File size: 7,488 Bytes
f74cf62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f575dc
f74cf62
 
 
 
 
6f575dc
 
 
 
 
f74cf62
6f575dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f74cf62
 
 
6f575dc
f74cf62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35351a6
 
 
 
f74cf62
 
35351a6
 
 
 
 
f74cf62
 
 
 
 
35351a6
f74cf62
35351a6
f74cf62
35351a6
 
f74cf62
 
 
35351a6
 
 
 
 
 
 
 
 
 
 
 
 
f74cf62
 
 
35351a6
f74cf62
 
 
35351a6
 
 
 
f74cf62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Secure Model Server - Protects model weights from extraction
Never expose:
- File paths to checkpoints
- Model architecture details
- Debug routes
"""

import os
import sys
import torch
import numpy as np
from pathlib import Path
from typing import Tuple, Optional

# Secure path resolution (not hardcoded)
def get_model_checkpoint_path():
    """Get checkpoint path secretly, never expose to client"""
    base_dir = Path(__file__).parent
    checkpoint = base_dir / "segment-anything-2" / "checkpoints" / "sam2.1_hiera_small.pt"
    if not checkpoint.exists():
        raise FileNotFoundError(f"Model checkpoint not found")
    return str(checkpoint)

def get_finetuned_weights_path():
    """Get fine-tuned weights path secretly, never expose to client
    
    Attempts to download from Hugging Face if local copy doesn't exist
    and HF_TOKEN is available.
    """
    base_dir = Path(__file__).parent
    checkpoint_dir = base_dir / "segment-anything-2" / "checkpoints"
    checkpoint_dir.mkdir(parents=True, exist_ok=True)
    weights = checkpoint_dir / "VREyeSAM_uncertainity_best.torch"
    
    # If weights already exist locally, return path
    if weights.exists():
        return str(weights)
    
    # Try to download from Hugging Face using HF_TOKEN
    hf_token = os.getenv('HF_TOKEN', '')
    if hf_token:
        try:
            from huggingface_hub import hf_hub_download
            print("Downloading VREyeSAM weights from Hugging Face...")
            
            checkpoint_path = hf_hub_download(
                repo_id='devnagaich/VREyeSAM',
                filename='VREyeSAM_uncertainity_best.torch',
                token=hf_token,
                cache_dir=str(checkpoint_dir)
            )
            print(f"Successfully downloaded VREyeSAM weights")
            return checkpoint_path
        except Exception as e:
            print(f"Warning: Could not download VREyeSAM weights: {e}")
    
    # If download fails or no token, return path anyway (may exist from upload)
    if weights.exists():
        return str(weights)
    
    # Last resort - raise error
    raise FileNotFoundError(f"VREyeSAM weights not found and could not download")

def get_model_config_path():
    """Get model config path secretly, never expose to client"""
    return "configs/sam2.1/sam2.1_hiera_s.yaml"


class ProtectedModelServer:
    """
    Encapsulates model loading and inference
    Only exposes inference API, never raw weights or paths
    """
    
    _instance = None  # Singleton pattern
    _model = None
    _predictor = None
    
    def __new__(cls):
        # Singleton: only one instance ever
        if cls._instance is None:
            cls._instance = super().__new__(cls)
        return cls._instance
    
    def __init__(self):
        """Initialize model (only once)"""
        if self._predictor is None:
            self._load_model()
    
    def _load_model(self):
        """Load model weights securely - never called from frontend"""
        try:
            # Add segment-anything-2 to path (internally only)
            base_dir = Path(__file__).parent
            sam2_path = base_dir / "segment-anything-2"
            
            if not sam2_path.exists():
                raise FileNotFoundError(f"SAM2 installation not found at {sam2_path}")
            
            sys.path.insert(0, str(sam2_path))
            
            try:
                from sam2.build_sam import build_sam2
                from sam2.sam2_image_predictor import SAM2ImagePredictor
            except ImportError as e:
                raise ImportError("SAM2 not properly installed. Check build logs.") from e
            
            # Get paths internally - NEVER sent to client
            model_cfg = get_model_config_path()
            sam2_checkpoint = get_model_checkpoint_path()
            
            # Load device
            device = "cuda" if torch.cuda.is_available() else "cpu"
            print(f"Loading model on device: {device}")
            
            # Load base SAM2 model
            print(f"Loading SAM2 from {sam2_checkpoint}")
            self._model = build_sam2(model_cfg, sam2_checkpoint, device=device)
            self._predictor = SAM2ImagePredictor(self._model)
            
            # Try to load fine-tuned weights if available
            try:
                fine_tuned_weights = get_finetuned_weights_path()
                print(f"Loading fine-tuned weights from {fine_tuned_weights}")
                state_dict = torch.load(fine_tuned_weights, map_location=device)
                self._predictor.model.load_state_dict(state_dict)
                print("Fine-tuned weights loaded successfully")
            except FileNotFoundError:
                print("Warning: Fine-tuned weights not found. Using base SAM2 model.")
                print("To use fine-tuned model, upload VREyeSAM_uncertainity_best.torch to Space Files")
            except Exception as e:
                print(f"Warning: Could not load fine-tuned weights: {e}")
                print("Continuing with base SAM2 model")
            
            # Model is now loaded - weights are NOT accessible to clients
            self._predictor.model.eval()
            print("Model loaded successfully")
            
            return True
        except Exception as e:
            print(f"Error loading model: {e}")
            import traceback
            traceback.print_exc()
            raise RuntimeError(f"Model initialization failed: {str(e)}") from e
    
    def predict(self, image: np.ndarray, num_samples: int = 30) -> Tuple[np.ndarray, np.ndarray]:
        """
        Perform iris segmentation
        
        Args:
            image: Input image (numpy array)
            num_samples: Number of random points for inference
            
        Returns:
            binary_mask: Binary segmentation mask
            prob_mask: Probability map
        """
        if self._predictor is None:
            raise RuntimeError("Model not initialized")
        
        try:
            # Generate random points for inference
            input_points = np.random.randint(0, min(image.shape[:2]), (num_samples, 1, 2))
            
            # Inference
            with torch.no_grad():
                self._predictor.set_image(image)
                masks, scores, _ = self._predictor.predict(
                    point_coords=input_points,
                    point_labels=np.ones([input_points.shape[0], 1])
                )
            
            # Convert to numpy
            np_masks = np.array(masks[:, 0]).astype(np.float32)
            np_scores = scores[:, 0]
            
            # Normalize scores
            score_sum = np.sum(np_scores)
            if score_sum > 0:
                normalized_scores = np_scores / score_sum
            else:
                normalized_scores = np.ones_like(np_scores) / len(np_scores)
            
            # Generate probabilistic mask
            prob_mask = np.sum(np_masks * normalized_scores[:, None, None], axis=0)
            prob_mask = np.clip(prob_mask, 0, 1)
            
            # Threshold to get binary mask
            binary_mask = (prob_mask > 0.2).astype(np.uint8)
            
            return binary_mask, prob_mask
        
        except Exception as e:
            raise RuntimeError(f"Inference failed") from e


def get_predictor() -> ProtectedModelServer:
    """Get singleton model instance"""
    return ProtectedModelServer()