app.py

import time
from PIL import Image
import gradio as gr
from glob import glob
import torch
from transformers import AutoModel, AutoProcessor

DEFAULT_EXAMPLE_PATH = f'examples/example_0'

device = "cuda" if torch.cuda.is_available() else "cpu"
weight_dtype = torch.bfloat16 if device == "cuda" else torch.float32
print(f"Using device: {device} ({weight_dtype})")
print("Loading model...")
model_pretrained_name_or_path = "facebook/metaclip-h14-fullcc2.5b"
processor = AutoProcessor.from_pretrained(model_pretrained_name_or_path)
model = AutoModel.from_pretrained(model_pretrained_name_or_path, torch_dtype=weight_dtype).eval().to(device)
print("Model loaded.")


def calc_probs(prompt, images):
    print("Processing inputs...")
    image_inputs = processor(
        images=images,
        padding=True,
        truncation=True,
        max_length=77,
        return_tensors="pt",
    ).to(device)

    image_inputs = {k: v.to(weight_dtype) for k, v in image_inputs.items()}

    text_inputs = processor(
        text=prompt,
        padding=True,
        truncation=True,
        max_length=77,
        return_tensors="pt",
    ).to(device)

    with torch.no_grad():
        print("Embedding images and text...")
        image_embs = model.get_image_features(**image_inputs)
        image_embs = image_embs / torch.norm(image_embs, dim=-1, keepdim=True)

        text_embs = model.get_text_features(**text_inputs)
        text_embs = text_embs / torch.norm(text_embs, dim=-1, keepdim=True)

        print("Calculating scores...")
        scores = model.logit_scale.exp() * (text_embs.float() @ image_embs.float().T)[0]

        print("Calculating probabilities...")
        probs = torch.softmax(scores, dim=-1)

    return probs.cpu().tolist()


def predict(prompt, image_1, image_2):
    print(f"Starting prediction for prompt: {prompt}")
    start_time = time.time()
    probs = calc_probs(prompt, [image_1, image_2])
    print(f"Prediction: {probs} ({time.time() - start_time:.2f} seconds, ) ")
    if device == "cuda":
        print(f"GPU mem used: {round(torch.cuda.max_memory_allocated(device) / 1024 / 1024 / 1024, 2)}/{round(torch.cuda.get_device_properties(device).total_memory / 1024 / 1024 / 1024, 2)} GB")
    return str(round(probs[0], 3)), str(round(probs[1], 3))


with gr.Blocks(title="PickScore v1") as demo:
    gr.Markdown("# PickScore v1")
    gr.Markdown(
        "This is a demo for the PickScore model - see [paper](https://arxiv.org/abs/2305.01569), [code](https://github.com/yuvalkirstain/PickScore), [dataset](https://huggingface.co/datasets/pickapic-anonymous/pickapic_v1), and [model](https://huggingface.co/yuvalkirstain/PickScore_v1).")
    gr.Markdown("## Instructions")
    gr.Markdown("Write a prompt, place two images, and press run to get their PickScore!")
    with gr.Row():
        prompt = gr.inputs.Textbox(lines=1, label="Prompt",
                                   default=open(f'{DEFAULT_EXAMPLE_PATH}/prompt.txt').readline())
    with gr.Row():
        image_1 = gr.components.Image(type="pil", label="image 1",
                                      value=Image.open(f'{DEFAULT_EXAMPLE_PATH}/image_1.png'))
        image_2 = gr.components.Image(type="pil", label="image 2",
                                      value=Image.open(f'{DEFAULT_EXAMPLE_PATH}/image_2.png'))
    with gr.Row():
        pred_1 = gr.outputs.Textbox(label="Probability 1")
        pred_2 = gr.outputs.Textbox(label="Probability 2")

    btn = gr.Button("Run")
    btn.click(fn=predict, inputs=[prompt, image_1, image_2], outputs=[pred_1, pred_2])
    prompt.change(lambda: ("", ""), inputs=[], outputs=[pred_1, pred_2])

    gr.Examples(
        [[open(f'{path}/prompt.txt').readline(), f'{path}/image_1.png', f'{path}/image_2.png'] for path in
         glob(f'examples/*')],
        [prompt, image_1, image_2],
        [pred_1, pred_2],
        predict
    )

demo.queue(concurrency_count=5).launch()