deep-text-recognition-benchmark/dataset.py

import os
import sys
import re
import six
import lmdb
import torch

from PIL import Image
import numpy as np
from torch.utils.data import Dataset, ConcatDataset, Subset
from torch._utils import _accumulate
import torchvision.transforms as transforms


class Batch_Balanced_Dataset(object):

    def __init__(self, opt):
        """
        Modulate the data ratio in the batch.
        For example, when select_data is "MJ-ST" and batch_ratio is "0.5-0.5",
        the 50% of the batch is filled with MJ and the other 50% of the batch is filled with ST.
        """
        print('-' * 80)
        print(f'dataset_root: {opt.train_data}\nopt.select_data: {opt.select_data}\nopt.batch_ratio: {opt.batch_ratio}')
        assert len(opt.select_data) == len(opt.batch_ratio)

        _AlignCollate = AlignCollate(imgH=opt.imgH, imgW=opt.imgW)
        self.data_loader_list = []
        self.dataloader_iter_list = []
        batch_size_list = []
        Total_batch_size = 0
        for selected_d, batch_ratio_d in zip(opt.select_data, opt.batch_ratio):
            _batch_size = max(round(opt.batch_size * float(batch_ratio_d)), 1)
            print('-' * 80)
            _dataset = hierarchical_dataset(root=opt.train_data, opt=opt, select_data=[selected_d])
            total_number_dataset = len(_dataset)

            """
            The total number of data can be modified with opt.total_data_usage_ratio.
            ex) opt.total_data_usage_ratio = 1 indicates 100% usage, and 0.2 indicates 20% usage.
            See 4.2 section in our paper.
            """
            number_dataset = int(total_number_dataset * float(opt.total_data_usage_ratio))
            dataset_split = [number_dataset, total_number_dataset - number_dataset]
            indices = range(total_number_dataset)
            _dataset, _ = [Subset(_dataset, indices[offset - length:offset])
                           for offset, length in zip(_accumulate(dataset_split), dataset_split)]
            print(f'num total samples of {selected_d}: {total_number_dataset} x {opt.total_data_usage_ratio} (total_data_usage_ratio) = {len(_dataset)}')
            print(f'num samples of {selected_d} per batch: {opt.batch_size} x {float(batch_ratio_d)} (batch_ratio) = {_batch_size}')
            batch_size_list.append(str(_batch_size))
            Total_batch_size += _batch_size

            _data_loader = torch.utils.data.DataLoader(
                _dataset, batch_size=_batch_size,
                shuffle=True,
                num_workers=int(opt.workers),
                collate_fn=_AlignCollate, pin_memory=True)
            self.data_loader_list.append(_data_loader)
            self.dataloader_iter_list.append(iter(_data_loader))
        print('-' * 80)
        print('Total_batch_size: ', '+'.join(batch_size_list), '=', str(Total_batch_size))
        opt.batch_size = Total_batch_size
        print('-' * 80)

    def get_batch(self):
        balanced_batch_images = []
        balanced_batch_texts = []

        for i, data_loader_iter in enumerate(self.dataloader_iter_list):
            try:
                image, text = data_loader_iter.next()
                balanced_batch_images.append(image)
                balanced_batch_texts += text
            except StopIteration:
                self.dataloader_iter_list[i] = iter(self.data_loader_list[i])
                image, text = self.dataloader_iter_list[i].next()
                balanced_batch_images.append(image)
                balanced_batch_texts += text
            except ValueError:
                pass

        balanced_batch_images = torch.cat(balanced_batch_images, 0)

        return balanced_batch_images, balanced_batch_texts


def hierarchical_dataset(root, opt, select_data='/'):
    """ select_data='/' contains all sub-directory of root directory """
    dataset_list = []
    print(f'dataset_root:    {root}\t dataset: {select_data[0]}')
    for dirpath, dirnames, filenames in os.walk(root):
        if not dirnames:
            select_flag = False
            for selected_d in select_data:
                if selected_d in dirpath:
                    select_flag = True
                    break

            if select_flag:
                dataset = LmdbDataset(dirpath, opt)
                print(f'sub-directory:\t/{os.path.relpath(dirpath, root)}\t num samples: {len(dataset)}')
                dataset_list.append(dataset)

    concatenated_dataset = ConcatDataset(dataset_list)

    return concatenated_dataset


class LmdbDataset(Dataset):

    def __init__(self, root, opt):

        self.root = root
        self.opt = opt
        self.env = lmdb.open(root, max_readers=32, readonly=True, lock=False, readahead=False, meminit=False)
        if not self.env:
            print('cannot create lmdb from %s' % (root))
            sys.exit(0)

        with self.env.begin(write=False) as txn:
            nSamples = int(txn.get('num-samples'.encode()))
            self.nSamples = nSamples

    def __len__(self):
        return self.nSamples

    def __getitem__(self, index):
        assert index <= len(self), 'index range error'
        index += 1
        with self.env.begin(write=False) as txn:
            label_key = 'label-%09d'.encode() % index
            label = txn.get(label_key).decode('utf-8')
            img_key = 'image-%09d'.encode() % index
            imgbuf = txn.get(img_key)

            buf = six.BytesIO()
            buf.write(imgbuf)
            buf.seek(0)
            try:
                if self.opt.rgb:
                    img = Image.open(buf).convert('RGB')  # for color image
                else:
                    img = Image.open(buf).convert('L')

            except IOError:
                print(f'Corrupted image for {index}')
                return

            if len(label) > self.opt.batch_max_length:
                print(f'The length of the label is longer than max_length: length {len(label)}, {label} in dataset {self.root}')
                return

            if not self.opt.sensitive:
                label = label.lower()

            # We only train and evaluate on alphanumerics (or pre-defined character set in train.py)
            out_of_char = f'[^{self.opt.character}]'
            label = re.sub(out_of_char, '', label)

        return (img, label)


class ResizeNormalize(object):

    def __init__(self, size, interpolation=Image.BICUBIC):
        self.size = size
        self.interpolation = interpolation
        self.toTensor = transforms.ToTensor()

    def __call__(self, img):
        img = img.resize(self.size, self.interpolation)
        img = self.toTensor(img)
        img.sub_(0.5).div_(0.5)
        return img


class AlignCollate(object):

    def __init__(self, imgH=32, imgW=100):
        self.imgH = imgH
        self.imgW = imgW

    def __call__(self, batch):
        batch = filter(lambda x: x is not None, batch)
        images, labels = zip(*batch)

        transform = ResizeNormalize((self.imgW, self.imgH))
        image_tensors = [transform(image) for image in images]
        image_tensors = torch.cat([t.unsqueeze(0) for t in image_tensors], 0)

        return image_tensors, labels


def tensor2im(image_tensor, imtype=np.uint8):
    image_numpy = image_tensor.cpu().float().numpy()
    if image_numpy.shape[0] == 1:
        image_numpy = np.tile(image_numpy, (3, 1, 1))
    image_numpy = (np.transpose(image_numpy, (1, 2, 0)) + 1) / 2.0 * 255.0
    return image_numpy.astype(imtype)


def save_image(image_numpy, image_path):
    image_pil = Image.fromarray(image_numpy)
    image_pil.save(image_path)