Spaces:
Runtime error
Runtime error
File size: 14,890 Bytes
bb3e610 |
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 |
import os
import sys
import torch
from torchvision.transforms.functional import normalize, to_pil_image
from torchvision.transforms import ToTensor, Normalize
import matplotlib.pyplot as plt
import json
from models import get_model
from utils import resize_density_map, sliding_window_predict
from PIL import Image
import numpy as np
from scipy.ndimage import gaussian_filter
from sklearn.cluster import KMeans
import datetime
from typing import Optional
from typing import Union
project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.append(project_root)
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class ClipEBC:
"""
CLIP-EBC (Efficient Boundary Counting) ์ด๋ฏธ์ง ์ฒ๋ฆฌ ํด๋์ค์
๋๋ค.
CLIP ๋ชจ๋ธ์ ์ฌ์ฉํ์ฌ ์ด๋ฏธ์ง๋ฅผ ์ฒ๋ฆฌํ๋ฉฐ, ์ฌ๋ผ์ด๋ฉ ์๋์ฐ ์์ธก ๊ธฐ๋ฅ์ ํฌํจํ
๋ค์ํ ์ค์ ์ต์
์ ์ ๊ณตํฉ๋๋ค.
Attributes:
truncation (int): ์๋ผ๋ด๊ธฐ ๋งค๊ฐ๋ณ์. ๊ธฐ๋ณธ๊ฐ 4.
reduction (int): ์ถ์ ๋น์จ. ๊ธฐ๋ณธ๊ฐ 8.
granularity (str): ์ธ๋ถํ ์์ค. ๊ธฐ๋ณธ๊ฐ "fine".
anchor_points (str): ์ต์ปค ํฌ์ธํธ ๋ฐฉ๋ฒ. ๊ธฐ๋ณธ๊ฐ "average".
model_name (str): CLIP ๋ชจ๋ธ ์ด๋ฆ. ๊ธฐ๋ณธ๊ฐ "clip_vit_b_16".
input_size (int): ์
๋ ฅ ์ด๋ฏธ์ง ํฌ๊ธฐ. ๊ธฐ๋ณธ๊ฐ 224.
window_size (int): ์ฌ๋ผ์ด๋ฉ ์๋์ฐ ํฌ๊ธฐ. ๊ธฐ๋ณธ๊ฐ 224.
stride (int): ์ฌ๋ผ์ด๋ฉ ์๋์ฐ ์ด๋ ๊ฐ๊ฒฉ. ๊ธฐ๋ณธ๊ฐ 224.
prompt_type (str): ํ๋กฌํํธ ์ ํ. ๊ธฐ๋ณธ๊ฐ "word".
dataset_name (str): ๋ฐ์ดํฐ์
์ด๋ฆ. ๊ธฐ๋ณธ๊ฐ "qnrf".
num_vpt (int): ๋น์ฃผ์ผ ํ๋กฌํํธ ํ ํฐ ์. ๊ธฐ๋ณธ๊ฐ 32.
vpt_drop (float): ๋น์ฃผ์ผ ํ๋กฌํํธ ํ ํฐ ๋๋กญ์์ ๋น์จ. ๊ธฐ๋ณธ๊ฐ 0.0.
deep_vpt (bool): ๊น์ ๋น์ฃผ์ผ ํ๋กฌํํธ ํ ํฐ ์ฌ์ฉ ์ฌ๋ถ. ๊ธฐ๋ณธ๊ฐ True.
mean (tuple): ์ ๊ทํ๋ฅผ ์ํ ํ๊ท ๊ฐ. ๊ธฐ๋ณธ๊ฐ (0.485, 0.456, 0.406).
std (tuple): ์ ๊ทํ๋ฅผ ์ํ ํ์คํธ์ฐจ๊ฐ. ๊ธฐ๋ณธ๊ฐ (0.229, 0.224, 0.225).
