File size: 8,933 Bytes
8065faa | 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 | import argparse
import os
import random
import numpy as np
import torch
import cv2
import tifffile
import matplotlib.pyplot as plt
from typing import List, Tuple
from torch.utils.data import Dataset
from torchvision.transforms import v2
from torchvision import tv_tensors
import sys
# Allow running as `python scripts/<name>.py` from anywhere.
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from affmae.data.transforms import ElasticTransform
IMAGE_MEAN = [0.0]
IMAGE_STD = [1.0]
THRESHOLD = 0.5
class EMDatasetMultiClass(Dataset):
def __init__(
self,
base_path: str,
test_dataset: bool = False,
img_size: int = 512,
apply_transforms: bool = False,
indices: List[int] = [0, 1, 2],
input_ext: str = ".tif",
return_path: bool = True,
):
self.test_dataset = test_dataset
self.img_size = img_size
self.apply_transforms = apply_transforms
self.indices = indices
self.return_path = return_path
# Logic to determine folders
train_place_in = os.path.join(base_path, "train")
test_place_in = os.path.join(base_path, "test")
if test_dataset:
input_folder = os.path.join(test_place_in, "images")
output_folder = os.path.join(test_place_in, "masks")
else:
input_folder = os.path.join(train_place_in, "images")
output_folder = os.path.join(train_place_in, "masks")
if not os.path.exists(input_folder):
raise FileNotFoundError(f"Input folder not found: {input_folder}")
self.image_paths, self.mask_paths = self._load_image_names(input_folder, output_folder, input_ext)
self.img_mean = torch.tensor(IMAGE_MEAN, dtype=torch.float)
self.img_std = torch.tensor(IMAGE_STD, dtype=torch.float)
self._setup_transforms()
@staticmethod
def _load_image_names(input_folder: str, output_folder: str, input_ext: str = ".tif") -> Tuple[np.ndarray, np.ndarray]:
all_img = np.array(os.listdir(input_folder))
all_tiff = np.array(os.listdir(output_folder))
split_func = np.vectorize(lambda x: x.split(".")[0])
# intersection of filenames
img_paths = np.intersect1d(split_func(all_img), split_func(all_tiff))
append_input = np.vectorize(lambda x: os.path.join(input_folder, x + input_ext))
input_imgs = append_input(img_paths)
append_output = np.vectorize(lambda x: os.path.join(output_folder, x + ".tiff"))
output_imgs = append_output(img_paths)
return input_imgs, output_imgs
def _setup_transforms(self):
base_transforms = [
v2.ToImage(),
v2.ConvertImageDtype(torch.float32),
v2.Resize(size=(self.img_size, self.img_size), antialias=True, interpolation=v2.InterpolationMode.BILINEAR),
]
if self.test_dataset or not self.apply_transforms:
self.transforms = v2.Compose(base_transforms)
else:
augmentation_transforms = [
v2.RandomHorizontalFlip(p=0.5),
v2.RandomVerticalFlip(p=0.5),
v2.RandomAdjustSharpness(sharpness_factor=0.8, p=0.25),
v2.RandomAdjustSharpness(sharpness_factor=1.25, p=0.25),
v2.RandomAffine(
degrees=30,
translate=(0.0, 0.1),
scale=(0.9, 1.3),
shear=20.0
),
ElasticTransform(alpha=(100, 100), sigma=(10, 10), p=0.6)
]
self.transforms = v2.Compose(base_transforms + augmentation_transforms)
def __len__(self) -> int:
return len(self.image_paths)
def _load_segm_image(self, id: int) -> np.ndarray:
img = tifffile.imread(self.image_paths[id])
return img if len(img.shape) == 3 else img.reshape(img.shape + (1,))
def _load_segm_target(self, id: int) -> np.ndarray:
img = tifffile.imread(self.mask_paths[id])
if len(img.shape) == 4: img = img[:, :, :, 0]
return img if len(img.shape) == 2 else img[[self.indices] if isinstance(self.indices, int) else self.indices, :, :]
def __getitem__(self, index: int):
image_np = self._load_segm_image(index)
