CLIP ViT-B/32 xlm roberta base - LAION-5B

CLIP ViT-B/32 xlm roberta base - LAION-5B model converted from OpenCLIP to HuggingFace Transformers.

See https://gist.github.com/calpt/8e3555bd11f1916b5169c8125117e5ee for conversion script and more info.

Usage

Model uses custom code. Make sure to pass trust_remote_code=True when loading the model.

Example:

import torch
from PIL import Image
from transformers import AutoModel, AutoFeatureExtractor, AutoTokenizer

model = AutoModel.from_pretrained("calpt/CLIP-ViT-B-32-xlm-roberta-base-laion5B-s13B-b90k", trust_remote_code=True)

processor = AutoFeatureExtractor.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K")
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

image_input = processor(Image.open("CLIP.png"), return_tensors="pt")
text_input = tokenizer(["a diagram", "a dog", "a cat"], return_tensors="pt", padding=True)

with torch.no_grad():
    outputs = model(**image_input, **text_input)
    text_probs = (100.0 * outputs.logits_per_image.softmax(dim=-1))

print("Label probs:", text_probs)