File size: 31,622 Bytes
420f791
 
fa937f9
 
 
 
420f791
fa937f9
 
 
 
 
 
 
 
 
420f791
fa937f9
 
 
 
 
 
 
 
 
 
420f791
fa937f9
 
 
226318e
fa937f9
 
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45d4825
 
fa937f9
 
 
420f791
 
 
 
 
 
 
 
 
 
fa937f9
420f791
 
fa937f9
420f791
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
fa937f9
420f791
 
 
 
 
 
fa937f9
 
420f791
fa937f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
420f791
 
 
fa937f9
 
 
 
420f791
 
fa937f9
 
 
 
420f791
 
 
 
fa937f9
 
 
420f791
 
 
 
fa937f9
420f791
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
fa937f9
420f791
fa937f9
420f791
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
fa937f9
 
420f791
 
fa937f9
 
45d4825
420f791
fa937f9
420f791
 
 
fa937f9
f9d3bad
420f791
 
 
 
 
fa937f9
420f791
 
fa937f9
 
420f791
fa937f9
 
 
45d4825
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
45d4825
fa937f9
420f791
 
 
fa937f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
420f791
fa937f9
 
 
420f791
 
 
 
 
 
fa937f9
420f791
 
 
fa937f9
420f791
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
 
 
fa937f9
 
 
 
 
420f791
fa937f9
 
 
 
420f791
 
 
 
 
 
fa937f9
 
420f791
fa937f9
420f791
 
 
 
 
 
 
 
 
fa937f9
3529ca1
fa937f9
420f791
 
 
 
 
fa937f9
 
420f791
 
 
 
 
 
 
 
 
 
 
 
fa937f9
3529ca1
420f791
 
fa937f9
 
420f791
fa937f9
420f791
fa937f9
 
420f791
fa937f9
420f791
fa937f9
420f791
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
fa937f9
 
420f791
fa937f9
 
420f791
fa937f9
 
 
 
 
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
 
64114f8
420f791
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
 
420f791
 
fa937f9
420f791
fa937f9
 
 
 
420f791
 
 
 
 
 
 
 
 
 
 
 
fa937f9
 
 
 
 
 
 
420f791
 
 
fa937f9
 
 
420f791
fa937f9
 
 
 
420f791
 
 
 
 
 
fa937f9
420f791
 
 
fa937f9
420f791
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
 
420f791
fa937f9
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa937f9
420f791
fa937f9
420f791
 
fa937f9
420f791
fa937f9
 
 
 
 
 
 
 
 
 
420f791
fa937f9
 
 
 
 
 
420f791
 
 
fa937f9
 
 
420f791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
"""Complete generative inference module with model loading and inference capabilities."""

import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as transforms
import torchvision.models as models
from torchvision.models.resnet import ResNet50_Weights
from PIL import Image
import numpy as np
import os
import requests
import time
import copy
from collections import OrderedDict
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Union

# Check for available hardware acceleration
if torch.cuda.is_available():
    device = torch.device("cuda")
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
    device = torch.device("mps")  # Use Apple Metal Performance Shaders for M-series Macs
else:
    device = torch.device("cpu")
print(f"Using device: {device}")

# Constants for model URLs
MODEL_URLS = {
    'resnet50_robust': 'https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps3.ckpt',
    'resnet50_standard': 'https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.ckpt',
    'resnet50_robust_face': 'https://huggingface.co/ttoosi/resnet50_robust_face/resolve/main/resnet50_vggface2_L2_eps_0.50_checkpoint150.pt'
}

# Model-specific preprocessing configurations
MODEL_CONFIGS = {
    'resnet50_robust_face': {
        'input_size': 112,
        'norm_mean': [0.5, 0.5, 0.5],
        'norm_std': [0.5, 0.5, 0.5],
        'n_classes': 500,
        'dataset': 'VGGFace2'
    },
    'resnet50_standard': {
        'input_size': 224,
        'norm_mean': [0.485, 0.456, 0.406],
        'norm_std': [0.229, 0.224, 0.225],
        'n_classes': 1000,
        'dataset': 'ImageNet'
    },
    'resnet50_robust': {
        'input_size': 224,
        'norm_mean': [0.485, 0.456, 0.406],
        'norm_std': [0.229, 0.224, 0.225],
        'n_classes': 1000,
        'dataset': 'ImageNet'
    }
}

IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD = [0.229, 0.224, 0.225]

def get_iterations_to_show(n_itr):
    """Generate a dynamic list of iterations to show based on total iterations."""
    if n_itr <= 50:
        return [1, 5, 10, 20, 30, 40, 50, n_itr]
    elif n_itr <= 100:
        return [1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, n_itr]
    elif n_itr <= 200:
        return [1, 5, 10, 20, 30, 40, 50, 75, 100, 125, 150, 175, 200, n_itr]
    elif n_itr <= 500:
        return [1, 5, 10, 20, 30, 40, 50, 75, 100, 150, 200, 250, 300, 350, 400, 450, 500, n_itr]
    else:
        return [1, 5, 10, 20, 30, 40, 50, 75, 100, 150, 200, 250, 300, 350, 400, 450, 500, 
                int(n_itr*0.6), int(n_itr*0.7), int(n_itr*0.8), int(n_itr*0.9), n_itr]

def get_inference_configs(inference_type='IncreaseConfidence', eps=0.5, n_itr=50, step_size=1.0):
    """Generate inference configuration with customizable parameters."""
    
    config = {
        'loss_infer': inference_type,
        'n_itr': n_itr,
        'eps': eps,
        'step_size': step_size,
        'diffusion_noise_ratio': 0.0,
        'initial_inference_noise_ratio': 0.0,
        'top_layer': 'all',
        'inference_normalization': False,
        'recognition_normalization': False,
        'iterations_to_show': get_iterations_to_show(n_itr),
        'misc_info': {'keep_grads': False}
    }
    
    if inference_type == 'IncreaseConfidence':
        config['loss_function'] = 'CE'
    elif inference_type == 'Prior-Guided Drift Diffusion':
        config['loss_function'] = 'MSE'
        config['initial_inference_noise_ratio'] = 0.05
        config['diffusion_noise_ratio'] = 0.01
        config['top_layer'] = 'layer4'
    elif inference_type == 'GradModulation':
        config['loss_function'] = 'CE'
        config['misc_info']['grad_modulation'] = 0.5
    elif inference_type == 'CompositionalFusion':
        config['loss_function'] = 'CE'
        config['misc_info']['positive_classes'] = []
        config['misc_info']['negative_classes'] = []
    
    return config

def get_model_preprocessing(model_type: str) -> Dict:
    """Get preprocessing configuration for specific model type."""
    if model_type not in MODEL_CONFIGS:
        print(f"Fall-back: Unknown model type {model_type}, using ImageNet defaults")
        return MODEL_CONFIGS['resnet50_standard']
    return MODEL_CONFIGS[model_type]

class NormalizeByChannelMeanStd(nn.Module):
    """Normalization layer for models."""
    def __init__(self, mean, std):
        super(NormalizeByChannelMeanStd, self).__init__()
        if not isinstance(mean, torch.Tensor):
            mean = torch.tensor(mean)
        if not isinstance(std, torch.Tensor):
            std = torch.tensor(std)
        self.register_buffer("mean", mean)
        self.register_buffer("std", std)
        
    def forward(self, tensor):
        return self.normalize_fn(tensor, self.mean, self.std)
    
    def normalize_fn(self, tensor, mean, std):
        """Differentiable version of torchvision.functional.normalize"""
        mean = mean[None, :, None, None]
        std = std[None, :, None, None]
        return tensor.sub(mean).div(std)

class InferStep:
    """Inference step class for gradient-based optimization."""
    
    def __init__(self, orig_image: torch.Tensor, eps: float, step_size: float):
        self.orig_image = orig_image
        self.eps = eps
        self.step_size = step_size

    def project(self, x: torch.Tensor) -> torch.Tensor:
        """Project x onto epsilon-ball around original image."""
        diff = x - self.orig_image
        diff = torch.clamp(diff, -self.eps, self.eps)
        return torch.clamp(self.orig_image + diff, 0, 1)

    def step(self, x: torch.Tensor, grad: torch.Tensor) -> torch.Tensor:
        """Take a normalized gradient step."""
        dim = len(x.shape) - 1
        grad_norm = torch.norm(grad.reshape(grad.shape[0], -1), dim=1).reshape(-1, *([1] * dim))
        scaled_grad = grad / (grad_norm + 1e-10)
        return scaled_grad * self.step_size

def extract_middle_layers(model: nn.Module, layer_index: Union[str, int]) -> nn.Module:
    """Extract middle layers from a model up to a specified layer index."""
    if isinstance(layer_index, str) and layer_index == 'all':
        return model
    
