File size: 11,707 Bytes
12fa055 |
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 |
"""
SAM 2 Few-Shot Learning Model
This module implements a few-shot segmentation model that combines SAM 2 with CLIP
for domain adaptation using minimal labeled examples.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
from PIL import Image
import clip
from segment_anything_2 import sam_model_registry, SamPredictor
from transformers import CLIPTextModel, CLIPTokenizer
class SAM2FewShot(nn.Module):
"""
SAM 2 Few-Shot Learning Model
Combines SAM 2 with CLIP for few-shot and zero-shot segmentation
across different domains (satellite, fashion, robotics).
"""
def __init__(
self,
sam2_checkpoint: str,
clip_model_name: str = "ViT-B/32",
device: str = "cuda",
prompt_engineering: bool = True,
visual_similarity: bool = True,
temperature: float = 0.1
):
super().__init__()
self.device = device
self.temperature = temperature
self.prompt_engineering = prompt_engineering
self.visual_similarity = visual_similarity
# Initialize SAM 2
self.sam2 = sam_model_registry["vit_h"](checkpoint=sam2_checkpoint)
self.sam2.to(device)
self.sam2_predictor = SamPredictor(self.sam2)
# Initialize CLIP for text and visual similarity
self.clip_model, self.clip_preprocess = clip.load(clip_model_name, device=device)
self.clip_model.eval()
# Domain-specific prompt templates
self.domain_prompts = {
"satellite": {
"building": ["building", "house", "structure", "rooftop"],
"road": ["road", "street", "highway", "pavement"],
"vegetation": ["vegetation", "forest", "trees", "green area"],
"water": ["water", "lake", "river", "ocean", "pond"]
},
"fashion": {
"shirt": ["shirt", "t-shirt", "blouse", "top"],
"pants": ["pants", "trousers", "jeans", "legs"],
"dress": ["dress", "gown", "outfit"],
"shoes": ["shoes", "footwear", "sneakers", "boots"]
},
"robotics": {
"robot": ["robot", "automation", "mechanical arm"],
"tool": ["tool", "wrench", "screwdriver", "equipment"],
"safety": ["safety equipment", "helmet", "vest", "protection"]
}
}
# Few-shot memory bank
self.few_shot_memory = {}
def encode_text_prompts(self, domain: str, class_names: List[str]) -> torch.Tensor:
"""Encode text prompts for given domain and classes."""
prompts = []
for class_name in class_names:
if domain in self.domain_prompts and class_name in self.domain_prompts[domain]:
prompts.extend(self.domain_prompts[domain][class_name])
else:
prompts.append(class_name)
# Add domain-specific context
if domain == "satellite":
prompts = [f"satellite image of {p}" for p in prompts]
elif domain == "fashion":
prompts = [f"fashion item {p}" for p in prompts]
elif domain == "robotics":
prompts = [f"robotics environment {p}" for p in prompts]
text_tokens = clip.tokenize(prompts).to(self.device)
with torch.no_grad():
text_features = self.clip_model.encode_text(text_tokens)
text_features = F.normalize(text_features, dim=-1)
return text_features
def encode_image(self, image: Union[torch.Tensor, np.ndarray, Image.Image]) -> torch.Tensor:
"""Encode image using CLIP."""
if isinstance(image, torch.Tensor):
if image.dim() == 4:
image = image.squeeze(0)
image = image.permute(1, 2, 0).cpu().numpy()
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
# Preprocess for CLIP
clip_image = self.clip_preprocess(image).unsqueeze(0).to(self.device)
with torch.no_grad():
image_features = self.clip_model.encode_image(clip_image)
image_features = F.normalize(image_features, dim=-1)
return image_features
def compute_similarity(
self,
image_features: torch.Tensor,
text_features: torch.Tensor
) -> torch.Tensor:
"""Compute similarity between image and text features."""
similarity = torch.matmul(image_features, text_features.T) / self.temperature
return similarity
def add_few_shot_example(
self,
domain: str,
class_name: str,
image: torch.Tensor,
mask: torch.Tensor
):
"""Add a few-shot example to the memory bank."""
if domain not in self.few_shot_memory:
self.few_shot_memory[domain] = {}
if class_name not in self.few_shot_memory[domain]:
self.few_shot_memory[domain][class_name] = []
# Encode the example
image_features = self.encode_image(image)
self.few_shot_memory[domain][class_name].append({
'image_features': image_features,
'mask': mask,
'image': image
})
def get_few_shot_similarity(
self,
query_image: torch.Tensor,
domain: str,
class_name: str
) -> torch.Tensor:
"""Compute similarity with few-shot examples."""
