dataset.py

import torch
import torch.nn.functional as F
from torch.nn.utils.rnn import pad_sequence
import numpy as np
import imagesize
import logging
import glob
import os
from os.path import join
from collections import defaultdict
import pickle
import cv2
from transformers import PreTrainedTokenizerFast
from tqdm.auto import tqdm

from pix2tex.utils.utils import in_model_path
from pix2tex.dataset.transforms import train_transform, test_transform
from multiprocessing import Pool
from functools import partial
import math


class Im2LatexDataset:
    keep_smaller_batches = False
    shuffle = True
    batchsize = 16
    max_dimensions = (1024, 512)
    min_dimensions = (32, 32)
    max_seq_len = 1024
    pad_token = "[PAD]"
    bos_token = "[BOS]"
    eos_token = "[EOS]"
    pad_token_id = 0
    bos_token_id = 1
    eos_token_id = 2
    transform = train_transform
    data = defaultdict(lambda: [])

    def __init__(self, equations=None, images=None, tokenizer=None, shuffle=True, batchsize=16, max_seq_len=1024,
                 max_dimensions=(1024, 512), min_dimensions=(32, 32), pad=False, keep_smaller_batches=False, test=False):
        """Generates a torch dataset from pairs of `equations` and `images`.

        Args:
            equations (str, optional): Path to equations. Defaults to None.
            images (str, optional): Directory where images are saved. Defaults to None.
            tokenizer (str, optional): Path to saved tokenizer. Defaults to None.
            shuffle (bool, opitonal): Defaults to True. 
            batchsize (int, optional): Defaults to 16.
            max_seq_len (int, optional): Defaults to 1024.
            max_dimensions (tuple(int, int), optional): Maximal dimensions the model can handle
            min_dimensions (tuple(int, int), optional): Minimal dimensions the model can handle
            pad (bool): Pad the images to `max_dimensions`. Defaults to False.
            keep_smaller_batches (bool): Whether to also return batches with smaller size than `batchsize`. Defaults to False.
            test (bool): Whether to use the test transformation or not. Defaults to False.
        """

        if images is not None and equations is not None:
            assert tokenizer is not None
            self.images = [path.replace('\\', '/')
                           for path in glob.glob(join(images, '*.png'))]
            self.sample_size = len(self.images)
            eqs = open(equations, 'r').read().split('\n')
            self.indices = [int(os.path.basename(img).split('.')[0])
                            for img in self.images]
            self.tokenizer = PreTrainedTokenizerFast(tokenizer_file=tokenizer)
            self.shuffle = shuffle
            self.batchsize = batchsize
            self.max_seq_len = max_seq_len
            self.max_dimensions = max_dimensions
            self.min_dimensions = min_dimensions
            self.pad = pad
            self.keep_smaller_batches = keep_smaller_batches
            self.test = test
            # check the image dimension for every image and group them together
            try:
                for i, im in tqdm(enumerate(self.images), total=len(self.images)):
                    width, height = imagesize.get(im)
                    if min_dimensions[0] <= width <= max_dimensions[0] and min_dimensions[1] <= height <= max_dimensions[1]:
                        self.data[(width, height)].append(
                            (eqs[self.indices[i]], im))
            except KeyboardInterrupt:
                pass
            self.data = dict(self.data)
            self._get_size()

            iter(self)

    def __len__(self):
        return self.size

    def __iter__(self):
        self.i = 0
        self.transform = test_transform if self.test else train_transform
        self.pairs = []
        for k in self.data:
            info = np.array(self.data[k], dtype=object)
            p = torch.randperm(
                len(info)) if self.shuffle else torch.arange(len(info))
            for i in range(0, len(info), self.batchsize):
                batch = info[p[i:i+self.batchsize]]
                if len(batch.shape) == 1:
                    batch = batch[None, :]
                if len(batch) < self.batchsize and not self.keep_smaller_batches:
                    continue
                self.pairs.append(batch)
        if self.shuffle:
            self.pairs = np.random.permutation(
                np.array(self.pairs, dtype=object))
        else:
            self.pairs = np.array(self.pairs, dtype=object)
        self.size = len(self.pairs)
        return self

    def __next__(self):
        if self.i >= self.size:
            raise StopIteration
        self.i += 1
        return self.prepare_data(self.pairs[self.i-1])

    def prepare_data(self, batch):
        """loads images into memory

        Args:
            batch (numpy.array[[str, str]]): array of equations and image path pairs

        Returns:
            tuple(torch.tensor, torch.tensor): data in memory
        """

        eqs, ims = batch.T
        tok = self.tokenizer(list(eqs), return_token_type_ids=False)
        # pad with bos and eos token
        for k, p in zip(tok, [[self.bos_token_id, self.eos_token_id], [1, 1]]):
            tok[k] = pad_sequence([torch.LongTensor(
                [p[0]]+x+[p[1]]) for x in tok[k]], batch_first=True, padding_value=self.pad_token_id)
        # check if sequence length is too long
        if self.max_seq_len < tok['attention_mask'].shape[1]:
            return next(self)
        # images = []
        # for path in list(ims):
        #     im = cv2.imread(path)
        #     if im is None:
        #         print(path, 'not found!')
        #         continue
        #     im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
        #     if not self.test:
        #         # sometimes convert to bitmask
        #         if np.random.random() < .04:
        #             im[im != 255] = 0
        #     images.append(self.transform(image=im)['image'][:1])
        # images = self.pool.map(self.im_transfrom, ims)
        images = [ele[0]
                  for ele in self.pool.map(partial(self.im_transform,), ims)]
        # output_pool, images = [], []
        # batch_size = math.ceil(len(ims)/self.num_workers)
        # for i in range(self.num_workers):
        #     batch_datum = ims[i*batch_size:(i+1)*batch_size]
        #     output_pool.append(self.pool.apply_async(
        #         func=self.im_transform, args=(batch_datum,)))
        # for i in output_pool:
        #     images.extend(i.get())
        try:
            images = torch.cat(images).float().unsqueeze(1)
        except RuntimeError:
            logging.critical('Images not working: %s' % (' '.join(list(ims))))
            return None, None
        if self.pad:
            h, w = images.shape[2:]
            images = F.pad(
                images, (0, self.max_dimensions[0]-w, 0, self.max_dimensions[1]-h), value=1)
        return tok, images