    # Handle ResNet layer extraction
    modules = list(model.named_children())
    cutoff_idx = next(
        (i for i, (name, _) in enumerate(modules) if name == str(layer_index)),
        None
    )
    
    if cutoff_idx is not None:
        new_model = nn.Sequential(OrderedDict(modules[:cutoff_idx + 1]))
        return new_model
    else:
        print(f"Fall-back: Module {layer_index} not found, using full model")
        return model

def calculate_loss(output_model: torch.Tensor, class_indices: List[int], loss_inference: str) -> torch.Tensor:
    """Calculate loss for specified class indices."""
    losses = []
    for idx in class_indices:
        target = torch.full((1,), idx, dtype=torch.long, device=output_model.device)
        if loss_inference == 'CE':
            loss = nn.CrossEntropyLoss()(output_model, target)
        elif loss_inference == 'MSE':
            one_hot_target = torch.zeros_like(output_model)
            one_hot_target[0, target] = 1
            loss = nn.MSELoss()(output_model, one_hot_target)
        else:
            raise ValueError(f"Unsupported loss_inference: {loss_inference}")
        losses.append(loss)
    
    return torch.stack(losses).mean()

def download_model(model_type):
    """Download model if needed."""
    if model_type not in MODEL_URLS or MODEL_URLS[model_type] is None:
        return None
    
    os.makedirs("models", exist_ok=True)
    
    if model_type == 'resnet50_robust_face':
        model_path = Path("models/resnet50_vggface2_L2_eps_0.50_checkpoint150.pt")
    else:
        model_path = Path(f"models/{model_type}.pt")
    
    if not model_path.exists():
        print(f"Downloading {model_type} model...")
        url = MODEL_URLS[model_type]
        response = requests.get(url, stream=True)
        if response.status_code == 200:
            with open(model_path, 'wb') as f:
                for chunk in response.iter_content(chunk_size=8192):
                    f.write(chunk)
            print(f"Model downloaded and saved to {model_path}")
        else:
            raise RuntimeError(f"Failed to download model: {response.status_code}")
    return model_path

class GenerativeInferenceModel:
    """Complete generative inference model with model loading and inference."""
    
    def __init__(self):
        self.models = {}
        self.model_preproc = {}
        self.labels = self.get_imagenet_labels()
        
    def get_imagenet_labels(self):
        """Get ImageNet labels."""
        url = "https://raw.githubusercontent.com/anishathalye/imagenet-simple-labels/master/imagenet-simple-labels.json"
        try:
            response = requests.get(url, timeout=10)  # Add timeout to prevent hanging
            if response.status_code == 200:
                return response.json()
            else:
                print("Fall-back: Failed to fetch ImageNet labels, using placeholder")
                return [f"class_{i}" for i in range(1000)]
        except Exception as e:
            print(f"Fall-back: Error fetching labels: {e}")
            return [f"class_{i}" for i in range(1000)]
    
    def load_model(self, model_type):
        """Load and cache models for different model types."""
        if model_type in self.models:
            print(f"Using cached {model_type} model")
            return self.models[model_type]
    
        start_time = time.time()
        
        # Get model-specific preprocessing config
        preproc_config = get_model_preprocessing(model_type)
        self.model_preproc[model_type] = preproc_config
        
        # Create normalizer
        normalizer = NormalizeByChannelMeanStd(
            preproc_config['norm_mean'], 
            preproc_config['norm_std']
        ).to(device)
        
        # Create base model architecture
        num_classes = preproc_config['n_classes']
        resnet = models.resnet50(num_classes=num_classes)
        model = nn.Sequential(normalizer, resnet)
        
        # Download and load checkpoint
        model_path = download_model(model_type)
        
        if model_path:
            print(f"Loading {model_type} model from {model_path}...")
            try:
                checkpoint = torch.load(model_path, map_location=device)
                
