deep-text-recognition-bench.../dataset.py

197 lines
7.0 KiB
Python

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 = []
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}')
_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)
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:
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)