pix2tex.py

from dataset.dataset import test_transform
import cv2
import pandas.io.clipboard as clipboard
from PIL import ImageGrab
from PIL import Image
import os
import sys
import argparse
import logging
import yaml
import re

import numpy as np
import torch
from torchvision import transforms
from munch import Munch
from transformers import PreTrainedTokenizerFast
from timm.models.resnetv2 import ResNetV2
from timm.models.layers import StdConv2dSame

from dataset.latex2png import tex2pil
from models import get_model
from utils import *
from checkpoints.get_latest_checkpoint import download_checkpoints

last_pic = None


def minmax_size(img, max_dimensions=None, min_dimensions=None):
    if max_dimensions is not None:
        ratios = [a/b for a, b in zip(img.size, max_dimensions)]
        if any([r > 1 for r in ratios]):
            size = np.array(img.size)//max(ratios)
            img = img.resize(size.astype(int), Image.BILINEAR)
    if min_dimensions is not None:
        if any([s < min_dimensions[i] for i, s in enumerate(img.size)]):
            padded_im = Image.new('L', min_dimensions, 255)
            padded_im.paste(img, img.getbbox())
            img = padded_im
    return img


def initialize(arguments=None):
    if arguments is None:
        arguments = Munch({'config': 'settings/config.yaml', 'checkpoint': 'checkpoints/weights.pth', 'no_cuda': True, 'no_resize': False})
    logging.getLogger().setLevel(logging.FATAL)
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
    with open(arguments.config, 'r') as f:
        params = yaml.load(f, Loader=yaml.FullLoader)
    args = parse_args(Munch(params))
    args.update(**vars(arguments))
    args.wandb = False
    # args.device = "cpu"
    args.device = 'cuda' if torch.cuda.is_available() and not args.no_cuda else 'cpu'
    if not os.path.exists(args.checkpoint):
        download_checkpoints()
    model = get_model(args)
    model.load_state_dict(torch.load(args.checkpoint, map_location=args.device))

    if 'image_resizer.pth' in os.listdir(os.path.dirname(args.checkpoint)) and not arguments.no_resize:
        image_resizer = ResNetV2(layers=[2, 3, 3], num_classes=max(args.max_dimensions)//32, global_pool='avg', in_chans=1, drop_rate=.05,
                                 preact=True, stem_type='same', conv_layer=StdConv2dSame).to(args.device)
        image_resizer.load_state_dict(torch.load(os.path.join(os.path.dirname(args.checkpoint), 'image_resizer.pth'), map_location=args.device))
        image_resizer.eval()
    else:
        image_resizer = None
    tokenizer = PreTrainedTokenizerFast(tokenizer_file=args.tokenizer)
    return args, model, image_resizer, tokenizer


def call_model(args, model, image_resizer, tokenizer, img=None):
    global last_pic
    encoder, decoder = model.encoder, model.decoder
    if type(img) is bool:
        img = None
    if img is None:
        if last_pic is None:
            print('Provide an image.')
            return ''
        else:
            img = last_pic.copy()
    else:
        last_pic = img.copy()
    img = minmax_size(pad(img), args.max_dimensions, args.min_dimensions)
    if image_resizer is not None and not args.no_resize:
        with torch.no_grad():
            input_image = img.convert('RGB').copy()
            r, w, h = 1, input_image.size[0], input_image.size[1]
            for _ in range(10):
                h = int(h * r)  # height to resize
                img = pad(minmax_size(input_image.resize((w, h), Image.BILINEAR if r > 1 else Image.LANCZOS), args.max_dimensions, args.min_dimensions))
                t = test_transform(image=np.array(img.convert('RGB')))['image'][:1].unsqueeze(0)
                w = (image_resizer(t.to(args.device)).argmax(-1).item()+1)*32
                logging.info(r, img.size, (w, int(input_image.size[1]*r)))
                if (w == img.size[0]):
                    break
                r = w/img.size[0]
    else:
        img = np.array(pad(img).convert('RGB'))
        t = test_transform(image=img)['image'][:1].unsqueeze(0)
    im = t.to(args.device)