if domain not in self.few_shot_memory or class_name not in self.few_shot_memory[domain]:
return torch.zeros(1, device=self.device)
query_features = self.encode_image(query_image)
similarities = []
for example in self.few_shot_memory[domain][class_name]:
similarity = F.cosine_similarity(
query_features,
example['image_features'],
dim=-1
)
similarities.append(similarity)
return torch.stack(similarities).mean()
def generate_sam2_prompts(
self,
image: torch.Tensor,
domain: str,
class_names: List[str],
use_few_shot: bool = True
) -> List[Dict]:
"""Generate SAM 2 prompts based on text and few-shot similarity."""
prompts = []
# Text-based prompts
if self.prompt_engineering:
text_features = self.encode_text_prompts(domain, class_names)
image_features = self.encode_image(image)
text_similarities = self.compute_similarity(image_features, text_features)
# Generate point prompts based on text similarity
for i, class_name in enumerate(class_names):
if text_similarities[0, i] > 0.3: # Threshold for relevance
# Simple center point prompt (can be enhanced with attention maps)
h, w = image.shape[-2:]
point = [w // 2, h // 2]
prompts.append({
'type': 'point',
'data': point,
'label': 1,
'class': class_name,
'confidence': text_similarities[0, i].item()
})
# Few-shot based prompts
if use_few_shot and self.visual_similarity:
for class_name in class_names:
few_shot_sim = self.get_few_shot_similarity(image, domain, class_name)
if few_shot_sim > 0.5: # High similarity threshold
h, w = image.shape[-2:]
point = [w // 2, h // 2]
prompts.append({
'type': 'point',
'data': point,
'label': 1,
'class': class_name,
'confidence': few_shot_sim.item()
})
return prompts
def segment(
self,
image: torch.Tensor,
domain: str,
class_names: List[str],
use_few_shot: bool = True
) -> Dict[str, torch.Tensor]:
"""
Perform few-shot/zero-shot segmentation.
Args:
image: Input image tensor [C, H, W]
domain: Domain name (satellite, fashion, robotics)
class_names: List of class names to segment
use_few_shot: Whether to use few-shot examples
Returns:
Dictionary with masks for each class
"""
# Convert image for SAM 2
if isinstance(image, torch.Tensor):
image_np = image.permute(1, 2, 0).cpu().numpy()
else:
image_np = image
# Set image in SAM 2 predictor
self.sam2_predictor.set_image(image_np)
# Generate prompts
prompts = self.generate_sam2_prompts(image, domain, class_names, use_few_shot)
results = {}
for prompt in prompts:
class_name = prompt['class']
if prompt['type'] == 'point':
point = prompt['data']
label = prompt['label']
# Get SAM 2 prediction
masks, scores, logits = self.sam2_predictor.predict(
point_coords=np.array([point]),
point_labels=np.array([label]),
multimask_output=True
)
# Select best mask
best_mask_idx = np.argmax(scores)
mask = torch.from_numpy(masks[best_mask_idx]).float()
# Apply confidence threshold
if prompt['confidence'] > 0.3:
results[class_name] = mask
return results
def forward(
self,
image: torch.Tensor,
domain: str,
class_names: List[str],
use_few_shot: bool = True
) -> Dict[str, torch.Tensor]:
"""Forward pass for training."""
return self.segment(image, domain, class_names, use_few_shot)
class FewShotTrainer:
"""Trainer for few-shot segmentation."""
def __init__(self, model: SAM2FewShot, learning_rate: float = 1e-4):
self.model = model
self.optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
self.criterion = nn.BCELoss()
def train_step(
self,
support_images: List[torch.Tensor],
support_masks: List[torch.Tensor],
query_image: torch.Tensor,
query_mask: torch.Tensor,
domain: str,
class_name: str
):
"""Single training step."""
self.model.train()
# Add support examples to memory
for img, mask in zip(support_images, support_masks):
self.model.add_few_shot_example(domain, class_name, img, mask)
# Forward pass
predictions = self.model(query_image, domain, [class_name], use_few_shot=True)
if class_name in predictions:
pred_mask = predictions[class_name]
loss = self.criterion(pred_mask, query_mask)
else:
# If no prediction, use zero loss (can be improved)
loss = torch.tensor(0.0, device=self.model.device, requires_grad=True)
# Backward pass
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return loss.item() |