"""
def __init__(self,
truncation=4,
reduction=8,
granularity="fine",
anchor_points="average",
model_name="clip_vit_b_16",
input_size=224,
window_size=224,
stride=224,
prompt_type="word",
dataset_name="qnrf",
num_vpt=32,
vpt_drop=0.,
deep_vpt=True,
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
config_dir="configs"):
"""CLIPEBC ํด๋์ค๋ฅผ ์ค์ ๋งค๊ฐ๋ณ์์ ํจ๊ป ์ด๊ธฐํํฉ๋๋ค."""
self.truncation = truncation
self.reduction = reduction
self.granularity = granularity
self.anchor_points_type = anchor_points # ์๋ ์
๋ ฅ๊ฐ ์ ์ฅ
self.model_name = model_name
self.input_size = input_size
self.window_size = window_size
self.stride = stride
self.prompt_type = prompt_type
self.dataset_name = dataset_name
self.num_vpt = num_vpt
self.vpt_drop = vpt_drop
self.deep_vpt = deep_vpt
self.mean = mean
self.std = std
self.config_dir = config_dir
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.bins = None
self.anchor_points = None
self.model = None
# ์ด๊ธฐ ์ค์ ๋ก๋ ๋ฐ ๋ชจ๋ธ ์ด๊ธฐํ
self._load_config()
self._initialize_model()
def _load_config(self):
"""์ค์ ํ์ผ์ ๋ก๋ํ๊ณ bins์ anchor_points๋ฅผ ์ค์ ํฉ๋๋ค."""
config_path = os.path.join(self.config_dir, f"reduction_{self.reduction}.json")
with open(config_path, "r") as f:
config = json.load(f)[str(self.truncation)][self.dataset_name]
self.bins = config["bins"][self.granularity]
self.bins = [(float(b[0]), float(b[1])) for b in self.bins]
if self.anchor_points_type == "average":
self.anchor_points = config["anchor_points"][self.granularity]["average"]
else:
self.anchor_points = config["anchor_points"][self.granularity]["middle"]
self.anchor_points = [float(p) for p in self.anchor_points]
def _initialize_model(self):
"""CLIP ๋ชจ๋ธ์ ์ด๊ธฐํํฉ๋๋ค."""
self.model = get_model(
backbone=self.model_name,
input_size=self.input_size,
reduction=self.reduction,
bins=self.bins,
anchor_points=self.anchor_points,
prompt_type=self.prompt_type,
num_vpt=self.num_vpt,
vpt_drop=self.vpt_drop,
deep_vpt=self.deep_vpt
)
ckpt_path = "assets/CLIP_EBC_nwpu_rmse.pth"
ckpt = torch.load(ckpt_path, map_location=device)
self.model.load_state_dict(ckpt)
self.model = self.model.to(device)
self.model.eval()
def visualize_density_map(self, alpha: float = 0.5, save: bool = False,
save_path: Optional[str] = None):
"""
ํ์ฌ ์ ์ฅ๋ ์์ธก ๊ฒฐ๊ณผ๋ฅผ ์๊ฐํํฉ๋๋ค.
Args:
alpha (float): density map์ ํฌ๋ช
๋ (0~1). ๊ธฐ๋ณธ๊ฐ 0.5
save (bool): ์๊ฐํ ๊ฒฐ๊ณผ๋ฅผ ์ด๋ฏธ์ง๋ก ์ ์ฅํ ์ง ์ฌ๋ถ. ๊ธฐ๋ณธ๊ฐ False
save_path (str, optional): ์ ์ฅํ ๊ฒฝ๋ก. None์ผ ๊ฒฝ์ฐ ํ์ฌ ๋๋ ํ ๋ฆฌ์ ์๋ ์์ฑ๋ ์ด๋ฆ์ผ๋ก ์ ์ฅ.
