TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance
[ICCV 2023] - TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance
TinyCLIP is a novel cross-modal distillation method for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. This work unleashes the capacity of small CLIP models, fully leveraging large-scale models as well as pre-training data and striking the best trade-off between speed and accuracy.
Use with Transformers
from PIL import Image
import requests
from transformers import CLIPProcessor, CLIPModel
model = CLIPModel.from_pretrained("wkcn/TinyCLIP-ViT-61M-32-Text-29M-LAION400M")
processor = CLIPProcessor.from_pretrained("wkcn/TinyCLIP-ViT-61M-32-Text-29M-LAION400M")
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True)
outputs = model(**inputs)
logits_per_image = outputs.logits_per_image # this is the image-text similarity score
probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities
Highlights
- TinyCLIP ViT-45M/32 uses only half parameters of ViT-B/32 to achieves comparable zero-shot performance.
- TinyCLIP ResNet-19M reduces the parameters by 50% while getting 2x inference speedup, and obtains 56.4% accuracy on ImageNet.
Model Zoo
| Model | Weight inheritance | Pretrain | IN-1K Acc@1(%) | MACs(G) | Throughput(pairs/s) | Link |
|---|---|---|---|---|---|---|
| TinyCLIP ViT-39M/16 Text-19M | manual | YFCC-15M | 63.5 | 9.5 | 1,469 | Model |
| TinyCLIP ViT-8M/16 Text-3M | manual | YFCC-15M | 41.1 | 2.0 | 4,150 | Model |
| TinyCLIP ResNet-30M Text-29M | manual | LAION-400M | 59.1 | 6.9 | 1,811 | Model |
| TinyCLIP ResNet-19M Text-19M | manual | LAION-400M | 56.4 | 4.4 | 3,024 | Model |
| TinyCLIP ViT-61M/32 Text-29M | manual | LAION-400M | 62.4 | 5.3 | 3,191 | Model |
| TinyCLIP ViT-40M/32 Text-19M | manual | LAION-400M | 59.8 | 3.5 | 4,641 | Model |
| TinyCLIP ViT-63M/32 Text-31M | auto | LAION-400M | 63.9 | 5.6 | 2,905 | Model |
| TinyCLIP ViT-45M/32 Text-18M | auto | LAION-400M | 61.4 | 3.7 | 3,682 | Model |
| TinyCLIP ViT-22M/32 Text-10M | auto | LAION-400M | 53.7 | 1.9 | 5,504 | Model |
| TinyCLIP ViT-63M/32 Text-31M | auto | LAION+YFCC-400M | 64.5 | 5.6 | 2,909 | Model |
| TinyCLIP ViT-45M/32 Text-18M | auto | LAION+YFCC-400M | 62.7 | 1.9 | 3,685 | Model |
Note: The configs of models with auto inheritance are generated automatically.
Official PyTorch Implementation
https://github.com/microsoft/Cream/tree/main/TinyCLIP
Citation
If this repo is helpful for you, please consider to cite it. :mega: Thank you! :)
@InProceedings{tinyclip,
title = {TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance},
author = {Wu, Kan and Peng, Houwen and Zhou, Zhenghong and Xiao, Bin and Liu, Mengchen and Yuan, Lu and Xuan, Hong and Valenzuela, Michael and Chen, Xi (Stephen) and Wang, Xinggang and Chao, Hongyang and Hu, Han},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {21970-21980}
}
Acknowledge
Our code is based on CLIP, OpenCLIP, CoFi and PyTorch. Thank contributors for their awesome contribution!