    with torch.no_grad():
        model.eval()
        device = args.device
        encoded = encoder(im.to(device))
        dec = decoder.generate(torch.LongTensor([args.bos_token])[:, None].to(device), args.max_seq_len,
                               eos_token=args.eos_token, context=encoded.detach(), temperature=args.get('temperature', .25))
        pred = post_process(token2str(dec, tokenizer)[0])
    try:
        clipboard.copy(pred)
    except:
        pass
    return pred


def output_prediction(pred, args):
    print(pred, '\n')
    if args.show or args.katex:
        try:
            if args.katex:
                raise ValueError
            tex2pil([f'$${pred}$$'])[0].show()
        except Exception as e:
            # render using katex
            import webbrowser
            from urllib.parse import quote
            url = 'https://katex.org/?data=' + \
                quote('{"displayMode":true,"leqno":false,"fleqn":false,"throwOnError":true,"errorColor":"#cc0000",\
"strict":"warn","output":"htmlAndMathml","trust":false,"code":"%s"}' % pred.replace('\\', '\\\\'))
            webbrowser.open(url)


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Use model', add_help=False)
    parser.add_argument('-t', '--temperature', type=float, default=.333, help='Softmax sampling frequency')
    parser.add_argument('-c', '--config', type=str, default='settings/config.yaml')
    parser.add_argument('-m', '--checkpoint', type=str, default='checkpoints/weights.pth')
    parser.add_argument('-s', '--show', action='store_true', help='Show the rendered predicted latex code')
    parser.add_argument('-f', '--file', type=str, default=None, help='Predict LaTeX code from image file instead of clipboard')
    parser.add_argument('-k', '--katex', action='store_true', help='Render the latex code in the browser')
    parser.add_argument('--no-cuda', action='store_true', help='Compute on CPU')
    parser.add_argument('--no-resize', action='store_true', help='Resize the image beforehand')
    arguments = parser.parse_args()
    latexocr_path = os.path.dirname(sys.argv[0])
    if latexocr_path != '':
        sys.path.insert(0, latexocr_path)
        os.chdir(latexocr_path)

    args, *objs = initialize(arguments)
    while True:
        instructions = input('Predict LaTeX code for image ("?"/"h" for help). ')
        possible_file = instructions.strip()
        ins = possible_file.lower()
        if ins == 'x':
            break
        elif ins in ['?', 'h', 'help']:
            print('''pix2tex help:

Usage:
    On Windows and macOS you can copy the image into memory and just press ENTER to get a prediction.
    Alternatively you can paste the image file path here and submit.

    You might get a different prediction every time you submit the same image. If the result you got was close you
    can just predict the same image by pressing ENTER again. If that still does not work you can change the temperature
    or you have to take another picture with another resolution (e.g. zoom out and take a screenshot with lower resolution). 

    Press "x" to close the program.
    You can interrupt the model if it takes too long by pressing Ctrl+C.

Visualization:
    You can either render the code into a png using XeLaTeX (see README) to get an image file back.
    This is slow and requires a working installation of XeLaTeX. To activate type 'show' or set the flag --show
    Alternatively you can render the expression in the browser using katex.org. Type 'katex' or set --katex

Settings:
    to toggle one of these settings: 'show', 'katex', 'no_resize' just type it into the console
    Change the temperature (default=0.333) type: "t=0.XX" to set a new temperature.
                ''')
            continue
        elif ins in ['show', 'katex', 'no_resize']:
            setattr(args, ins, not getattr(args, ins, False))
            print('set %s to %s' % (ins, getattr(args, ins)))
            continue
        elif os.path.isfile(os.path.realpath(possible_file)):
            args.file = possible_file
        else:
            t = re.match(r't=([\.\d]+)', ins)
            if t is not None:
                t = t.groups()[0]
                args.temperature = float(t)+1e-8
                print('new temperature: T=%.3f' % args.temperature)
                continue
        try:
            img = None
            if args.file:
                img = Image.open(args.file)
            else:
                try:
                    img = ImageGrab.grabclipboard()
                except:
                    pass
            pred = call_model(args, *objs, img=img)
            output_prediction(pred, args)
        except KeyboardInterrupt:
            pass
        args.file = None