Spaces:
Running
Running
File size: 10,483 Bytes
891e05c | 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 | {
"cells": [
{
"cell_type": "markdown",
"id": "531bfd29",
"metadata": {},
"source": [
"In this notebook, we transform raw datasets to parquet format to enable faster loading speed during training and evaluation.\n",
"\n",
"The raw format of released datasets is as follows:\n",
"```python\n",
"# train set\n",
"/train/real/...\n",
"/train/fake/...\n",
"/train/masks/...\n",
"# valid set\n",
"/valid/real/...\n",
"/valid/fake/...\n",
"/valid/masks/...\n",
"```"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "8bd7e9d5",
"metadata": {},
"outputs": [],
"source": [
"import os\n",
"from datasets import Dataset, DatasetDict\n",
"from datasets import Features, Image, Value\n",
"from typing import List, Optional\n",
"\n",
"\n",
"def load_images_from_dir(directory: str) -> List[str]:\n",
" return [\n",
" os.path.join(directory, fname)\n",
" for fname in os.listdir(directory)\n",
" if fname.endswith((\"jpg\", \"jpeg\", \"png\", \"tif\"))\n",
" ]\n",
"\n",
"\n",
"def create_split(root_dir: str, split: str) -> Optional[Dataset]:\n",
" fake_dir = os.path.join(root_dir, split, \"fake\")\n",
" masks_dir = os.path.join(root_dir, split, \"masks\")\n",
" real_dir = os.path.join(root_dir, split, \"real\")\n",
"\n",
" if all(not os.path.isdir(p) for p in [fake_dir, masks_dir, real_dir]):\n",
" return None\n",
"\n",
" print(f\"Split: {split},\", end=\" \")\n",
" fake_images, real_images, mask_images = [], [], []\n",
" if os.path.isdir(fake_dir):\n",
" fake_images = load_images_from_dir(fake_dir)\n",
" print(f\"Fake images: {len(fake_images)}\", end=\"\")\n",
" if os.path.isdir(masks_dir):\n",
" mask_images = load_images_from_dir(masks_dir)\n",
" print(f\", Masks: {len(mask_images)}\", end=\"\")\n",
" assert len(fake_images) == len(mask_images)\n",
" if os.path.isdir(real_dir):\n",
" real_images = load_images_from_dir(real_dir)\n",
" print(f\", Real images: {len(real_images)}\", end=\"\")\n",
" print()\n",
"\n",
" return Dataset.from_dict(\n",
" {\n",
" \"path\": fake_images + real_images,\n",
" \"image\": fake_images + real_images,\n",
" \"mask\": mask_images + [None] * len(real_images),\n",
" },\n",
" features=Features(\n",
" {\"path\": Value(dtype=\"string\"), \"image\": Image(), \"mask\": Image()}\n",
" ),\n",
" )\n",
"\n",
"\n",
"def create_dataset(root_dir: str) -> DatasetDict:\n",
" return DatasetDict(\n",
" {\n",
" split: d\n",
" for split in [\"train\", \"valid\", \"test\"]\n",
" if (d := create_split(root_dir, split)) is not None\n",
" }\n",
" )\n",
"\n",
"\n",
"# replace with your own dataset path\n",
"root_dir = \"/gemini/space/lye/track1\"\n",
"save_dir = \"/gemini/space/jyc/track1\""
]
},
{
"cell_type": "markdown",
"id": "a1d6f1c7",
"metadata": {},
"source": [
"We merge `real/` and `fake/` into `images` column for simplity. A image is real if there is no corresponding mask."
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "07009f1e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Split: train, Fake images: 798831, Masks: 798831, Real images: 156100\n",
"Split: valid, Fake images: 199708, Masks: 199708, Real images: 39025\n",
"Split: test, Images: 222847\n"
]
},
{
"data": {
"text/plain": [
"DatasetDict({\n",
" train: Dataset({\n",
" features: ['path', 'image', 'mask'],\n",
" num_rows: 954931\n",
" })\n",
" valid: Dataset({\n",
" features: ['path', 'image', 'mask'],\n",
" num_rows: 238733\n",
" })\n",
" test: Dataset({\n",
" features: ['path', 'image'],\n",
" num_rows: 222847\n",
" })\n",
"})"
]
},
"execution_count": 14,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"dataset = create_dataset(root_dir)\n",
"dataset"
]
},
{
"cell_type": "markdown",
"id": "3aa7de84",
"metadata": {},
"source": [
"Then save processed datasets to parquet."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd6b20bc",
"metadata": {},
"outputs": [],
"source": [
"os.makedirs(save_dir, exist_ok=True)\n",
"for split in dataset:\n",
" dataset[split].to_parquet(os.path.join(save_dir, f\"{split}.parquet\"))\n",
" print(f\"Saved {split} split to {save_dir}/{split}.parquet\")"
]
},
{
"cell_type": "markdown",
"id": "f63933c8",
"metadata": {},
"source": [
"Load from processed datasets to do whatever you want."
