app.py

import spaces
import gradio as gr
import numpy as np
import requests
import base64
import os
from datetime import datetime
from pytz import timezone

import torch
import diffusers
from diffusers import DDPMPipeline
from transformers import AutoTokenizer, AutoModel

from utils import normalize_formula

tz = timezone('EST')

API_ENDPOINT = os.getenv('API_ENDPOINT')
API_KEY = os.getenv('API_KEY')

print (API_ENDPOINT)
print (API_KEY)

title = "<h1><center>Markup-to-Image Diffusion Models with Scheduled Sampling</center></h1>"
authors = "<center>Yuntian Deng, Noriyuki Kojima, Alexander M. Rush</center>"
info = '<center><a href="https://openreview.net/pdf?id=81VJDmOE2ol">Paper</a> <a href="https://github.com/da03/markup2im">Code</a></center>'
#notice = "<p><center><strong>Notice:</strong> Due to resource constraints, we've transitioned from GPU to CPU processing for this demo, which results in significantly longer inference times. We appreciate your understanding.</center></p>"
notice = "<p><center>Acknowledgment: This demo is powered by GPU resources supported by the Hugging Face Community Grant.</center></p>"


# setup
def setup():
    img_pipe = DDPMPipeline.from_pretrained("yuntian-deng/latex2im_ss_finetunegptneo")
    model_type = "EleutherAI/gpt-neo-125M"
    #encoder = AutoModel.from_pretrained(model_type).to(device)
    encoder = img_pipe.unet.text_encoder
    if False:
        l = len(img_pipe.unet.down_blocks)
        for i in range(l):
            img_pipe.unet.down_blocks[i] = torch.compile(img_pipe.unet.down_blocks[i])
        l = len(img_pipe.unet.up_blocks)
        for i in range(l):
            img_pipe.unet.up_blocks[i] = torch.compile(img_pipe.unet.up_blocks[i])
    tokenizer = AutoTokenizer.from_pretrained(model_type, max_length=1024)
    eos_id = tokenizer.encode(tokenizer.eos_token)[0]
    
    def forward_encoder(latex):
        device = ("cuda" if torch.cuda.is_available() else "cpu")
        img_pipe.to(device)
        encoded = tokenizer(latex, return_tensors='pt', truncation=True, max_length=1024)
        input_ids = encoded['input_ids']
        input_ids = torch.cat((input_ids, torch.LongTensor([eos_id,]).unsqueeze(0)), dim=-1)
        input_ids = input_ids.to(device)
        attention_mask = encoded['attention_mask']
        attention_mask = torch.cat((attention_mask, torch.LongTensor([1,]).unsqueeze(0)), dim=-1)
        attention_mask = attention_mask.to(device)
        with torch.no_grad():
            outputs = encoder(input_ids=input_ids, attention_mask=attention_mask)
            last_hidden_state = outputs.last_hidden_state 
            last_hidden_state = attention_mask.unsqueeze(-1) * last_hidden_state # shouldn't be necessary
        return last_hidden_state
    return img_pipe, forward_encoder

img_pipe, forward_encoder = setup()

with gr.Blocks() as demo:
    gr.Markdown(title)
    gr.Markdown(authors)
    gr.Markdown(info)
    gr.Markdown(notice)
    with gr.Row():
        with gr.Column(scale=2):
            textbox = gr.Textbox(label=r'Type LaTeX formula below and click "Generate"', lines=1, max_lines=1, placeholder='Type LaTeX formula here and click "Generate"', value=r'\sum_{t=1}^T\E_{y_t \sim {\tilde P(y_t| y_0)}} \left\| \frac{y_t - \sqrt{\bar{\alpha}_t}y_0}{\sqrt{1-\bar{\alpha}_t}} - \epsilon_\theta(y_t, t)\right\|^2.')
            submit_btn = gr.Button("Generate", elem_id="btn")
        with gr.Column(scale=3):
            slider = gr.Slider(0, 1000, value=0, label='step (out of 1000)')
            image = gr.Image(label="Rendered Image", show_label=False, elem_id="image")
    inputs = [textbox]
    outputs = [slider, image, submit_btn]

    # duration is set to default to avoid quota issues
    @spaces.GPU(duration=90)
    def infer(formula):
        current_time = datetime.now(tz)
        print (current_time, formula)
        data = {'formula': formula, 'api_key': API_KEY}
        try:
            formula_normalized = normalize_formula(formula)
        except Exception as e:
            print (e)
            formula_normalized = formula
        print ('normalized', formula_normalized)
        encoder_hidden_states = forward_encoder(formula_normalized)

        try:
            i = 0
            results = []
            for _, image_clean in img_pipe.run_clean(batch_size=1, generator=torch.manual_seed(0), encoder_hidden_states=encoder_hidden_states, output_type="numpy"):
                i += 1
                image_clean = image_clean[0]
                image_clean = np.ascontiguousarray(image_clean)
                #s = base64.b64encode(image_clean).decode('ascii')
                #yield s
                q = image_clean
                q = q.reshape((64, 320, 3))
                #print (q.min(), q.max())
                yield i, q, gr.update(visible=False)
            yield i, q, gr.update(visible=True)
            #with requests.post(url=API_ENDPOINT, data=data, timeout=600, stream=True) as r:
            #    i = 0
            #    for line in r.iter_lines():
            #        response = line.decode('ascii').strip()
            #        r = base64.decodebytes(response.encode('ascii'))
            #        q = np.frombuffer(r, dtype=np.float32).reshape((64, 320, 3))
            #        i += 1
            #        yield i, q, gr.update(visible=False)
            #    yield i, q, gr.update(visible=True)
        except Exception as e:
            yield 1000, 255*np.ones((64, 320, 3)), gr.update(visible=True)
    submit_btn.click(fn=infer, inputs=inputs, outputs=outputs, concurrency_limit=1)
demo.queue(max_size=20).launch()