extract_features(captions).to('cuda')
with torch.no_grad(), torch.cuda.amp.autocast():
image_features = model.encode_image(image)
text_features = model.encode_text(text_features)
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)
BibTeX & Citation
@misc{huang2024llm2clippowerfullanguagemodel,
title={LLM2CLIP: Powerful Language Model Unlock Richer Visual Representation},
author={Weiquan Huang and Aoqi Wu and Yifan Yang and Xufang Luo and Yuqing Yang and Liang Hu and Qi Dai and Xiyang Dai and Dongdong Chen and Chong Luo and Lili Qiu},
year={2024},
eprint={2411.04997},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2411.04997},
}