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "4af7f346",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"Dataset({\n",
" features: ['path', 'image', 'mask'],\n",
" num_rows: 954931\n",
"})"
]
},
"execution_count": 10,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"import os\n",
"from datasets import load_dataset\n",
"\n",
"trainset = load_dataset(\"parquet\", data_dir=save_dir, split=\"train\")\n",
"trainset"
]
},
{
"cell_type": "markdown",
"id": "b3c84f0a",
"metadata": {},
"source": [
"Since the forged components are usually smaller in proportion compared to the real ones, this leads to class imbalance.\n",
"For optimal training performance, hyper parameters such as `pixel_forge_weight` and `cls_forge_weight` in `src.loupe.configuration_loupe.LoupeConfig` must be appropriately configured. These parameters control the weights of forged pixels and forged images.\n",
"\n",
"Once suitable parameters are found using the following code snippet, you can set them in `configs/model/cls.yaml` or `configs/model/seg.yaml`.\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "40a5ec91",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "19d416f59f20464692ee95bddefdaded",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Computing mask stats (num_proc=8): 0%| | 0/5000 [00:00<?, ? examples/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"cls_forge_weight: 0.16920000000000002\n",
"patch_forge_weight: 0.9294853830073696\n",
"pixel_forge_weight: 0.9160308902282281\n"
]
}
],
"source": [
"import numpy as np\n",
"from PIL import Image\n",
"from tqdm.notebook import tqdm\n",
"\n",
"cls_forge_weight: float # the ratio of forged images to total images.\n",
"# the ratio of forged patches to total patches across all images.\n",
"patch_forge_weight: float\n",
"# the ratio of forged pixels to total pixels across fake images.\n",
"pixel_forge_weight: float\n",
"\n",
"num_subset_samples = min(5000, len(trainset))\n",
"subset = trainset.shuffle().select(range(num_subset_samples))\n",
"image_size, patch_size = 336, 14\n",
"\n",
"\n",
"def compute_mask_stats(example):\n",
"\n",
" if example[\"mask\"] is None:\n",
" return {\n",
" \"is_forge\": 0,\n",
" \"forge_pixel_sum\": 0.0,\n",
" \"total_pixel_count\": 0,\n",
" \"forge_patch_sum\": 0.0,\n",
" }\n",
"\n",
" mask = example[\"mask\"].convert(\"L\").resize((image_size, image_size), Image.NEAREST)\n",
" mask_np = np.array(mask, dtype=np.float32)\n",
"\n",
" if mask_np.max() != mask_np.min():\n",
" mask_np = (mask_np - mask_np.min()) / (mask_np.max() - mask_np.min())\n",
" else:\n",
" mask_np[:] = 0.0\n",
"\n",
" forged_pixel_sum = mask_np.sum()\n",
" total_pixels = mask_np.size\n",
"\n",
" reshaped = mask_np.reshape(\n",
" image_size // patch_size, patch_size, image_size // patch_size, patch_size\n",
" )\n",
" patches = reshaped.transpose(0, 2, 1, 3)\n",
" forged_patch_sum = (patches != 0).sum(axis=(2, 3)) / (patch_size * patch_size)\n",
" forged_patch_sum = forged_patch_sum.sum()\n",
"\n",
" return {\n",
" \"is_forge\": 1,\n",
" \"forge_pixel_sum\": forged_pixel_sum,\n",
" \"total_pixel_count\": total_pixels,\n",
" \"forge_patch_sum\": forged_patch_sum,\n",
" }\n",
"\n",
"\n",
"processed = subset.map(compute_mask_stats, num_proc=8, desc=\"Computing mask stats\")\n",
"\n",
"num_forge_images = sum(processed[\"is_forge\"])\n",
"num_forge_pixels = sum(processed[\"forge_pixel_sum\"])\n",
"num_total_pixels = sum(processed[\"total_pixel_count\"])\n",
"num_forge_patches = sum(processed[\"forge_patch_sum\"])\n",
"num_total_patches = len(processed) * (image_size // patch_size) ** 2\n",
"\n",
"cls_forge_weight = 1 - num_forge_images / len(processed)\n",
"patch_forge_weight = 1 - num_forge_patches / num_total_patches\n",
"pixel_forge_weight = 1 - num_forge_pixels / num_total_pixels\n",
"\n",
"print(\"cls_forge_weight:\", cls_forge_weight)\n",
"print(\"patch_forge_weight:\", patch_forge_weight)\n",
"print(\"pixel_forge_weight:\", pixel_forge_weight)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "loupe2",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.9"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|