    def __getstate__(self):
        self_dict = self.__dict__.copy()
        del self_dict['pool']
        return self_dict

    def im_transform(self, ims, **kwargs):
        images = []
        ims = [ims] if isinstance(ims, str) else ims
        for path in ims:
            im = cv2.imread(path)
            if im is None:
                print(path, 'not found!')
                raise FileNotFoundError
            im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
            if not self.test:
                # sometimes convert to bitmask
                if np.random.random() < .04:
                    im[im != 255] = 0
            im = self.transform(image=im)['image'][:1]
            images.append(im)
        print(im[0].shape)
        return images

    def _get_size(self):
        self.size = 0
        for k in self.data:
            div, mod = divmod(len(self.data[k]), self.batchsize)
            self.size += div  # + (1 if mod > 0 else 0)

    def load(self, filename, args=[], **kwargs):
        """returns a pickled version of a dataset

        Args:
            filename (str): Path to dataset
        """
        if not os.path.exists(filename):
            with in_model_path():
                tempf = os.path.join('..', filename)
                if os.path.exists(tempf):
                    filename = os.path.realpath(tempf)
        with open(filename, 'rb') as file:
            x = pickle.load(file)
        x.num_workers = kwargs.get('num_workers', 1)
        x.pool = Pool(processes=x.num_workers)
        return x

    def combine(self, x):
        """Combine Im2LatexDataset with another Im2LatexDataset

        Args:
            x (Im2LatexDataset): Dataset to absorb
        """
        for key in x.data.keys():
            if key in self.data.keys():
                self.data[key].extend(x.data[key])
                self.data[key] = list(set(self.data[key]))
            else:
                self.data[key] = x.data[key]
        self._get_size()
        iter(self)

    def save(self, filename):
        """save a pickled version of a dataset

        Args:
            filename (str): Path to dataset
        """
        with open(filename, 'wb') as file:
            pickle.dump(self, file)

    def update(self, **kwargs):
        for k in ['batchsize', 'shuffle', 'pad', 'keep_smaller_batches', 'test', 'max_seq_len']:
            if k in kwargs:
                setattr(self, k, kwargs[k])
        if 'max_dimensions' in kwargs or 'min_dimensions' in kwargs:
            if 'max_dimensions' in kwargs:
                self.max_dimensions = kwargs['max_dimensions']
            if 'min_dimensions' in kwargs:
                self.min_dimensions = kwargs['min_dimensions']
            temp = {}
            for k in self.data:
                if self.min_dimensions[0] <= k[0] <= self.max_dimensions[0] and self.min_dimensions[1] <= k[1] <= self.max_dimensions[1]:
                    temp[k] = self.data[k]
            self.data = temp
        if 'tokenizer' in kwargs:
            tokenizer_file = kwargs['tokenizer']
            if not os.path.exists(tokenizer_file):
                with in_model_path():
                    tokenizer_file = os.path.realpath(tokenizer_file)
            self.tokenizer = PreTrainedTokenizerFast(
                tokenizer_file=tokenizer_file)
        self._get_size()
        iter(self)


def generate_tokenizer(equations, output, vocab_size):
    from tokenizers import Tokenizer, pre_tokenizers
    from tokenizers.models import BPE
    from tokenizers.trainers import BpeTrainer
    tokenizer = Tokenizer(BPE())
    tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
    trainer = BpeTrainer(special_tokens=[
                         "[PAD]", "[BOS]", "[EOS]"], vocab_size=vocab_size, show_progress=True)
    tokenizer.train(equations, trainer)
    tokenizer.save(path=output, pretty=False)


if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser(description='Train model', add_help=False)
    parser.add_argument('-i', '--images', type=str, nargs='+',
                        default=None, help='Image folders')
    parser.add_argument('-e', '--equations', type=str,
                        nargs='+', default=None, help='equations text files')
    parser.add_argument('-t', '--tokenizer', default=None,
                        help='Pretrained tokenizer file')
    parser.add_argument('-o', '--out', type=str,
                        required=True, help='output file')
    parser.add_argument('-s', '--vocab-size', default=8000,
                        type=int, help='vocabulary size when training a tokenizer')
    args = parser.parse_args()
    if args.tokenizer is None:
        with in_model_path():
            args.tokenizer = os.path.realpath(
                os.path.join('dataset', 'tokenizer.json'))
    if args.images is None and args.equations is not None:
        print('Generate tokenizer')
        generate_tokenizer(args.equations, args.out, args.vocab_size)
    elif args.images is not None and args.equations is not None:
        print('Generate dataset')
        dataset = None
        for images, equations in zip(args.images, args.equations):
            if dataset is None:
                dataset = Im2LatexDataset(equations, images, args.tokenizer)
            else:
                dataset.combine(Im2LatexDataset(
                    equations, images, args.tokenizer))
        dataset.update(batchsize=1, keep_smaller_batches=True)
        dataset.save(args.out)
    else:
        print('Not defined')