target_np_multi_channel = self._load_segm_target(index)
# CLAHE preprocessing
clahe = cv2.createCLAHE(clipLimit=4.25, tileGridSize=(8, 8))
try:
if image_np.dtype != np.uint8:
norm_img = cv2.normalize(image_np, None, 0, 255, cv2.NORM_MINMAX, dtype=cv2.CV_8U)
image_np = clahe.apply(norm_img)
else:
image_np = clahe.apply(image_np)
except:
if len(image_np.shape) == 3 and image_np.shape[2] == 3:
image_np = cv2.cvtColor(image_np, cv2.COLOR_BGR2GRAY)
image_np = clahe.apply(image_np)
if len(image_np.shape) == 2:
image_np = image_np.reshape(image_np.shape + (1,))
if len(target_np_multi_channel.shape) == 3:
C, H, W = target_np_multi_channel.shape
target_np_multi_class = np.zeros((H, W), dtype=np.int64)
for c in range(C):
class_id = c + 1
is_active = target_np_multi_channel[c, :, :] > THRESHOLD
target_np_multi_class[is_active] = class_id
else:
target_np_multi_class = target_np_multi_channel
image = image_np
target = tv_tensors.Mask(target_np_multi_class)
image, target = self.transforms(image, target)
image = image.float().sub_(self.img_mean).div_(self.img_std)
if self.return_path:
return image, target, (self.image_paths[index], self.mask_paths[index])
return image, target
def parse_args():
parser = argparse.ArgumentParser(
description="Render a grid of augmented training samples to eyeball the "
"augmentation pipeline.")
parser.add_argument("--base-path", required=True,
help="Dataset root (the directory holding train/test).")
parser.add_argument("--output", default="aug_viz.png",
help="Output image path.")
parser.add_argument("--img-size", type=int, default=512)
parser.add_argument("--input-ext", default=".tiff",
help="Image file extension.")
parser.add_argument("--config", default=None,
help="Optional YAML config to take base_path, img_size and "
"input_ext from instead of the flags above.")
return parser.parse_args()
def visualize_grid(base_path, output_path, img_size, input_ext):
print(f"Loading dataset from: {base_path}")
try:
dataset = EMDatasetMultiClass(
base_path=base_path,
test_dataset=False,
img_size=img_size,
apply_transforms=True,
input_ext=input_ext,
return_path=True
)
except Exception as e:
print(f"Error initializing dataset: {e}")
return
print(f"Found {len(dataset)} images.")
if len(dataset) < 25:
print("Dataset is too small for a 5x5 grid, using all images.")
indices = list(range(len(dataset)))
else:
indices = random.sample(range(len(dataset)), 25)
fig, axes = plt.subplots(5, 5, figsize=(15, 15))
fig.suptitle("5x5 Random Augmentation Grid", fontsize=16)
print("Generating grid...")
for idx, ax in zip(indices, axes.flat):
img_tensor, mask_tensor, (img_path, mask_path) = dataset[idx]
rel_path = os.path.relpath(img_path, base_path)
# Un-normalize for display
img = img_tensor.cpu().numpy()
img = np.transpose(img, (1, 2, 0))
mean = np.array(IMAGE_MEAN)
std = np.array(IMAGE_STD)
img = (img * std) + mean
img = np.clip(img, 0.0, 1.0)
if img.shape[2] == 1:
img = img.squeeze(2)
ax.imshow(img, cmap='gray')
else:
ax.imshow(img)
# Set Title with relative path
ax.set_title(rel_path, fontsize=6)
ax.axis('off')
plt.tight_layout()
out_dir = os.path.dirname(os.path.abspath(output_path))
os.makedirs(out_dir, exist_ok=True)
plt.savefig(output_path, dpi=125)
print(f"Saved visualization to: {output_path}")
def main():
args = parse_args()
base_path, img_size, input_ext = args.base_path, args.img_size, args.input_ext
if args.config:
from affmae.config import load_config
cfg = load_config(args.config)
base_path = cfg.base_path
img_size = cfg.img_size
input_ext = cfg.input_ext
visualize_grid(base_path, args.output, img_size, input_ext)
if __name__ == "__main__":
main() |