app.py

import os
os.system('pip install git+https://github.com/huggingface/transformers --upgrade')

import gradio as gr
from transformers import ImageGPTFeatureExtractor, ImageGPTForCausalImageModeling
import torch
import numpy as np
import requests
from PIL import Image
import matplotlib.pyplot as plt

feature_extractor = ImageGPTFeatureExtractor.from_pretrained("openai/imagegpt-medium")
model = ImageGPTForCausalImageModeling.from_pretrained("openai/imagegpt-medium")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# load image examples
urls = ['https://i.imgflip.com/4/4t0m5.jpg',
        'https://cdn.openai.com/image-gpt/completions/igpt-xl-miscellaneous-2-orig.png',
        'https://cdn.openai.com/image-gpt/completions/igpt-xl-miscellaneous-29-orig.png',
        'https://cdn.openai.com/image-gpt/completions/igpt-xl-openai-cooking-0-orig.png'
        ]
for idx, url in enumerate(urls):
  image = Image.open(requests.get(url, stream=True).raw)
  image.save(f"image_{idx}.png")

def process_image(image):
    # prepare 7 images, shape (7, 1024)
    batch_size = 7
    encoding = feature_extractor([image for _ in range(batch_size)], return_tensors="pt")

    # create primers
    samples = encoding.input_ids.numpy()
    n_px = feature_extractor.size
    clusters = feature_extractor.clusters
    n_px_crop = 16
    primers = samples.reshape(-1,n_px*n_px)[:,:n_px_crop*n_px] # crop top n_px_crop rows. These will be the conditioning tokens
    
    # get conditioned image (from first primer tensor), padded with black pixels to be 32x32
    primers_img = np.reshape(np.rint(127.5 * (clusters[primers[0]] + 1.0)), [n_px_crop,n_px, 3]).astype(np.uint8) 
    primers_img = np.pad(primers_img, pad_width=((0,16), (0,0), (0,0)), mode="constant")
    
    # generate (no beam search)
    context = np.concatenate((np.full((batch_size, 1), model.config.vocab_size - 1), primers), axis=1)
    context = torch.tensor(context).to(device)
    output = model.generate(input_ids=context, max_length=n_px*n_px + 1, temperature=1.0, do_sample=True, top_k=40)

    # decode back to images (convert color cluster tokens back to pixels)
    samples = output[:,1:].cpu().detach().numpy()
    samples_img = [np.reshape(np.rint(127.5 * (clusters[s] + 1.0)), [n_px, n_px, 3]).astype(np.uint8) for s in samples] 
    
    samples_img = [primers_img] + samples_img
    
    # stack images horizontally
    row1 = np.hstack(samples_img[:4])
    row2 = np.hstack(samples_img[4:])
    result = np.vstack([row1, row2])
    
    # return as PIL Image
    completion = Image.fromarray(result)

    return completion

title = "Interactive demo: ImageGPT"
description = "Demo for OpenAI's ImageGPT: Generative Pretraining from Pixels. To use it, simply upload an image or use the example image below and click 'submit'. Results will show up in a few seconds."
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2109.10282'>ImageGPT: Generative Pretraining from Pixels</a> | <a href='https://openai.com/blog/image-gpt/'>Official blog</a></p>"
examples =[f"image_{idx}.png" for idx in range(len(urls))]

iface = gr.Interface(fn=process_image, 
                     inputs=gr.inputs.Image(type="pil"), 
                     outputs=gr.outputs.Image(type="pil", label="Model input + completions"),
                     title=title,
                     description=description,
                     article=article,
                     examples=examples,
                     enable_queue=True)
iface.launch(debug=True)