                # Handle different checkpoint formats
                if 'model' in checkpoint:
                    state_dict = checkpoint['model']
                    print("Using 'model' key from checkpoint")
                elif 'state_dict' in checkpoint:
                    state_dict = checkpoint['state_dict']
                    print("Using 'state_dict' key from checkpoint")
                else:
                    state_dict = checkpoint
                    print("Using checkpoint directly as state_dict")
                
                # Extract ResNet state dict
                resnet_state_dict = {}
                resnet_keys = set(resnet.state_dict().keys())
                
                # For face model, prioritize 'module.model.model.' structure (seen in actual checkpoint)
                if model_type == 'resnet50_robust_face':
                    # Check for 'module.model.model.' structure first (face checkpoints use this)
                    module_model_model_keys = [key for key in state_dict.keys() if key.startswith('module.model.model.')]
                    if module_model_model_keys:
                        print(f"Found 'module.model.model.' structure with {len(module_model_model_keys)} parameters (face model)")
                        for source_key, value in state_dict.items():
                            if source_key.startswith('module.model.model.'):
                                target_key = source_key[len('module.model.model.'):]
                                if target_key in resnet_keys:
                                    resnet_state_dict[target_key] = value
                        print(f"Extracted {len(resnet_state_dict)} parameters from module.model.model.")
                    
                    # Also check for 'module.model.' structure as fallback
                    if len(resnet_state_dict) < len(resnet_keys):
                        module_model_keys = [key for key in state_dict.keys() if key.startswith('module.model.') and not key.startswith('module.model.model.')]
                        if module_model_keys:
                            print(f"Found additional 'module.model.' structure with {len(module_model_keys)} parameters")
                            for source_key, value in state_dict.items():
                                if source_key.startswith('module.model.') and not source_key.startswith('module.model.model.'):
                                    target_key = source_key[len('module.model.'):]
                                    # Remove extra 'model.' if present
                                    if target_key.startswith('model.'):
                                        target_key = target_key[len('model.'):]
                                    if target_key in resnet_keys and target_key not in resnet_state_dict:
                                        resnet_state_dict[target_key] = value
                            print(f"Now have {len(resnet_state_dict)} parameters after adding module.model. keys")
                
                # Handle different key prefixes in checkpoints (for other models)
                if len(resnet_state_dict) == 0:
                    prefixes_to_try = ['', 'module.', 'model.', 'attacker.model.', 'attacker.']
                    
                    for source_key, value in state_dict.items():
                        target_key = source_key
                        
                        # Try removing various prefixes
                        for prefix in prefixes_to_try:
                            if source_key.startswith(prefix):
                                target_key = source_key[len(prefix):]
                                break
                        
                        # Handle nested model keys
                        if target_key.startswith('model.'):
                            target_key = target_key[len('model.'):]
                        
                        # If the target key is in ResNet keys, add it
                        if target_key in resnet_keys:
                            resnet_state_dict[target_key] = value
                
                # Load the state dict
                if resnet_state_dict:
                    result = resnet.load_state_dict(resnet_state_dict, strict=False)
                    missing_keys, unexpected_keys = result
                    
                    loaded_percent = (len(resnet_state_dict) / len(resnet_keys)) * 100
                    print(f"Model loading: {len(resnet_state_dict)}/{len(resnet_keys)} parameters ({loaded_percent:.1f}%)")
                    
                    if loaded_percent < 50:
                        print(f"Fall-back: Loading too incomplete ({loaded_percent:.1f}%), using PyTorch pretrained")
                        if model_type != 'resnet50_robust_face':
                            resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
                            model = nn.Sequential(normalizer, resnet)
                        
                else:
                    print("Fall-back: No matching keys found in checkpoint, using PyTorch pretrained")
                    if model_type != 'resnet50_robust_face':
                        resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
                        model = nn.Sequential(normalizer, resnet)
                        
            except Exception as e:
                print(f"Fall-back: Error loading checkpoint: {e}")
                if model_type != 'resnet50_robust_face':
                    print("Fall-back: Using PyTorch pretrained model")
                    resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
                    model = nn.Sequential(normalizer, resnet)
                else:
                    print("Fall-back: Face model checkpoint failed, model may not work properly")
                    
        else:
            # Use PyTorch's pretrained model for ImageNet models
            if model_type != 'resnet50_robust_face':
                print(f"No checkpoint for {model_type}, using PyTorch pretrained")
                resnet = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
                model = nn.Sequential(normalizer, resnet)
            else:
                print("Fall-back: Face model requires checkpoint, model may not work properly")
        
        model = model.to(device)
        model.eval()
        
