File size: 10,405 Bytes
b72fefd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
Validation Runner: Runs trained model on validation set and saves predictions
"""

import os
import json
import torch
import numpy as np
from PIL import Image
from pathlib import Path
import logging
from transformers import AutoProcessor
from stage_4 import MultiHeadSiglipClassifier, CKPT, load_task_config
import pandas as pd
from tqdm import tqdm

# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

def _is_incomplete_classification(classification: dict, task_config: dict) -> bool:
    """Check if classification contains incomplete data (empty or '...' values)"""
    if not task_config or 'tasks' not in task_config:
        return True
        
    required_tasks = [task['key'] for task in task_config['tasks']]
    
    for task_key in required_tasks:
        if task_key not in classification:
            return True
        
        value = classification[task_key]
        # Check for incomplete markers
        if not value or value == "..." or value == "" or value is None:
            return True
            
    return False

def load_trained_model(checkpoint_dir: str = './checkpoints'):
    """Load the trained model and processor"""
    checkpoint_path = Path(checkpoint_dir)
    
    # Load task configuration
    task_config_path = checkpoint_path / 'task_config.json'
    if not task_config_path.exists():
        # Fallback to root directory
        task_config_path = './task_config.json'
    
    task_config = load_task_config(str(task_config_path))
    
    # Load processor
    processor = AutoProcessor.from_pretrained(CKPT)
    
    # Load model with task config
    model = MultiHeadSiglipClassifier(task_config)
    model_state = torch.load(checkpoint_path / 'multi_head_siglip2_classifier.pth', map_location='cpu')
    model.load_state_dict(model_state)
    
    # Set to evaluation mode
    model.eval()
    
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model.to(device)
    
    logger.info(f"Model loaded successfully on device: {device}")
    return model, processor, device, task_config

def load_validation_data(data_dir: str = './data', task_config: dict = None):
    """Load validation samples from stage 2 metadata files"""
    data_path = Path(data_dir)
    
    # Load from stage 2 metadata files
    metadata_dir = data_path / 'metadata'
    if not metadata_dir.exists():
        logger.error("Metadata directory not found. Run stages 1 and 2 first.")
        return []
    
    metadata_files = list(metadata_dir.glob('meta_*_stage2.json'))
    if not metadata_files:
        logger.error("No stage 2 metadata files found. Run stage 2 first.")
        return []
    
    samples = []
    skipped_incomplete = 0
    
    for meta_file in tqdm(metadata_files, desc="Loading validation data"):
        try:
            with open(meta_file, 'r') as f:
                metadata = json.load(f)
            
            # Check if classification is complete
            if not metadata.get('stage2_complete', False):
                logger.warning(f"Skipping {meta_file} - classification not complete")
                skipped_incomplete += 1
                continue
            
            # Check if classification contains incomplete data
            classification = metadata.get('classification', {})
            if not classification or _is_incomplete_classification(classification, task_config):
                logger.warning(f"Skipping {meta_file} - incomplete classification data")
                skipped_incomplete += 1
                continue
            
            # Check if image exists
            image_path = metadata['image_path']
            if not os.path.exists(image_path):
                logger.warning(f"Image not found: {image_path}")
                skipped_incomplete += 1
                continue
            
            samples.append({
                'idx': metadata['idx'],
                'image_path': metadata['image_path'],
                'caption': metadata['caption'],
                'url': metadata['url'],
                'hash': metadata['hash'],
                'ground_truth': metadata['classification']
            })
            
        except Exception as e:
            logger.warning(f"Error loading {meta_file}: {e}")
            skipped_incomplete += 1
    
    if skipped_incomplete > 0:
        logger.warning(f"Skipped {skipped_incomplete} incomplete samples")
    logger.info(f"Loaded {len(samples)} valid validation samples")
    return samples

def predict_batch(model, processor, images, device, task_config, batch_size=8):
    """Run predictions on a batch of images"""
    predictions = []
    tasks = {task['key']: task for task in task_config['tasks']}
    
    for i in range(0, len(images), batch_size):
        batch_images = images[i:i+batch_size]
        
        # Process images
        inputs = processor(images=batch_images, return_tensors="pt")
        pixel_values = inputs['pixel_values'].to(device)
        
        with torch.no_grad():
            outputs = model(pixel_values)
            
            # Convert outputs to probabilities and predictions
            batch_preds = []
            for j in range(len(batch_images)):
                pred = {}
                
