File size: 20,598 Bytes
dae5c90 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 | import os
import random
from torchvision import datasets, transforms
from timm.data.constants import \
IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
from timm.data import create_transform
import pandas as pd
from torch.utils.data import Dataset
from torchvision import transforms
from PIL import Image
from sklearn.model_selection import train_test_split
import cv2
import numpy as np
from .data_processing import ita_to_group
class LocalISICDataset(Dataset):
def __init__(self, root, transform = None, skin_color_csv = None, augment_transforms = None, split = 'train', cielab=False, skin_former = False, segment_out_skin = False):
"""
Args:
root (str or ``pathlib.Path``): Root directory path.
transform (callable, optional): A function/transform that takes in a PIL image
and returns a transformed version. E.g, ``transforms.RandomCrop``
skin_color_csv (str): Path to the CSV file containing information about each image ita, fitzpatrick scale and group it belongs
- ita: The individual typology angle (ITA) is a measure of the skin's reaction to sun exposure.
- fitzpatrick_scale: The Fitzpatrick scale is a numerical classification schema for human skin color.
- group : Each patient is classified into one of the groups based on his skin color type.
augment_transforms (dict): Dictionary containing augmentations to be applied on malignant cases.
transform (callable, optional): Optional transform to be applied on an image.
split (str): 'train' or 'valid', determines which subset to use.
cielab (bool): If True, convert images to CIELAB color space.
skin_former (bool): If True, use skin-former augmentation.
segment_out_skin (bool): If True, segment out skin from the image.
"""
self.root = root
self.transform = transform
self.split = split
self.skin_fomer = skin_former
self.segment_out_skin = segment_out_skin
# Probability of applying skin transformations to dark
self.group_shift_probs = [0.2, 0.4, 0, 0]
if augment_transforms is None:
self.augment_transforms = None
self.oversample_ratio = 1
else:
self.augment_transforms = augment_transforms
self.oversample_ratio = len(self.augment_transforms.keys())
split_dir = os.path.join(self.root, split)
benign_dir = os.path.join(split_dir, 'benign')
malignant_dir = os.path.join(split_dir, 'malignant')
benign_mask_dir = os.path.join(self.root, 'masks', split, 'benign')
malignant_mask_dir = os.path.join(self.root, 'masks', split, 'malignant')
if os.path.exists(benign_mask_dir) and os.path.exists(malignant_mask_dir):
benign_masks = [os.path.join(benign_mask_dir, mask) for mask in os.listdir(benign_mask_dir)
if mask.lower().endswith(('.jpg'))]
malignant_masks = [os.path.join(malignant_mask_dir, mask) for mask in os.listdir(malignant_mask_dir)
if mask.lower().endswith(('.jpg'))]
benign_masks = sorted(benign_masks)
malignant_masks = sorted(malignant_masks)
else:
benign_masks = [None] * len(os.listdir(benign_dir))
malignant_masks = [None] * len(os.listdir(malignant_dir))
benign_images = [(os.path.join(benign_dir, img), 0) for img in os.listdir(benign_dir)
if img.lower().endswith(('.jpg'))]
malignant_images = [(os.path.join(malignant_dir, img), 1) for img in os.listdir(malignant_dir)
if img.lower().endswith(('.jpg'))]
benign_images = sorted(benign_images, key=lambda x: x[0])
malignant_images = sorted(malignant_images, key=lambda x: x[0])
self.samples = []
self.samples.extend(zip(benign_images, benign_masks))
self.samples.extend(zip(malignant_images, malignant_masks))
self.benign_count = len(benign_images)
self.malignant_count = len(malignant_images)
self.classes = 2
if self.augment_transforms is None:
self.class_distribution = (self.benign_count, self.malignant_count)
else:
self.class_distribution = (self.benign_count, self.malignant_count * self.oversample_ratio)
self.use_cielab = cielab
if skin_color_csv is not None:
self.skin_data = pd.read_csv(skin_color_csv, sep=';')
self.samples_with_skin = []
skin_info_dict = {}
for _, row in self.skin_data.iterrows():
img_name = row['image_name']
if 'group' in self.skin_data.columns:
skin_info_dict[img_name] = {'ita': row['ita'], 'group': row['group']}
missing = 0
for (path, label), mask in self.samples:
img_filename = os.path.basename(path)
if img_filename in skin_info_dict:
if not skin_former and not segment_out_skin:
