File size: 11,522 Bytes
0ab9351 | 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 |
"""
SAM2 Click Refinement RL Environment
"""
import gymnasium as gym
from gymnasium import spaces
import numpy as np
import torch
from PIL import Image
from scipy.ndimage import distance_transform_edt
def compute_dice(pred_mask, gt_mask):
pred = pred_mask.astype(bool)
gt = gt_mask.astype(bool)
intersection = (pred & gt).sum()
total = pred.sum() + gt.sum()
if total == 0:
return 1.0 if intersection == 0 else 0.0
return float(2 * intersection / (total + 1e-8))
def compute_iou(pred_mask, gt_mask):
pred = pred_mask.astype(bool)
gt = gt_mask.astype(bool)
intersection = (pred & gt).sum()
union = (pred | gt).sum()
if union == 0:
return 1.0
return float(intersection / (union + 1e-8))
def oracle_click(pred_mask, gt_mask, noise_range=3):
pred = pred_mask.astype(bool)
gt = gt_mask.astype(bool)
fn_mask = (~pred) & gt
fp_mask = pred & (~gt)
fn_dist = distance_transform_edt(fn_mask) if fn_mask.any() else np.zeros_like(pred_mask, dtype=float)
fp_dist = distance_transform_edt(fp_mask) if fp_mask.any() else np.zeros_like(pred_mask, dtype=float)
if fn_dist.max() >= fp_dist.max() and fn_mask.any():
click_pos = np.unravel_index(fn_dist.argmax(), fn_dist.shape)
label = 1
elif fp_mask.any():
click_pos = np.unravel_index(fp_dist.argmax(), fp_dist.shape)
label = 0
else:
if gt.any():
gt_dist = distance_transform_edt(gt)
click_pos = np.unravel_index(gt_dist.argmax(), gt_dist.shape)
label = 1
else:
click_pos = (pred_mask.shape[0] // 2, pred_mask.shape[1] // 2)
label = 0
if noise_range > 0:
noise = np.random.randint(-noise_range, noise_range + 1, size=2)
click_pos = (
int(np.clip(click_pos[0] + noise[0], 0, pred_mask.shape[0] - 1)),
int(np.clip(click_pos[1] + noise[1], 0, pred_mask.shape[1] - 1)),
)
return click_pos[0], click_pos[1], label
def binary_erosion_safe(mask):
from scipy.ndimage import binary_erosion
if not mask.any():
return mask
result = binary_erosion(mask, iterations=2)
if not result.any():
return mask
return result
class SAM2ClickEnv(gym.Env):
metadata = {"render_modes": ["rgb_array"]}
def __init__(self, dataset=None, sam_predictor=None, obs_size=128, grid_size=32,
max_clicks=5, click_radius=3, boundary_reward_weight=0.3,
initial_click_noise=5, use_sam=True, split="train", render_mode=None):
super().__init__()
self.dataset = dataset
self.sam_predictor = sam_predictor
self.obs_size = obs_size
self.grid_size = grid_size
self.max_clicks = max_clicks
self.click_radius = click_radius
self.boundary_reward_weight = boundary_reward_weight
self.initial_click_noise = initial_click_noise
self.use_sam = use_sam
self.render_mode = render_mode
self.observation_space = spaces.Box(low=0, high=255, shape=(obs_size, obs_size, 6), dtype=np.uint8)
self.action_space = spaces.Discrete(grid_size * grid_size * 2)
self.current_image = None
self.current_gt = None
self.current_mask = None
self.click_coords = []
self.click_labels = []
self.prev_dice = 0.0
self.n_clicks = 0
self.orig_h = 0
self.orig_w = 0
self._dataset_indices = None
self._current_idx = 0
def _load_random_sample(self):
if self._dataset_indices is None:
self._dataset_indices = np.random.permutation(len(self.dataset))
self._current_idx = 0
idx = int(self._dataset_indices[self._current_idx])
self._current_idx = (self._current_idx + 1) % len(self._dataset_indices)
if self._current_idx == 0:
self._dataset_indices = np.random.permutation(len(self.dataset))
sample = self.dataset[idx]
image = sample["image"]
mask = sample["mask"]
if isinstance(image, Image.Image):
image = image.convert("RGB")
image = np.array(image)
if isinstance(mask, Image.Image):
mask = mask.convert("L")
mask = np.array(mask)
mask = (mask > 127).astype(np.uint8)
return image, mask
def _resize_for_obs(self, img, is_mask=False):
pil_img = Image.fromarray(img)
if is_mask:
pil_img = pil_img.resize((self.obs_size, self.obs_size), Image.NEAREST)
else:
pil_img = pil_img.resize((self.obs_size, self.obs_size), Image.BILINEAR)
return np.array(pil_img)
def _make_click_heatmap(self, clicks_yx, orig_h, orig_w):
heatmap = np.zeros((self.obs_size, self.obs_size), dtype=np.uint8)
for (y, x) in clicks_yx:
obs_y = int(y * self.obs_size / orig_h)
obs_x = int(x * self.obs_size / orig_w)
obs_y = np.clip(obs_y, 0, self.obs_size - 1)
obs_x = np.clip(obs_x, 0, self.obs_size - 1)
for dy in range(-self.click_radius, self.click_radius+1):
for dx in range(-self.click_radius, self.click_radius+1):
if dy**2 + dx**2 <= self.click_radius**2:
ny, nx = obs_y + dy, obs_x + dx
if 0 <= ny < self.obs_size and 0 <= nx < self.obs_size:
heatmap[ny, nx] = 255
return heatmap
def _get_obs(self):