                # Process each task dynamically
                for task_key, task_info in tasks.items():
                    logits = outputs[task_key][j]
                    probs = torch.softmax(logits, dim=0)
                    pred_class = torch.argmax(logits).item()
                    confidence = probs[pred_class].item()
                    
                    if task_info['type'] == 'binary':
                        # Binary classification
                        pred[f'{task_key}_prediction'] = 'yes' if pred_class == 1 else 'no'
                        pred[f'{task_key}_confidence'] = confidence
                        pred[f'{task_key}_prob_yes'] = probs[1].item()
                        pred[f'{task_key}_prob_no'] = probs[0].item()
                        
                    elif task_info['type'] == 'multi_class':
                        # Multi-class classification
                        pred_label = task_info['labels'][pred_class]
                        pred[f'{task_key}_prediction'] = pred_label
                        pred[f'{task_key}_confidence'] = confidence
                        
                        # Add probabilities for all classes
                        for idx, label in enumerate(task_info['labels']):
                            pred[f'{task_key}_prob_{label}'] = probs[idx].item()
                
                batch_preds.append(pred)
            
            predictions.extend(batch_preds)
    
    return predictions

def calculate_accuracies(predictions, ground_truths, task_config):
    """Calculate accuracies for each task"""
    accuracies = {}
    tasks = {task['key']: task for task in task_config['tasks']}
    
    for task_key, task_info in tasks.items():
        pred_key = f'{task_key}_prediction'
        
        correct = sum(1 for pred, gt in zip(predictions, ground_truths) 
                     if pred[pred_key] == gt[task_key])
        total = len(predictions)
        accuracies[f'{task_key}_accuracy'] = correct / total if total > 0 else 0
    
    return accuracies

def run_validation(data_dir: str = './data', checkpoint_dir: str = './checkpoints', 
                   output_file: str = './validation_results.json'):
    """Run complete validation and save results"""
    logger.info("Starting validation run...")
    
    # Load model and data
    model, processor, device, task_config = load_trained_model(checkpoint_dir)
    samples = load_validation_data(data_dir, task_config)
    
    if not samples:
        logger.error("No validation samples found!")
        return
    
    # Prepare images for batch processing
    images = []
    for sample in tqdm(samples, desc="Loading images"):
        try:
            img = Image.open(sample['image_path']).convert('RGB')
            images.append(img)
        except Exception as e:
            logger.error(f"Error loading image {sample['image_path']}: {e}")
            images.append(None)
    
    # Filter out failed images
    valid_samples = []
    valid_images = []
    for sample, img in zip(samples, images):
        if img is not None:
            valid_samples.append(sample)
            valid_images.append(img)
    
    logger.info(f"Running predictions on {len(valid_samples)} valid samples...")
    
    # Run predictions
    predictions = predict_batch(model, processor, valid_images, device, task_config)
    
    # Calculate accuracies
    ground_truths = [sample['ground_truth'] for sample in valid_samples]
    accuracies = calculate_accuracies(predictions, ground_truths, task_config)
    
    # Combine results
    validation_results = []
    for sample, prediction in zip(valid_samples, predictions):
        result = {
            **sample,
            'predictions': prediction
        }
        validation_results.append(result)
    
    # Create final output
    output_data = {
        'metadata': {
            'total_samples': len(validation_results),
            'model_checkpoint': checkpoint_dir,
            'validation_accuracies': accuracies,
            'task_config': task_config,
            'timestamp': pd.Timestamp.now().isoformat()
        },
        'results': validation_results
    }
    
    # Save results
    output_path = Path(output_file)
    with open(output_path, 'w') as f:
        json.dump(output_data, f, indent=2)
    
    logger.info(f"Validation results saved to {output_path}")
    logger.info("Validation Accuracies:")
    for key, value in accuracies.items():
        logger.info(f"  {key}: {value:.4f}")
    
    return output_data

def main():
    """Main execution"""
    logger.info("Starting validation runner...")
    
    # Check if model exists
    if not Path('./checkpoints/multi_head_siglip2_classifier.pth').exists():
        logger.error("Trained model not found! Run stage 4 first.")
        return
    
    # Run validation
    results = run_validation()
    
    if results:
        logger.info("Validation completed successfully!")
    else:
        logger.error("Validation failed!")

if __name__ == "__main__":
    main()