self.samples_with_skin.append((path, label, skin_info_dict[img_filename]['group']))
else:
self.samples_with_skin.append((path, label, skin_info_dict[img_filename]['group'], skin_info_dict[img_filename]['ita'], mask))
else:
missing+=1
self.samples = self.samples_with_skin
self.groups = 1
if skin_color_csv is not None:
self.groups = len(self.skin_data['group'].unique())
self.group = self.skin_data['group'].value_counts()
def __len__(self):
if self.augment_transforms is None or self.split != 'train':
return len(self.samples)
if self.split == 'train':
return self.benign_count + self.malignant_count * self.oversample_ratio
def __getitem__(self, idx):
if self.augment_transforms is not None and self.split == 'train' and idx >= self.benign_count:
adjusted_idx = self.benign_count + ((idx - self.benign_count) // self.oversample_ratio)
augment_type_idx = (idx - self.benign_count) % self.oversample_ratio
augment_type = list(self.augment_transforms.keys())[augment_type_idx]
if adjusted_idx >= len(self.samples):
adjusted_idx = self.benign_count + ((adjusted_idx - self.benign_count) % self.malignant_count)
else:
adjusted_idx = idx
augment_type = "original"
if isinstance(self.samples[adjusted_idx], tuple) and len(self.samples[adjusted_idx]) == 3 and not self.skin_fomer:
image_path, target, group = self.samples[adjusted_idx]
elif isinstance(self.samples[adjusted_idx], tuple) and len(self.samples[adjusted_idx]) == 5 and (self.skin_fomer or self.segment_out_skin):
image_path, target, group, ita, mask = self.samples[adjusted_idx]
else:
image_path, target = self.samples[adjusted_idx]
group = -1
if not self.use_cielab:
image = Image.open(image_path).convert('RGB')
elif (self.use_cielab and not self.skin_fomer and not self.segment_out_skin) or (self.split != 'train' and not self.segment_out_skin):
image = Image.fromarray(cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2LAB))
elif self.use_cielab and self.skin_fomer:
np_image = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2LAB)
shift_prob = self.group_shift_probs[group]
if random.random() < shift_prob:
# Apply transformation
mask = 1 - cv2.imread(mask)[:, :, 0] / 255
assert len(mask.shape) == 2 or mask.shape[-1] == 1, "Mask has to be grayscale"
assert mask.min() >= 0 and mask.max() <= 1, "Mask values have to be in [0,1]"
# Max ita in our darkest groups (28)
target_ita = random.random() * 38 - 10
delta_ita = float(ita.replace(",", ".")) - target_ita
np_image[:, :, 2] += (mask * delta_ita * 0.5).astype(np.uint) # Shift b
np_image[:, :, 0] -= (mask * delta_ita * 0.12).astype(np.uint) # Shift L
# New label
group = ita_to_group(target_ita)
image = Image.fromarray(np_image)
else:
np_image = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2LAB)
mask = cv2.imread(mask)[:, :, 0] / 255
np_image = np_image.astype(np.float32)
np_image *= mask[:, :, np.newaxis]
np_image = np.clip(np_image, 0, 255).astype(np.uint8)
image = Image.fromarray(np_image)
if target == 1 and self.split == 'train' and augment_type != "original" and self.augment_transforms is not None:
image = self.augment_transforms[augment_type](image)
if self.transform:
image = self.transform(image)
return image, target, group
def get_class_distribution(self):
return self.class_distribution
def get_labels(self):
if self.augment_transforms is None or self.split != 'train':
labels = []
for sample in self.samples:
if isinstance(sample, tuple) and len(sample) >= 2:
labels.append(sample[1])
else:
img_data, _ = sample
_, target = img_data
labels.append(target)
return labels
else:
labels = []
for idx in range(len(self)):
if idx < self.benign_count:
sample = self.samples[idx]
else:
adjusted_idx = self.benign_count + ((idx - self.benign_count) // self.oversample_ratio)
if adjusted_idx >= len(self.samples):
adjusted_idx = self.benign_count + ((adjusted_idx - self.benign_count) % self.malignant_count)
sample = self.samples[adjusted_idx]
if isinstance(sample, tuple) and len(sample) >= 2:
labels.append(sample[1])
else:
img_data, _ = sample
_, target = img_data
labels.append(target)
return labels
class KaggleISICDataset(Dataset):
def __init__(self, csv_file, image_dir, skin_color_csv = None, transform=None, augment_transforms = None,split='train', test_size=0.2,seed=42):
"""
Args:
csv_file (str): Path to the CSV file containing image names and targets.