๊ธฐ๋ณธ๊ฐ None
Returns:
Tuple[matplotlib.figure.Figure, np.ndarray]:
- density map์ด ์ค๋ฒ๋ ์ด๋ matplotlib Figure ๊ฐ์ฒด
- RGB ํ์์ ์๊ฐํ๋ ์ด๋ฏธ์ง ๋ฐฐ์ด (H, W, 3)
Raises:
ValueError: density_map ๋๋ processed_image๊ฐ None์ธ ๊ฒฝ์ฐ (predict ๋ฉ์๋๊ฐ ์คํ๋์ง ์์ ๊ฒฝ์ฐ)
"""
if self.density_map is None or self.processed_image is None:
raise ValueError("๋จผ์ predict ๋ฉ์๋๋ฅผ ์คํํ์ฌ ์์ธก์ ์ํํด์ผ ํฉ๋๋ค.")
fig, ax = plt.subplots(dpi=200, frameon=False)
ax.imshow(self.processed_image)
ax.imshow(self.density_map, cmap="jet", alpha=alpha)
ax.axis("off")
if save:
if save_path is None:
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
save_path = f"crowd_density_{timestamp}.png"
# ์ฌ๋ฐฑ ์ ๊ฑฐํ๊ณ ์ ์ฅ
plt.savefig(save_path, bbox_inches='tight', pad_inches=0, dpi=200)
print(f"Image saved to: {save_path}")
fig.canvas.draw()
image_from_plot = np.frombuffer(fig.canvas.buffer_rgba(), dtype=np.uint8)
image_from_plot = image_from_plot.reshape(fig.canvas.get_width_height()[::-1] + (4,))
image_from_plot = image_from_plot[:,:,:3] # RGB๋ก ๋ณํ
return fig , image_from_plot
def visualize_dots(self, dot_size: int = 20, sigma: float = 1, percentile: float = 97,
save: bool = False, save_path: Optional[str] = None):
"""
์์ธก๋ ๊ตฐ์ค ์์น๋ฅผ ์ ์ผ๋ก ํ์ํ์ฌ ์๊ฐํํฉ๋๋ค.
Args:
dot_size (int): ์ ์ ํฌ๊ธฐ. ๊ธฐ๋ณธ๊ฐ 20
sigma (float): Gaussian ํํฐ์ sigma ๊ฐ. ๊ธฐ๋ณธ๊ฐ 1
percentile (float): ์๊ณ๊ฐ์ผ๋ก ์ฌ์ฉํ ๋ฐฑ๋ถ์์ (0-100). ๊ธฐ๋ณธ๊ฐ 97
save (bool): ์๊ฐํ ๊ฒฐ๊ณผ๋ฅผ ์ด๋ฏธ์ง๋ก ์ ์ฅํ ์ง ์ฌ๋ถ. ๊ธฐ๋ณธ๊ฐ False
save_path (str, optional): ์ ์ฅํ ๊ฒฝ๋ก. None์ผ ๊ฒฝ์ฐ ํ์ฌ ๋๋ ํ ๋ฆฌ์ ์๋ ์์ฑ๋ ์ด๋ฆ์ผ๋ก ์ ์ฅ.
๊ธฐ๋ณธ๊ฐ None
Returns:
Tuple[matplotlib.backends.backend_agg.FigureCanvasBase, np.ndarray]:
- matplotlib figure์ canvas ๊ฐ์ฒด
- RGB ํ์์ ์๊ฐํ๋ ์ด๋ฏธ์ง ๋ฐฐ์ด (H, W, 3)
Raises:
ValueError: density_map ๋๋ processed_image๊ฐ None์ธ ๊ฒฝ์ฐ (predict ๋ฉ์๋๊ฐ ์คํ๋์ง ์์ ๊ฒฝ์ฐ)
"""
if self.density_map is None or self.processed_image is None:
raise ValueError("๋จผ์ predict ๋ฉ์๋๋ฅผ ์คํํ์ฌ ์์ธก์ ์ํํด์ผ ํฉ๋๋ค.")
