likaixin/IconClip-ViT-L-14

Model Description

A CLIP ViT-B/32 model trained with the IconStack dataset using OpenCLIP.

It scores 80.24% on zero-shot classification on icon-dataset.

Installation

You need to install open_clip to use this model:

pip install open_clip_torch

Icon-to-Text Zero-Shot Classification

import torch
from PIL import Image
import open_clip

CLIP_TEXT_TEMPLATE = "an icon of {}"
ICON_CLASSES = ["add", "close", "play", ...]  # Modify your class names here

model_checkpoint = "<path_to_your_local_model>"
model, _, preprocess = open_clip.create_model_and_transforms('ViT-B-32', pretrained=model_checkpoint)
model.eval()
tokenizer = open_clip.get_tokenizer('ViT-B-32')

image = preprocess(Image.open("icon.png")).unsqueeze(0)
text = tokenizer([CLIP_TEXT_TEMPLATE.format(cls) for cls in ICON_CLASSES])

with torch.no_grad(), torch.autocast("cuda"):
    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)  # prints something like: [[1., 0., 0., ...]]