image_dir (str): Directory containing the image files.
skin_color_csv (str): Path to the CSV file containing information about each image ita, fitzpatrick scale and group it belongs
- ita: The individual typology angle (ITA) is a measure of the skin's reaction to sun exposure.
- fitzpatrick_scale: The Fitzpatrick scale is a numerical classification schema for human skin color.
- group : Each patient is classified into one of the groups based on his skin color type.
transform (callable, optional): Optional transform to be applied on an image.
augment_transforms (dict): Dictionary containing augmentations to be applied on malignant cases.
split (str): 'train' or 'valid', determines which subset to use.
test_size (float): Proportion of the dataset to allocate for validation.
seed (int): Random seed for reproducibility.
"""
self.image_dir = image_dir
self.transform = transform
self.split = split
if augment_transforms is None:
self.augment_transforms = None
self.oversample_ratio = 1
else:
self.augment_transforms = augment_transforms
self.oversample_ratio = len(self.augment_transforms.keys())
if skin_color_csv is not None:
print(f"Skin color csv is defined. Using predefined train/valid splits.")
df = pd.read_csv(skin_color_csv, sep=';')
train_df = df[df['split'] == "train"]
valid_df = df[df['split'] == "train"]
else:
df = pd.read_csv(csv_file)
train_df, valid_df = train_test_split(
df, test_size=test_size, stratify=df['target'], random_state=seed
)
self.data = train_df if split == 'train' else valid_df
if skin_color_csv is not None:
self.samples = [(row['image_name'], row['target'], row['group']) for _, row in self.data.iterrows()]
self.groups = len(self.data['group'].unique())
self.group = self.data['group'].value_counts()
else:
self.samples = [(row['image_name'], row['target'], None ) for _, row in self.data.iterrows()]
self.classes = 2
if self.augment_transforms is None:
self.class_distribution = (len(self.data[self.data['target'] == 0]), len(self.data[self.data['target'] == 1]))
else:
self.class_distribution = (len(self.data[self.data['target'] == 0]), len(self.data[self.data['target'] == 1]) * self.oversample_ratio)
def __len__(self):
if self.augment_transforms is None or self.split != 'train':
return len(self.data)
return len(self.data[self.data['target'] == 0]) + self.oversample_ratio * len(self.data[self.data['target'] == 1])
def __getitem__(self, idx):
if self.augment_transforms is not None and self.split == 'train':
actual_idx = min(idx // self.oversample_ratio, len(self.data) - 1) if idx % self.oversample_ratio != 0 else min(idx, len(self.data) - 1)
row = self.data.iloc[actual_idx]
augment_type = list(self.augment_transforms.keys())[idx % self.oversample_ratio]
else:
row = self.data.iloc[idx]
augment_type = "original"
image_name = row['image_name']
if not image_name.lower().endswith('.jpg'):
image_name = image_name + '.jpg'
image_path = os.path.join(self.image_dir, image_name)
target = row['target']
group = row['group'] if 'group' in row.index else -1
image = Image.open(image_path).convert('RGB')
if target == 1 and self.split == 'train' and self.augment_transforms is not None:
image = self.augment_transforms[augment_type](image)
if self.transform:
image = self.transform(image)
return image, target, group
def build_dataset(is_train, args, transform=None, nb_classes=2):
if not args.convert_to_ffcv and transform is None:
transform = build_transform(is_train, args)
print("Transform = ")
if isinstance(transform, tuple):
for trans in transform:
print(" - - - - - - - - - - ")
for t in trans.transforms:
print(t)
else:
for t in transform.transforms:
