app.py

import os
import open_clip

from PIL import Image
import time
import numpy as np
import torch

model_name = "ViT-B-32-quickgelu"
pretrained_path = "laion400m_e32"
clip_model, _, clip_preprocess = open_clip.create_model_and_transforms(model_name, pretrained_path)
tokenizer = open_clip.get_tokenizer(model_name)
# device = torch.device("cuda:0")
# clip_model = clip_model.to(device)

def clip_inference(image_path, texts):
    start = time.time()
    images = clip_preprocess(Image.open(image_path)).unsqueeze(0)
    # images = images.to(device)
    texts = [element.strip() for element in texts.split(",")]
    # texts_token = tokenizer(texts).to(device)
    texts_token = tokenizer(texts)
    with torch.no_grad():
        image_features = clip_model.encode_image(images)
        image_features /= image_features.norm(dim=-1, keepdim=True)
        text_features = clip_model.encode_text(texts_token)
        text_features /= text_features.norm(dim=-1, keepdim=True)
        text_probs = (100.0 * image_features @ text_features.T)
        text_probs = text_probs.cpu().numpy().tolist()

    results = []
    for text_prob in text_probs:
        for confidence, text in zip(text_prob, texts):
            result = {text: confidence}
            results.append(result)
    print("classification time:", time.time()-start)
    return results

import gradio as gr

block = gr.Blocks().queue()
with block:
    gr.Markdown("# Anomaly Detection Demo")
    gr.Markdown("### Using OVOD and OVIC")
    
    with gr.Tab("OVIC"):
        with gr.Row():
            with gr.Column():
                clip_input_image = gr.Image(source='upload', type="filepath")
                clip_prompt= gr.Textbox(label="Classification Prompt")
                clip_run_button = gr.Button(label="Run")
            with gr.Column():
                clip_output = gr.outputs.Textbox(label="Output")

        clip_run_button.click(fn=clip_inference, inputs=[clip_input_image, clip_prompt], outputs=[clip_output])

block.launch()