        # Verify model
        self.verify_model_integrity(model, model_type)
        
        # Cache the model
        self.models[model_type] = model
        
        end_time = time.time()
        print(f"Model {model_type} loaded in {end_time - start_time:.2f} seconds")
        return model

    def verify_model_integrity(self, model, model_type):
        """Verify model integrity."""
        try:
            print(f"Fall-back: Running model integrity check for {model_type}")
            config = get_model_preprocessing(model_type)
            H = W = config['input_size']
            
            test_input = torch.zeros(1, 3, H, W, device=device)
            test_input[0, 0, H//4:3*H//4, W//4:3*W//4] = 0.5
            
            with torch.no_grad():
                output = model(test_input)
            
            expected_classes = config['n_classes']
            if output.shape != (1, expected_classes):
                print(f"Fall-back: Unexpected output shape: {output.shape}, expected (1, {expected_classes})")
                return False
                
            probs = torch.nn.functional.softmax(output, dim=1)
            confidence, prediction = torch.max(probs, 1)
            
            print(f"Model integrity check passed:")
            print(f"- Output shape: {output.shape}")
            print(f"- Top prediction: Class {prediction.item()} with {confidence.item()*100:.2f}% confidence")
            
            return True
            
        except Exception as e:
            print(f"Fall-back: Model integrity check failed with error: {e}")
            return False

    def inference(self, image, model_type, config):
        """Run generative inference."""
        inference_start = time.time()
        
        # Load the model
        model = self.load_model(model_type)
        
        # Handle image input
        if isinstance(image, str):
            if os.path.exists(image):
                image = Image.open(image).convert('RGB')
            else:
                raise ValueError(f"Image path does not exist: {image}")
        elif isinstance(image, np.ndarray):
            if image.dtype != np.uint8:
                if image.max() <= 1.0:
                    image = (image * 255).astype(np.uint8)
                else:
                    image = image.astype(np.uint8)
            if len(image.shape) == 3:
                if image.shape[0] == 3 or image.shape[0] == 1:
                    image = np.transpose(image, (1, 2, 0))
                if image.shape[2] == 4:
                    image = image[:, :, :3]
                elif image.shape[2] == 1:
                    image = np.repeat(image, 3, axis=2)
            image = Image.fromarray(image)
        elif not isinstance(image, Image.Image):
            try:
                image = Image.fromarray(np.array(image)).convert('RGB')
            except Exception as e:
                raise ValueError(f"Cannot convert image type {type(image)} to PIL Image: {e}")
        
        if isinstance(image, Image.Image) and image.mode != 'RGB':
            image = image.convert('RGB')
        
        # Get preprocessing config
        preproc_config = get_model_preprocessing(model_type)
        input_size = preproc_config['input_size']
        norm_mean = torch.tensor(preproc_config['norm_mean'])
        norm_std = torch.tensor(preproc_config['norm_std'])
        n_classes = preproc_config['n_classes']
        
        # Create transform
        if config.get('inference_normalization', False):
            transform = transforms.Compose([
                transforms.Resize(input_size),
                transforms.CenterCrop(input_size),
                transforms.ToTensor(),
                transforms.Normalize(norm_mean.tolist(), norm_std.tolist()),
            ])
            print(f"Fall-back: Using normalization with mean={norm_mean.tolist()}, std={norm_std.tolist()}")
        else:
            transform = transforms.Compose([
                transforms.Resize(input_size),
                transforms.CenterCrop(input_size),
                transforms.ToTensor(),
            ])
            print(f"Normalization OFF - feeding raw [0,1] tensors to model (normalization applied in the model)")
        