print(t)
print("---------------------------")
else:
import warnings
warnings.warn("As the transformations are built using factory method timm.data.create_transform,\
such a factory method should be implemented for ffcv module as well.\
Currently transformations cannot be dynamically allocated for ffcv module \
and any ffcv dataset that is built will contain no augmented images, if not specified here otherwise.",
UserWarning)
if args.data_set == 'CIFAR':
if not args.convert_to_ffcv :
dataset = datasets.CIFAR100(args.data_path, train=is_train, transform=transform, download=True)
else:
dataset = datasets.CIFAR100(args.data_path, train=is_train, download=True)
elif args.data_set == 'IMNET':
root = os.path.join(args.data_path, 'train' if is_train else 'val')
if not args.convert_to_ffcv :
dataset = datasets.ImageFolder(root, transform=transform)
else:
dataset = datasets.ImageFolder(root)
elif args.data_set == 'IMAGENET1K':
root = os.path.join(args.data_path, 'train' if is_train else 'val')
if not args.convert_to_ffcv :
dataset = datasets.ImageFolder(root, transform=transform)
else:
dataset = datasets.ImageFolder(root)
assert args.nb_classes == nb_classes
elif args.data_set == 'IMAGENET100':
root = os.path.join(args.data_path, 'train' if is_train else 'val')
if not args.convert_to_ffcv :
dataset = datasets.ImageFolder(root, transform=transform)
else:
dataset = datasets.ImageFolder(root)
assert args.nb_classes == nb_classes
elif args.data_set == 'TINY_IMAGENET':
root = os.path.join(args.data_path, 'train' if is_train else 'val')
if not args.convert_to_ffcv :
dataset = datasets.ImageFolder(root, transform=transform)
else:
dataset = datasets.ImageFolder(root)
assert args.nb_classes == nb_classes
elif args.data_set == "image_folder":
root = args.data_path if is_train else args.eval_data_path
if not args.convert_to_ffcv:
dataset = datasets.ImageFolder(root, transform=transform)
else:
dataset = datasets.ImageFolder(root)
nb_classes = args.nb_classes
assert len(dataset.class_to_idx) == nb_classes
else:
raise NotImplementedError()
print(f"Dataset type : {args.data_path}")
print("Reading from datapath", args.data_path)
print("Number of classes = %d" % nb_classes)
return dataset, nb_classes
def build_transform(is_train, args):
resize_im = args.input_size > 32
imagenet_default_mean_and_std = args.imagenet_default_mean_and_std
mean = IMAGENET_INCEPTION_MEAN if not imagenet_default_mean_and_std else IMAGENET_DEFAULT_MEAN
std = IMAGENET_INCEPTION_STD if not imagenet_default_mean_and_std else IMAGENET_DEFAULT_STD
if is_train:
transform = create_transform(
input_size=args.input_size,
is_training=True,
color_jitter=args.color_jitter,
auto_augment=args.aa,
interpolation=args.train_interpolation,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
mean=mean,
std=std,
)
if not resize_im:
transform.transforms[0] = transforms.RandomCrop(
args.input_size, padding=4)
return transform
t = []
if resize_im:
# warping (no cropping) when evaluated at 384 or larger
if args.input_size >= 384:
t.append(
transforms.Resize((args.input_size, args.input_size),
interpolation=transforms.InterpolationMode.BICUBIC),
)
print(f"Warping {args.input_size} size input images...")
else:
if args.crop_pct is None and args.input_size==224:
args.crop_pct = 224 / 256
elif args.crop_pct is None:
args.crop_pct = 1.0
size = int(args.input_size / args.crop_pct)
t.append(
transforms.Resize((size,size), interpolation=transforms.InterpolationMode.BICUBIC),
)
t.append(transforms.CenterCrop(args.input_size))
t.append(transforms.ToTensor())
t.append(transforms.Normalize(mean, std))
return transforms.Compose(t)
|