adjusted_pred_count = int(round(self.count))
fig, ax = plt.subplots(dpi=200, frameon=False)
ax.imshow(self.processed_image)
filtered_density = gaussian_filter(self.density_map, sigma=sigma)
threshold = np.percentile(filtered_density, percentile)
candidate_pixels = np.column_stack(np.where(filtered_density >= threshold))
if len(candidate_pixels) > adjusted_pred_count:
kmeans = KMeans(n_clusters=adjusted_pred_count, random_state=42, n_init=10)
kmeans.fit(candidate_pixels)
head_positions = kmeans.cluster_centers_.astype(int)
else:
head_positions = candidate_pixels
y_coords, x_coords = head_positions[:, 0], head_positions[:, 1]
ax.scatter(x_coords, y_coords,
c='red',
s=dot_size,
alpha=1.0,
edgecolors='white',
linewidth=1)
ax.axis("off")
if save:
if save_path is None:
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
save_path = f"crowd_dots_{timestamp}.png"
plt.savefig(save_path, bbox_inches='tight', pad_inches=0, dpi=200)
print(f"Image saved to: {save_path}")
# Figure๋ฅผ numpy ๋ฐฐ์ด๋ก ๋ณํ
fig.canvas.draw()
image_from_plot = np.frombuffer(fig.canvas.buffer_rgba(), dtype=np.uint8)
image_from_plot = image_from_plot.reshape(fig.canvas.get_width_height()[::-1] + (4,))
image_from_plot = image_from_plot[:,:,:3] # RGB๋ก ๋ณํ
# plt.close(fig)
# return image_from_plot
return fig.canvas, image_from_plot
def _process_image(self, image: Union[str, np.ndarray]) -> torch.Tensor:
"""
์ด๋ฏธ์ง๋ฅผ ์ ์ฒ๋ฆฌํฉ๋๋ค. ์ด๋ฏธ์ง ๊ฒฝ๋ก, ๋ํ์ด ๋ฐฐ์ด, Streamlit UploadedFile ๋ชจ๋ ์ฒ๋ฆฌ ๊ฐ๋ฅํฉ๋๋ค.
Args:
image: ์
๋ ฅ ์ด๋ฏธ์ง. ๋ค์ ํ์ ์ค ํ๋์ฌ์ผ ํฉ๋๋ค:
- str: ์ด๋ฏธ์ง ํ์ผ ๊ฒฝ๋ก
- np.ndarray: (H, W, 3) ํํ์ RGB ์ด๋ฏธ์ง
- UploadedFile: Streamlit์ ์
๋ก๋๋ ํ์ผ
Returns:
torch.Tensor: ์ ์ฒ๋ฆฌ๋ ์ด๋ฏธ์ง ํ
์, shape (1, 3, H, W)
Raises:
ValueError: ์ง์ํ์ง ์๋ ์ด๋ฏธ์ง ํ์์ด ์
๋ ฅ๋ ๊ฒฝ์ฐ
Exception: ์ด๋ฏธ์ง ํ์ผ์ ์ด ์ ์๋ ๊ฒฝ์ฐ
"""
to_tensor = ToTensor()
normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
# ์๋ณธ ์ด๋ฏธ์ง ์ ์ฅ
self.original_image = image
# ์
๋ ฅ ํ์
์ ๋ฐ๋ฅธ ์ฒ๋ฆฌ
if isinstance(image, str):
# ํ์ผ ๊ฒฝ๋ก์ธ ๊ฒฝ์ฐ
with open(image, "rb") as f:
pil_image = Image.open(f).convert("RGB")
elif isinstance(image, np.ndarray):
# ๋ํ์ด ๋ฐฐ์ด์ธ ๊ฒฝ์ฐ
if image.dtype == np.uint8:
pil_image = Image.fromarray(image)
else:
# float ํ์
์ธ ๊ฒฝ์ฐ [0, 1] ๋ฒ์๋ก ๊ฐ์ ํ๊ณ ๋ณํ
pil_image = Image.fromarray((image * 255).astype(np.uint8))
else:
# Streamlit UploadedFile ๋๋ ๊ธฐํ ํ์ผ ๊ฐ์ฒด์ธ ๊ฒฝ์ฐ
try:
pil_image = Image.open(image).convert("RGB")
except Exception as e:
raise ValueError(f"์ง์ํ์ง ์๋ ์ด๋ฏธ์ง ํ์์
๋๋ค: {type(image)}") from e
# ํ
์ ๋ณํ ๋ฐ ์ ๊ทํ
tensor_image = to_tensor(pil_image)
normalized_image = normalize(tensor_image)
batched_image = normalized_image.unsqueeze(0) # (1, 3, H, W)
batched_image = batched_image.to(self.device)
return batched_image
def _post_process_image(self, image):
"""์ด๋ฏธ์ง ํ์ฒ๋ฆฌ๋ฅผ ์ํํฉ๋๋ค."""
