|
| 1 | +from argparse import Namespace |
| 2 | + |
| 3 | +import albumentations as A |
| 4 | +import cv2 |
| 5 | +import ignite.distributed as idist |
| 6 | +import numpy as np |
| 7 | +import torch |
| 8 | +from albumentations.pytorch import ToTensorV2 as ToTensor |
| 9 | +from ignite.utils import convert_tensor |
| 10 | +from PIL import Image |
| 11 | +from torch.utils.data import Dataset |
| 12 | +from torchvision.datasets.voc import VOCSegmentation |
| 13 | + |
| 14 | + |
| 15 | +class TransformedDataset(Dataset): |
| 16 | + def __init__(self, ds, transform_fn): |
| 17 | + assert isinstance(ds, Dataset) |
| 18 | + assert callable(transform_fn) |
| 19 | + self.ds = ds |
| 20 | + self.transform_fn = transform_fn |
| 21 | + |
| 22 | + def __len__(self): |
| 23 | + return len(self.ds) |
| 24 | + |
| 25 | + def __getitem__(self, index): |
| 26 | + dp = self.ds[index] |
| 27 | + return self.transform_fn(**dp) |
| 28 | + |
| 29 | + |
| 30 | +class VOCSegmentationPIL(VOCSegmentation): |
| 31 | + |
| 32 | + target_names = [ |
| 33 | + "background", |
| 34 | + "aeroplane", |
| 35 | + "bicycle", |
| 36 | + "bird", |
| 37 | + "boat", |
| 38 | + "bottle", |
| 39 | + "bus", |
| 40 | + "car", |
| 41 | + "cat", |
| 42 | + "chair", |
| 43 | + "cow", |
| 44 | + "diningtable", |
| 45 | + "dog", |
| 46 | + "horse", |
| 47 | + "motorbike", |
| 48 | + "person", |
| 49 | + "plant", |
| 50 | + "sheep", |
| 51 | + "sofa", |
| 52 | + "train", |
| 53 | + "tv/monitor", |
| 54 | + ] |
| 55 | + |
| 56 | + def __init__(self, *args, return_meta=False, **kwargs): |
| 57 | + super().__init__(*args, **kwargs) |
| 58 | + self.return_meta = return_meta |
| 59 | + |
| 60 | + def __getitem__(self, index): |
| 61 | + img = np.asarray(Image.open(self.images[index]).convert("RGB")) |
| 62 | + assert img is not None, f"Image at '{self.images[index]}' has a problem" |
| 63 | + mask = np.asarray(Image.open(self.masks[index])) |
| 64 | + |
| 65 | + if self.return_meta: |
| 66 | + return { |
| 67 | + "image": img, |
| 68 | + "mask": mask, |
| 69 | + "meta": { |
| 70 | + "index": index, |
| 71 | + "image_path": self.images[index], |
| 72 | + "mask_path": self.masks[index], |
| 73 | + }, |
| 74 | + } |
| 75 | + |
| 76 | + return {"image": img, "mask": mask} |
| 77 | + |
| 78 | + |
| 79 | +def setup_data(config: Namespace): |
| 80 | + dataset_train = VOCSegmentationPIL( |
| 81 | + root=config.data_path, year="2012", image_set="train", download=False |
| 82 | + ) |
| 83 | + dataset_eval = VOCSegmentationPIL( |
| 84 | + root=config.data_path, year="2012", image_set="val", download=False |
| 85 | + ) |
| 86 | + |
| 87 | + val_img_size = 513 |
| 88 | + train_img_size = 480 |
| 89 | + |
| 90 | + mean = (0.485, 0.456, 0.406) |
| 91 | + std = (0.229, 0.224, 0.225) |
| 92 | + |
| 93 | + transform_train = A.Compose( |
| 94 | + [ |
| 95 | + A.RandomScale( |
| 96 | + scale_limit=(0.0, 1.5), interpolation=cv2.INTER_LINEAR, p=1.0 |
| 97 | + ), |
| 98 | + A.PadIfNeeded( |
| 99 | + val_img_size, val_img_size, border_mode=cv2.BORDER_CONSTANT |
| 100 | + ), |
| 101 | + A.RandomCrop(train_img_size, train_img_size), |
| 102 | + A.HorizontalFlip(), |
| 103 | + A.Blur(blur_limit=3), |
| 104 | + A.Normalize(mean=mean, std=std), |
| 105 | + ignore_mask_boundaries, |
| 106 | + ToTensor(), |
| 107 | + ] |
| 108 | + ) |
| 109 | + |
| 110 | + transform_eval = A.Compose( |
| 111 | + [ |
| 112 | + A.PadIfNeeded( |
| 113 | + val_img_size, val_img_size, border_mode=cv2.BORDER_CONSTANT |
| 114 | + ), |
| 115 | + A.Normalize(mean=mean, std=std), |
| 116 | + ignore_mask_boundaries, |
| 117 | + ToTensor(), |
| 118 | + ] |
| 119 | + ) |
| 120 | + |
| 121 | + dataset_train = TransformedDataset( |
| 122 | + dataset_train, transform_fn=transform_train |
| 123 | + ) |
| 124 | + dataset_eval = TransformedDataset(dataset_eval, transform_fn=transform_eval) |
| 125 | + |
| 126 | + dataloader_train = idist.auto_dataloader( |
| 127 | + dataset_train, |
| 128 | + shuffle=True, |
| 129 | + batch_size=config.train_batch_size, |
| 130 | + num_workers=config.num_workers, |
| 131 | + drop_last=True, |
| 132 | + ) |
| 133 | + dataloader_eval = idist.auto_dataloader( |
| 134 | + dataset_eval, |
| 135 | + shuffle=False, |
| 136 | + batch_size=config.train_batch_size, |
| 137 | + num_workers=config.num_workers, |
| 138 | + drop_last=False, |
| 139 | + ) |
| 140 | + |
| 141 | + return dataloader_train, dataloader_eval |
| 142 | + |
| 143 | + |
| 144 | +def ignore_mask_boundaries(force_apply, **kwargs): |
| 145 | + assert "mask" in kwargs, "Input should contain 'mask'" |
| 146 | + mask = kwargs["mask"] |
| 147 | + mask[mask == 255] = 0 |
| 148 | + kwargs["mask"] = mask |
| 149 | + return kwargs |
| 150 | + |
| 151 | + |
| 152 | +def denormalize(t, mean, std, max_pixel_value=255): |
| 153 | + assert isinstance(t, torch.Tensor), f"{type(t)}" |
| 154 | + assert t.ndim == 3 |
| 155 | + d = t.device |
| 156 | + mean = torch.tensor(mean, device=d).unsqueeze(-1).unsqueeze(-1) |
| 157 | + std = torch.tensor(std, device=d).unsqueeze(-1).unsqueeze(-1) |
| 158 | + tensor = std * t + mean |
| 159 | + tensor *= max_pixel_value |
| 160 | + return tensor |
| 161 | + |
| 162 | + |
| 163 | +def prepare_image_mask(batch, device, non_blocking): |
| 164 | + x, y = batch["image"], batch["mask"] |
| 165 | + x = convert_tensor(x, device, non_blocking=non_blocking) |
| 166 | + y = convert_tensor(y, device, non_blocking=non_blocking).long() |
| 167 | + return x, y |
| 168 | + |
| 169 | + |
| 170 | +def download_datasets(data_path): |
| 171 | + VOCSegmentation(data_path, image_set="train", download=True) |
| 172 | + VOCSegmentation(data_path, image_set="val", download=True) |
0 commit comments