img_resized = self._resize_for_obs(self.current_image)
mask_resized = self._resize_for_obs((self.current_mask * 255).astype(np.uint8), is_mask=True)
fg_yx = [(y, x) for (x, y), l in zip(self.click_coords, self.click_labels) if l == 1]
bg_yx = [(y, x) for (x, y), l in zip(self.click_coords, self.click_labels) if l == 0]
fg_heatmap = self._make_click_heatmap(fg_yx, self.orig_h, self.orig_w)
bg_heatmap = self._make_click_heatmap(bg_yx, self.orig_h, self.orig_w)
obs = np.stack([
img_resized[:, :, 0], img_resized[:, :, 1], img_resized[:, :, 2],
mask_resized, fg_heatmap, bg_heatmap,
], axis=-1).astype(np.uint8)
return obs
def _run_sam(self):
if not self.use_sam or self.sam_predictor is None:
return self._simulate_mask()
coords = np.array(self.click_coords, dtype=np.float32)
labels = np.array(self.click_labels, dtype=np.int32)
device = "cuda" if torch.cuda.is_available() else "cpu"
ctx = torch.autocast(device, dtype=torch.bfloat16) if device == "cuda" else torch.inference_mode()
with torch.inference_mode(), ctx:
masks, scores, logits = self.sam_predictor.predict(
point_coords=coords, point_labels=labels,
multimask_output=(len(self.click_coords) == 1),
)
if len(masks.shape) == 3 and masks.shape[0] > 1:
best_idx = np.argmax(scores)
mask = masks[best_idx]
else:
mask = masks[0] if len(masks.shape) == 3 else masks
return mask.astype(np.uint8)
def _simulate_mask(self):
if not hasattr(self, '_noise_mask'):
noise = np.random.random(self.current_gt.shape) < 0.15
self._noise_mask = self.current_gt.copy()
from scipy.ndimage import binary_dilation, binary_erosion
if np.random.random() < 0.5:
self._noise_mask = binary_dilation(self._noise_mask, np.ones((7,7))).astype(np.uint8)
else:
self._noise_mask = binary_erosion(self._noise_mask, np.ones((5,5))).astype(np.uint8)
self._noise_mask = (self._noise_mask ^ noise.astype(np.uint8)).astype(np.uint8)
mask = self._noise_mask.copy()
for (x, y), label in zip(self.click_coords, self.click_labels):
radius = 20
y0, y1 = max(0, int(y)-radius), min(mask.shape[0], int(y)+radius)
x0, x1 = max(0, int(x)-radius), min(mask.shape[1], int(x)+radius)
mask[y0:y1, x0:x1] = self.current_gt[y0:y1, x0:x1]
return mask
def reset(self, seed=None, options=None):
super().reset(seed=seed)
self.current_image, self.current_gt = self._load_random_sample()
self.orig_h, self.orig_w = self.current_image.shape[:2]
self.click_coords = []
self.click_labels = []
self.n_clicks = 0
if hasattr(self, '_noise_mask'):
del self._noise_mask
if self.use_sam and self.sam_predictor is not None:
device = "cuda" if torch.cuda.is_available() else "cpu"
ctx = torch.autocast(device, dtype=torch.bfloat16) if device == "cuda" else torch.inference_mode()
with torch.inference_mode(), ctx:
self.sam_predictor.set_image(self.current_image)
init_y, init_x, init_label = oracle_click(
np.zeros_like(self.current_gt), self.current_gt,
noise_range=self.initial_click_noise
)
self.click_coords.append((int(init_x), int(init_y)))
self.click_labels.append(init_label)
self.current_mask = self._run_sam()
self.prev_dice = compute_dice(self.current_mask, self.current_gt)
return self._get_obs(), {"dice": self.prev_dice, "iou": compute_iou(self.current_mask, self.current_gt)}
def step(self, action):
label = action % 2
pos = action // 2
grid_y = pos // self.grid_size
grid_x = pos % self.grid_size
orig_x = int((grid_x + 0.5) * self.orig_w / self.grid_size)
orig_y = int((grid_y + 0.5) * self.orig_h / self.grid_size)
orig_x = np.clip(orig_x, 0, self.orig_w - 1)
orig_y = np.clip(orig_y, 0, self.orig_h - 1)
self.click_coords.append((orig_x, orig_y))
self.click_labels.append(1 if label == 0 else 0)
self.current_mask = self._run_sam()
new_dice = compute_dice(self.current_mask, self.current_gt)
delta_dice = new_dice - self.prev_dice
boundary_bonus = 0.0
pred = self.current_mask.astype(bool)
gt = self.current_gt.astype(bool)
error_mask = pred != gt
if error_mask.any():
error_dist = distance_transform_edt(~error_mask)
click_error_dist = error_dist[orig_y, orig_x]
if click_error_dist < 10:
boundary_bonus = 0.05 * (1.0 - click_error_dist / 10.0)
reward = delta_dice + self.boundary_reward_weight * boundary_bonus
self.prev_dice = new_dice
self.n_clicks += 1
terminated = self.n_clicks >= self.max_clicks
truncated = False
if new_dice > 0.95 and not terminated:
reward += 0.1
info = {
"dice": new_dice, "iou": compute_iou(self.current_mask, self.current_gt),
"delta_dice": delta_dice, "n_clicks": self.n_clicks + 1,
"click_pos": (orig_x, orig_y), "click_label": self.click_labels[-1],
}
return self._get_obs(), float(reward), terminated, truncated, info
def render(self):
if self.render_mode == "rgb_array":
return self._get_obs()[:, :, :3]
return None
|