image = normalize(image, mean=(0., 0., 0.),
std=(1. / self.std[0], 1. / self.std[1], 1. / self.std[2]))
image = normalize(image, mean=(-self.mean[0], -self.mean[1], -self.mean[2]),
std=(1., 1., 1.))
processed_image = to_pil_image(image.squeeze(0))
return processed_image
@torch.no_grad()
def predict(self, image: torch.Tensor) -> Image.Image:
"""
๋ชจ๋ธ ์ถ๋ ฅ ์ด๋ฏธ์ง์ ํ์ฒ๋ฆฌ๋ฅผ ์ํํฉ๋๋ค.
Args:
image (torch.Tensor): ํ์ฒ๋ฆฌํ ์ด๋ฏธ์ง ํ
์, shape (1, 3, H, W)
Returns:
PIL.Image.Image: ํ์ฒ๋ฆฌ๋ PIL ์ด๋ฏธ์ง
Note:
์ด๋ฏธ์ง ํ
์์ ๋ํด ์ ๊ทํ๋ฅผ ์ญ๋ณํํ๊ณ PIL ์ด๋ฏธ์ง ํ์์ผ๋ก ๋ณํํฉ๋๋ค.
self.mean๊ณผ self.std ๊ฐ์ ์ฌ์ฉํ์ฌ ์๋ณธ ์ด๋ฏธ์ง์ ์ค์ผ์ผ๋ก ๋ณต์ํฉ๋๋ค.
"""
processed_image = self._process_image(image)
image_height, image_width = processed_image.shape[-2:]
processed_image = processed_image.to(self.device)
pred_density = sliding_window_predict(self.model, processed_image,
self.window_size, self.stride)
pred_count = pred_density.sum().item()
resized_pred_density = resize_density_map(pred_density,
(image_height, image_width)).cpu()
self.processed_image = self._post_process_image(processed_image)
self.density_map = resized_pred_density.squeeze().numpy()
self.count = pred_count
return pred_count
def crowd_count(self):
"""
๊ฐ์ฅ ์ต๊ทผ ์์ธก์ ๊ตฐ์ค ์๋ฅผ ๋ฐํํฉ๋๋ค.
Returns:
float: ์์ธก๋ ๊ตฐ์ค ์
None: ์์ง ์์ธก์ด ์ํ๋์ง ์์ ๊ฒฝ์ฐ
"""
return self.count
def get_density_map(self):
"""
๊ฐ์ฅ ์ต๊ทผ ์์ธก์ ๋ฐ๋ ๋งต์ ๋ฐํํฉ๋๋ค.
Returns:
numpy.ndarray: ๋ฐ๋ ๋งต
None: ์์ง ์์ธก์ด ์ํ๋์ง ์์ ๊ฒฝ์ฐ
"""
return self.density_map
|