File size: 9,563 Bytes
3fd9d26
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
DAM Model Classes for Demo
Simplified versions of DAM inference for Hugging Face Space deployment
"""

import os
import torch
import time
from PIL import Image
from collections import defaultdict
from typing import Dict, Tuple, Optional
from transformers import AutoModel

# Simplified utility functions
def resize_keep_aspect(img: Image.Image, max_size: int = 1024) -> Image.Image:
    """Resize image while keeping aspect ratio."""
    W, H = img.size
    if max(W, H) <= max_size:
        return img
    
    if W > H:
        new_W, new_H = max_size, int(H * max_size / W)
    else:
        new_W, new_H = int(W * max_size / H), max_size
    
    return img.resize((new_W, new_H), Image.LANCZOS)

def create_full_image_mask(width: int, height: int) -> Image.Image:
    """Create a full white mask for the entire image."""
    return Image.new("L", (width, height), 255)

def get_windows(width: int, height: int, window_size: int, stride: int):
    """Generate sliding window coordinates."""
    windows = []
    for y in range(0, height - window_size + 1, stride):
        for x in range(0, width - window_size + 1, stride):
            windows.append((x, y, min(x + window_size, width), min(y + window_size, height)))
    
    # Add remaining edge windows
    if width % stride != 0:
        for y in range(0, height - window_size + 1, stride):
            windows.append((width - window_size, y, width, min(y + window_size, height)))
    
    if height % stride != 0:
        for x in range(0, width - window_size + 1, stride):
            windows.append((x, height - window_size, min(x + window_size, width), height))
    
    return windows

def aggregate_votes(votes: Dict[str, float]) -> str:
    """Aggregate votes and return the answer with highest weight."""
    if not votes:
        return ""
    return max(votes.items(), key=lambda x: x[1])[0]

class DAMOriginal:
    """Original DAM model using full image."""
    
    def __init__(self, device: str = "auto"):
        if device == "auto":
            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        else:
            self.device = torch.device(device)
        
        print(f"Loading DAM model on {self.device}...")
        self.dam_model = AutoModel.from_pretrained(
            "nvidia/DAM-3B-Self-Contained",
            trust_remote_code=True,
        ).to(self.device)
        
        self.dam = self.dam_model.init_dam(conv_mode="v1", prompt_mode="full+focal_crop")
        print("DAM Original model loaded successfully!")
    
    def predict(self, img: Image.Image, question: str, max_new_tokens: int = 100) -> Tuple[str, float]:
        """
        Generate prediction for the question using full image.
        
        Returns:
            Tuple of (answer, inference_time)
        """
        # Resize image
        img = resize_keep_aspect(img, 1024)
        W, H = img.size
        
        # Create full image mask
        mask = create_full_image_mask(W, H)
        
        # Format prompt
        prompt = (
            "<image>\n"
            "Answer each question concisely in a single word or short phrase, "
            "without any lengthy descriptions or explanations.\n"
            "Rely only on information that is clearly visible in the provided image.\n"
            "If the answer cannot be determined from the image, respond with \"unanswerable\".\n"
            f"Question: {question}\nAnswer:"
        )
        
        # Inference parameters
        params = {
            "streaming": False,
            "temperature": 1e-7,
            "top_p": 0.5,
            "num_beams": 1,
            "max_new_tokens": max_new_tokens
        }
        
        start_time = time.time()
        try:
            tokens = self.dam.get_description(img, mask, prompt, **params)
            inference_time = time.time() - start_time
            
            if isinstance(tokens, str):
                answer = tokens.strip()
            else:
                answer = "".join(tokens).strip()
                
            return answer, inference_time
            
        except Exception as e:
            inference_time = time.time() - start_time
            print(f"Error in DAM Original prediction: {e}")
            return f"Error: {str(e)}", inference_time


class DAMSlidingWindow:
    """DAM model with sliding window approach."""
    