        # Helper function to safely apply transform with fallback for numpy compatibility
        def safe_transform(img):
            try:
                return transform(img)
            except TypeError as e:
                if "expected np.ndarray" in str(e) or "got numpy.ndarray" in str(e):
                    # Fallback: manually convert PIL to tensor
                    print(f"[WARNING] Transform failed with numpy compatibility issue, using manual conversion")
                    # Apply resize and center crop manually
                    resize_transform = transforms.Resize(input_size)
                    crop_transform = transforms.CenterCrop(input_size)
                    img = crop_transform(resize_transform(img))
                    # Convert to numpy array and then to tensor using torch.tensor() to avoid numpy compatibility issues
                    img_array = np.array(img, dtype=np.uint8)
                    # Use torch.tensor() instead of torch.from_numpy() to avoid compatibility issues
                    # Convert to float and normalize to [0, 1], then convert from HWC to CHW format
                    img_tensor = torch.tensor(img_array, dtype=torch.float32).div(255.0).permute(2, 0, 1)
                    # Apply normalization if needed
                    if config.get('inference_normalization', False):
                        img_tensor = transforms.Normalize(norm_mean.tolist(), norm_std.tolist())(img_tensor)
                    return img_tensor
                else:
                    raise
        
        # Prepare image tensor with safe transform
        image_tensor = safe_transform(image).unsqueeze(0).to(device)
        image_tensor.requires_grad = True

        # Get model components
        is_sequential = isinstance(model, nn.Sequential)
        if is_sequential and isinstance(model[0], NormalizeByChannelMeanStd):
            core_model = model[1]
        else:
            core_model = model

        # Prepare model for layer extraction
        if config.get('top_layer', 'all') != 'all':
            new_model = extract_middle_layers(core_model, config['top_layer'])
        else:
            new_model = model

        # Get original predictions
        with torch.no_grad():
            if config.get('inference_normalization', False):
                output_original = model(image_tensor)
            else:
                output_original = core_model(image_tensor)
            
            probs_orig = F.softmax(output_original, dim=1)
            conf_orig, classes_orig = torch.max(probs_orig, 1)
            
            # Get least confident classes for IncreaseConfidence
            if config['loss_infer'] == 'IncreaseConfidence':
                _, least_confident_classes = torch.topk(probs_orig, k=int(n_classes / 10), largest=False)

        # Setup for Prior-Guided Drift Diffusion
        noisy_features = None
        if config['loss_infer'] == 'Prior-Guided Drift Diffusion':
            print(f"Setting up Prior-Guided Drift Diffusion...")
            added_noise = config.get('initial_inference_noise_ratio', 0.05) * torch.randn_like(image_tensor).to(device)
            noisy_image_tensor = image_tensor + added_noise
            noisy_features = new_model(noisy_image_tensor)

        # Initialize inference step
        infer_step = InferStep(image_tensor, config['eps'], config['step_size'])
        
        # Storage for inference steps
        x = image_tensor.clone().detach().requires_grad_(True)
        all_steps = [image_tensor[0].detach().cpu()]
        
        selected_inferred_patterns = []
        perceived_categories = []
        confidence_list = []
        
        # Main inference loop
        print(f"Starting inference loop with {config['n_itr']} iterations for {config['loss_infer']}...")
        
        for i in range(config['n_itr']):
            # Reset gradients
            x.grad = None
            
            if i == 0:
                # Get predictions for first iteration
                if config.get('inference_normalization', False):
                    output = model(x)
                else:
                    output = core_model(x)
                
                if isinstance(output, torch.Tensor) and output.size(-1) == n_classes:
                    probs = F.softmax(output, dim=1)
                    conf, classes = torch.max(probs, 1)
                else:
                    probs = 0
                    conf = 0
                    classes = 'N/A'
            else:
                # Calculate loss and gradients
                try:
                    # Forward pass through new_model for feature extraction
                    features = new_model(x)
                    
                    if config['loss_infer'] == 'Prior-Guided Drift Diffusion':
                        assert config.get('loss_function', 'MSE') == 'MSE', "Prior-Guided Drift Diffusion requires MSE loss"
                        if noisy_features is not None:
                            loss = F.mse_loss(features, noisy_features)
                            grad = torch.autograd.grad(loss, x)[0]
                            adjusted_grad = infer_step.step(x, grad)
                        else:
                            raise ValueError("Noisy features not computed for Prior-Guided Drift Diffusion")
                    
                    elif config['loss_infer'] == 'IncreaseConfidence':
                        # Calculate loss using least confident classes
                        num_target_classes = min(int(n_classes / 10), least_confident_classes.size(1))
                        target_classes = least_confident_classes[0, :num_target_classes]
                        
