Spaces:
Paused
Paused
| """ | |
| Grad-CAM overlay rendering. | |
| Produces the side-by-side figure the brief asks for: the original scene | |
| next to the same scene with the attribution map blended over it, red where | |
| influence is high and blue where it is low, with the named regions boxed. | |
| Written to `reports/gradcam/`. | |
| """ | |
| from __future__ import annotations | |
| import logging | |
| from pathlib import Path | |
| import cv2 | |
| import matplotlib | |
| matplotlib.use("Agg") | |
| import matplotlib.pyplot as plt | |
| import numpy as np | |
| from explainability.gradcam_explainer.config import GradCamConfig | |
| from explainability.gradcam_explainer.gradcam import GradCamExplanation | |
| logger = logging.getLogger(__name__) | |
| def render_heatmap(heatmap: np.ndarray, config: GradCamConfig) -> np.ndarray: | |
| """Colourise a normalised attribution map as RGB (red high, blue low).""" | |
| colormap = getattr(cv2, f"COLORMAP_{config.colormap}", cv2.COLORMAP_JET) | |
| coloured = cv2.applyColorMap((np.clip(heatmap, 0, 1) * 255).astype(np.uint8), colormap) | |
| return cv2.cvtColor(coloured, cv2.COLOR_BGR2RGB) | |
| def blend_overlay( | |
| image: np.ndarray, heatmap: np.ndarray, config: GradCamConfig | |
| ) -> np.ndarray: | |
| """Blend the colourised attribution over the scene.""" | |
| if image.shape[:2] != heatmap.shape[:2]: | |
| heatmap = cv2.resize(heatmap, (image.shape[1], image.shape[0])) | |
| coloured = render_heatmap(heatmap, config) | |
| alpha = float(np.clip(config.overlay_alpha, 0.0, 1.0)) | |
| return cv2.addWeighted(coloured, alpha, image.astype(np.uint8), 1.0 - alpha, 0) | |
| def save_overlay( | |
| image_path: str | Path, | |
| explanation: GradCamExplanation, | |
| config: GradCamConfig, | |
| *, | |
| filename: str | None = None, | |
| ) -> Path: | |
| """Render original + overlay side by side with the named regions boxed.""" | |
| from ai_models.satellite_model.preprocessing import load_rgb | |
| config.ensure_output_dirs() | |
| image = load_rgb(image_path) | |
| overlay = blend_overlay(image, explanation.heatmap, config) | |
| fig, (left, right) = plt.subplots(1, 2, figsize=(13, 6.4)) | |
| left.imshow(image) | |
| left.set_title(f"Satellite scene — {Path(image_path).stem}", fontsize=10) | |
| left.axis("off") | |
| right.imshow(overlay) | |
| right.set_title( | |
| f"Grad-CAM ({explanation.model_name}, layer {explanation.target_layer})\n" | |
| f"satellite risk {explanation.satellite_risk:.2f} — " | |
| f"{explanation.coverage_label} high-influence coverage " | |
| f"({explanation.coverage * 100:.1f}% of scene)", | |
| fontsize=10, | |
| ) | |
| right.axis("off") | |
| height, width = image.shape[:2] | |
| for region in explanation.regions: | |
| x0, y0, x1, y1 = region.bbox | |
| rectangle = plt.Rectangle( | |
| (x0 * width, y0 * height), | |
| (x1 - x0) * width, | |
| (y1 - y0) * height, | |
| fill=False, edgecolor="white", linewidth=1.6, | |
| ) | |
| right.add_patch(rectangle) | |
| right.text( | |
| x0 * width, max(y0 * height - 6, 10), region.name, | |
| color="white", fontsize=8, | |
| bbox=dict(facecolor="black", alpha=0.55, pad=1.5, edgecolor="none"), | |
| ) | |
| fig.suptitle("Red = high influence on the classification · Blue = low", fontsize=9, y=0.03) | |
| fig.tight_layout() | |
| path = config.reports_dir / (filename or f"gradcam_{Path(image_path).stem}.png") | |
| fig.savefig(path, dpi=150, bbox_inches="tight") | |
| plt.close(fig) | |
| logger.info("Grad-CAM overlay -> %s", path) | |
| return path | |