File size: 6,197 Bytes
ada7b0f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d298346
 
 
 
ada7b0f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d298346
 
 
ada7b0f
 
 
 
 
 
 
 
 
9163dd2
 
 
 
ada7b0f
 
 
 
9163dd2
 
 
 
 
ada7b0f
 
 
 
 
9163dd2
 
 
 
ada7b0f
 
 
 
 
0c81eb2
ada7b0f
 
 
 
 
 
 
0c81eb2
 
ada7b0f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0c81eb2
ada7b0f
 
 
d298346
 
 
 
 
 
ada7b0f
 
d298346
ada7b0f
 
 
 
 
d298346
ada7b0f
 
 
d298346
ada7b0f
 
d298346
ada7b0f
 
 
 
d298346
ada7b0f
 
 
0c81eb2
ada7b0f
 
d298346
ada7b0f
 
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
"""Explainability module for LULC models."""

import torch
import torch.nn.functional as F
import numpy as np
import cv2
import base64
import matplotlib.pyplot as plt
from io import BytesIO
import PIL.Image

from pytorch_grad_cam import GradCAM, GradCAMPlusPlus
from pytorch_grad_cam.utils.image import show_cam_on_image
from lime import lime_image
from skimage.segmentation import mark_boundaries
import logging

# Set up logging
logger = logging.getLogger("lulc-explainers")

def array_to_base64(img_array):
    """Convert numpy RGB array to base64 jpeg."""
    # Convert RGB to BGR for cv2
    img_bgr = cv2.cvtColor(img_array, cv2.COLOR_RGB2BGR)
    _, buffer = cv2.imencode('.jpg', img_bgr)
    return base64.b64encode(buffer).decode('utf-8')


def get_saliency_map(model, image_tensor):
    """Generate vanilla Saliency map."""
    # Make a copy attached to gradients
    x = image_tensor.clone()
    x.requires_grad = True
    model.eval()
    
    # Forward pass
    logits = model(x)
    
    # Get top class
    score, indices = torch.max(logits, 1)
    
    # Backward pass
    model.zero_grad()
    score.backward()
    
    # Saliency map = max absolute gradient across channels
    saliency, _ = torch.max(x.grad.data.abs(), dim=1)
    saliency = saliency.squeeze().cpu().numpy()
    
    # Normalize to 0-1
    saliency = (saliency - saliency.min()) / (saliency.max() - saliency.min() + 1e-8)
    
    # Apply colormap (jet)
    saliency_heatmap = cv2.applyColorMap(np.uint8(255 * saliency), cv2.COLORMAP_JET)
    saliency_heatmap = cv2.cvtColor(saliency_heatmap, cv2.COLOR_BGR2RGB)
    
    # Resize to original tensor size for consistency
    _, _, H, W = image_tensor.shape
    saliency_heatmap = cv2.resize(saliency_heatmap, (W, H))
    
    return array_to_base64(saliency_heatmap)


def get_gradcam_maps(model, image_tensor, orig_image_np):
    """Generate GradCAM and GradCAM++."""
    # Target layer for AMFRNet is the last block
    target_layers = [model.block3]
    
    # Normalize original image for overlay
    orig_img_norm = orig_image_np.astype(np.float32) / 255.0
    h, w = orig_img_norm.shape[:2]
    
    # 1. GradCAM
    cam_algo = GradCAM(model=model, target_layers=target_layers)
    grayscale_cam = cam_algo(input_tensor=image_tensor, targets=None)[0]
    
    # Force resize grayscale_cam to match orig_img_norm exactly to avoid broadcast errors
    if grayscale_cam.shape[0] != h or grayscale_cam.shape[1] != w:
        logger.warning(f"[explainers] Shape mismatch! Resizing GradCAM heatmap from {grayscale_cam.shape} to ({h}, {w})")
        grayscale_cam = cv2.resize(grayscale_cam, (w, h))
        
    cam_image = show_cam_on_image(orig_img_norm, grayscale_cam, use_rgb=True)
    
    # 2. GradCAM++
    cam_plus_algo = GradCAMPlusPlus(model=model, target_layers=target_layers)
    grayscale_cam_plus = cam_plus_algo(input_tensor=image_tensor, targets=None)[0]
    
    if grayscale_cam_plus.shape[0] != h or grayscale_cam_plus.shape[1] != w:
        grayscale_cam_plus = cv2.resize(grayscale_cam_plus, (w, h))
        
    cam_plus_image = show_cam_on_image(orig_img_norm, grayscale_cam_plus, use_rgb=True)
    
    return array_to_base64(cam_image), array_to_base64(cam_plus_image)


def get_lime_map(model, preprocessor, image_bytes, orig_image_np, device, model_type):
    """Generate LIME explanation."""
    explainer = lime_image.LimeImageExplainer()
    
    def predict_fn(images):
        tensors = []
        for img_arr in images:
            img_pil = PIL.Image.fromarray(img_arr)
            transform = preprocessor.get_transforms(model_type, img_pil.width, img_pil.height)
            tensor = transform(img_pil).to(device)
            tensors.append(tensor)
        
        batch_tensor = torch.stack(tensors)
        with torch.no_grad():
            logits = model(batch_tensor)
            probs = torch.softmax(logits, dim=1)
        return probs.cpu().numpy()
    
    # Run explanation (num_samples=100 for decent speed, default 1000 is too slow for API)
    explanation = explainer.explain_instance(
        orig_image_np, 
        predict_fn, 
        top_labels=1, 
        hide_color=0, 
        num_samples=150
    )
    
    temp, mask = explanation.get_image_and_mask(
        explanation.top_labels[0], 
        positive_only=True, 
        num_features=5, 
        hide_rest=False
    )
    
    img_boundry = mark_boundaries(temp / 255.0, mask)
    img_boundry_uint8 = np.uint8(img_boundry * 255)
    
    return array_to_base64(img_boundry_uint8)

def generate_all_explanations(model, preprocessor, image_tensor, image_bytes, model_type):
    """Wrapper function to generate all maps and return a dict of base64 strings."""
    device = image_tensor.device
    
    # Dynamically determine the target size from the image tensor (1, 3, H, W)
    _, _, H, W = image_tensor.shape
    target_size = (W, H) # PIL uses (width, height)
    
    logger.info(f"[explainers] Target size for explanations: {target_size}")
    
    # Original image decoding for overlay
    orig_pil = PIL.Image.open(BytesIO(image_bytes)).convert("RGB")
    orig_pil = orig_pil.resize(target_size)
    orig_np = np.array(orig_pil)
    
    result = {}
    
    try:
        logger.info("[explainers] generating Saliency...")
        saliency_b64 = get_saliency_map(model, image_tensor)
        result["Saliency"] = f"data:image/jpeg;base64,{saliency_b64}"
    except Exception as e:
        logger.error(f"[explainers] Saliency failed: {e}")
        
    try:
        logger.info("[explainers] generating GradCAM & GradCAM++...")
        gc_b64, gcp_b64 = get_gradcam_maps(model, image_tensor, orig_np)
        result["GradCAM"] = f"data:image/jpeg;base64,{gc_b64}"
        result["GradCAM++"] = f"data:image/jpeg;base64,{gcp_b64}"
    except Exception as e:
        logger.error(f"[explainers] GradCAM failed: {e}")
        
    try:
        print("[explainers] generating LIME...")
        lime_b64 = get_lime_map(model, preprocessor, image_bytes, orig_np, device, model_type)
        result["LIME"] = f"data:image/jpeg;base64,{lime_b64}"
    except Exception as e:
        logger.error(f"[explainers] LIME failed: {e}")
    
    return result