patomp/thai-light-multimodal-clip-and-distill

Requirements

pip install pythainlp
pip install gensim>=4.3.1
pip install git+https://github.com/openai/CLIP.git

Usage

Encode a text by

from transformers import AutoModel

text = 'หมากำลังวิ่งในสนามหญ้า'
model = AutoModel.from_pretrained("patomp/thai-light-multimodal-clip-and-distill", trust_remote_code=True)

embeddings = model(text)
print("Text features shape:", embeddings.shape)

Encode an image by

import torch
import clip
import requests
from PIL import Image

device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
image = preprocess(image).unsqueeze(0).to(device)

with torch.no_grad():
    image_features = model.encode_image(image)

print("Image features shape:", image_features.shape) 

Benchmark

On the test set of Thai MS COCO 2014 dataset

Model \ Metrics	text-find-image recall@1	text-find-image recall@10	image-find-text recall@1	image-find-text recall@10	# text samples per second*
Multilingual Encoder
clip-ViT-B-32-multilingual-v1	0.075	0.242	0.096	0.286	251
XLM-Roberta-Large-Vit-B-32	0.226	0.565	0.265	0.596	20
Thai Encoder (WangchanBERTa-based)
Thai-Cross-CLIP	0.167	0.475	0.197	0.523	48
Thai Encoder (Thai2Fit-based)
thai-light-multimodal-clip-and-distill	0.082	0.328	0.118	0.401	450
thai-light-multimodal-distill	0.084	0.319	0.122	0.401	450

Reference

Some part of this content referenced from https://huggingface.co/M-CLIP/XLM-Roberta-Large-Vit-B-32.

For more detail, please visit https://github.com/calzonelover/Lightweight-Multi-modal-Encoder-for-Thai.