    def __init__(self, device: str = "auto", window_size: int = 512, stride: int = 256):
        if device == "auto":
            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        else:
            self.device = torch.device(device)
        
        self.window_size = window_size
        self.stride = stride
        
        print(f"Loading DAM model on {self.device}...")
        self.dam_model = AutoModel.from_pretrained(
            "nvidia/DAM-3B-Self-Contained",
            trust_remote_code=True,
        ).to(self.device)
        
        self.dam = self.dam_model.init_dam(conv_mode="v1", prompt_mode="full+focal_crop")
        print(f"DAM Sliding Window model loaded successfully! (window_size={window_size}, stride={stride})")
    
    def predict(self, img: Image.Image, question: str, max_new_tokens: int = 100, 
                unanswerable_weight: float = 1.0) -> Tuple[str, float, Dict]:
        """
        Generate prediction using sliding window approach with voting.
        
        Returns:
            Tuple of (answer, inference_time, voting_details)
        """
        # Resize image
        img = resize_keep_aspect(img, 1024)
        W, H = img.size
        
        # Format prompt
        prompt = (
            "<image>\n"
            "Answer each question concisely in a single word or short phrase, "
            "without any lengthy descriptions or explanations.\n"
            "Rely only on information that is clearly visible in the provided image.\n"
            "If the answer cannot be determined from the image, respond with \"unanswerable\".\n"
            f"Question: {question}\nAnswer:"
        )
        
        # Inference parameters
        params = {
            "streaming": False,
            "temperature": 1e-7,
            "top_p": 0.5,
            "num_beams": 1,
            "max_new_tokens": max_new_tokens
        }
        
        start_time = time.time()
        votes = defaultdict(float)
        voting_details = {"full_image": None, "windows": []}
        
        try:
            # Full image vote
            mask_full = create_full_image_mask(W, H)
            ans_full = self.dam.get_description(img, mask_full, prompt, **params)
            
            if isinstance(ans_full, str):
                ans_full = ans_full.strip()
            else:
                ans_full = "".join(ans_full).strip()
            
            if ans_full:
                weight = 1.0
                if ans_full.lower() == "unanswerable":
                    weight *= unanswerable_weight
                votes[ans_full] += weight
                voting_details["full_image"] = {"answer": ans_full, "weight": weight}
            
            # Sliding window votes
            windows = get_windows(W, H, self.window_size, self.stride)
            for i, (x0, y0, x1, y1) in enumerate(windows):
                crop = img.crop((x0, y0, x1, y1))
                mask_crop = Image.new("L", (x1-x0, y1-y0), 255)
                
                ans = self.dam.get_description(crop, mask_crop, prompt, **params)
                
                if isinstance(ans, str):
                    ans = ans.strip()
                else:
                    ans = "".join(ans).strip()
                
                if ans:
                    weight = ((x1-x0) * (y1-y0)) / (W * H)
                    if ans.lower() == "unanswerable":
                        weight *= unanswerable_weight
                    votes[ans] += weight
                    
                    voting_details["windows"].append({
                        "window_id": i,
                        "coords": (x0, y0, x1, y1),
                        "answer": ans,
                        "weight": weight
                    })
            
            # Aggregate votes
            prediction = aggregate_votes(votes)
            if not prediction:
                prediction = ans_full if 'ans_full' in locals() else "No answer"
            
            inference_time = time.time() - start_time
            
            # Add vote summary to details
            voting_details["vote_summary"] = dict(votes)
            voting_details["final_answer"] = prediction
            voting_details["total_windows"] = len(windows)
            
            return prediction, inference_time, voting_details
            
        except Exception as e:
            inference_time = time.time() - start_time
            print(f"Error in DAM Sliding Window prediction: {e}")
            return f"Error: {str(e)}", inference_time, {"error": str(e)}


# Global model instances (lazy loading)
_dam_original = None
_dam_sliding = None

def get_dam_original(device: str = "auto"):
    """Get or create DAM Original model instance."""
    global _dam_original
    if _dam_original is None:
        _dam_original = DAMOriginal(device)
    return _dam_original

def get_dam_sliding(device: str = "auto", window_size: int = 512, stride: int = 256):
    """Get or create DAM Sliding Window model instance."""
    global _dam_sliding
    if _dam_sliding is None:
        _dam_sliding = DAMSlidingWindow(device, window_size, stride)
    return _dam_sliding