                        loss = calculate_loss(features, target_classes.tolist(), config.get('loss_function', 'CE'))
                        grad = torch.autograd.grad(loss, x, retain_graph=True)[0]
                        adjusted_grad = infer_step.step(x, grad)
                    
                    else:
                        raise ValueError(f"Loss inference method {config['loss_infer']} not supported")
                    
                    if grad is None:
                        print("Fall-back: Direct gradient calculation failed")
                        random_noise = (torch.rand_like(x) - 0.5) * 2 * config['step_size']
                        x = infer_step.project(x.clone() + random_noise)
                    else:
                        # Add diffusion noise if specified
                        diffusion_noise = config.get('diffusion_noise_ratio', 0.0) * torch.randn_like(x).to(device)
                        x = infer_step.project(x.clone() + adjusted_grad + diffusion_noise)
                        
                except Exception as e:
                    print(f"Fall-back: Error in gradient calculation: {e}")
                    random_noise = (torch.rand_like(x) - 0.5) * 2 * config['step_size']
                    x = infer_step.project(x.clone() + random_noise)
            
            # Store step if in iterations_to_show
            if i+1 in config.get('iterations_to_show', []) or i+1 == config['n_itr']:
                all_steps.append(x[0].detach().cpu())
                selected_inferred_patterns.append(x[0].detach().cpu())
                
                # Get current predictions
                with torch.no_grad():
                    if config.get('inference_normalization', False):
                        current_output = model(x)
                    else:
                        current_output = core_model(x)
                    
                    if isinstance(current_output, torch.Tensor) and current_output.size(-1) == n_classes:
                        current_probs = F.softmax(current_output, dim=1)
                        current_conf, current_classes = torch.max(current_probs, 1)
                        perceived_categories.append(current_classes.item())
                        confidence_list.append(current_conf.item())
                    else:
                        perceived_categories.append('N/A')
                        confidence_list.append(0.0)
        
        # Final predictions
        with torch.no_grad():
            if config.get('inference_normalization', False):
                final_output = model(x)
            else:
                final_output = core_model(x)
                
            final_probs = F.softmax(final_output, dim=1)
            final_conf, final_classes = torch.max(final_probs, 1)
            
            total_time = time.time() - inference_start
            
            print(f"Original top class: {classes_orig.item()} ({conf_orig.item():.4f})")
            print(f"Final top class: {final_classes.item()} ({final_conf.item():.4f})")
            print(f"Total inference time: {total_time:.2f} seconds")
            
        # Return results in Code 1 format
        return {
            'final_image': x[0].detach().cpu(),
            'steps': all_steps,
            'original_class': classes_orig.item(),
            'original_confidence': conf_orig.item(),
            'final_class': final_classes.item(),
            'final_confidence': final_conf.item(),
            'all_categories': perceived_categories,
            'all_confidences': confidence_list,
        }

def show_inference_steps(steps, figsize=(15, 10)):
    """Show inference steps using matplotlib."""
    try:
        import matplotlib.pyplot as plt
        
        n_steps = len(steps)
        fig, axes = plt.subplots(1, n_steps, figsize=figsize)
        
        if n_steps == 1:
            axes = [axes]
        
        for i, step_img in enumerate(steps):
            if isinstance(step_img, torch.Tensor):
                img = step_img.permute(1, 2, 0).numpy()
                img = np.clip(img, 0, 1)
            else:
                img = step_img
                
            axes[i].imshow(img)
            axes[i].set_title(f"Step {i+1}")
            axes[i].axis('off')
        
        plt.tight_layout()
        return fig
        
    except ImportError:
        print("Fall-back: matplotlib not available for visualization")
        return None
    except Exception as e:
        print(f"Fall-back: Visualization failed: {e}")
        return None

# Export the main classes and functions
__all__ = ['GenerativeInferenceModel', 'get_inference_configs', 'show_inference_steps']