clean up

Author: elenaivadreyer

.DS_Store

@@ -59,7 +59,7 @@ repos:
         #Automatic linting fixes
         types_or: [ python, pyi, jupyter ]
         args: [ --fix ]
-        stages: [manual]
+        #stages: [manual]
       - id: ruff-format
         types_or: [ python, pyi, jupyter ]
-        stages: [manual]
+        #stages: [manual]

.pre-commit-config.yaml

@@ -1,13 +1,10 @@
+from pathlib import Path
+
 import numpy as np
 from PIL import Image
-from pathlib import Path
 
-def extract_pv_masks(
-    src_dir: Path,
-    dst_dir: Path,
-    pv_class: int = 0,
-    background_value: int = 255
-):
+
+def extract_pv_masks(src_dir: Path, dst_dir: Path, pv_class: int = 0, background_value: int = 255):
     """
     Extract masks that contain PV modules (class 0) and save a cleaned PV-only mask.
 
@@ -21,6 +18,7 @@ def extract_pv_masks(
         Class index representing PV modules (default: 0).
     background_value : int, optional
         Value used for non-PV pixels in the output mask (default: 255 for easy visualization).
+
     """
     dst_dir.mkdir(parents=True, exist_ok=True)
 
@@ -46,6 +44,7 @@ def extract_pv_masks(
 
     print(f"Saved {count_saved} PV-only masks to: {dst_dir}")
 
+
 if __name__ == "__main__":
     BASE_DIR = Path(__file__).resolve().parent
     SRC_MASK_DIR = BASE_DIR / "masks_superstructures_reviewed"
@@ -55,5 +54,5 @@ def extract_pv_masks(
         src_dir=SRC_MASK_DIR,
         dst_dir=DST_PV_MASK_DIR,
         pv_class=0,
-        background_value=255,   # white background for visualization
-    )
+        background_value=255,  # white background for visualization
+    )

data/generate_pv_masks.py

@@ -19,7 +19,7 @@
 from pathlib import Path
 
 SERVER = "rsync://m1655470@dataserv.ub.tum.de/m1655470"
-BASE_DIR = Path(__file__).resolve().parent     # data/
+BASE_DIR = Path(__file__).resolve().parent  # data/
 DATA_DIR = BASE_DIR
 
 FOLDERS = [
@@ -51,7 +51,7 @@ def rsync_folder(remote_name: str):
         "-av",
         "--progress",
         src,
-        str(dst) + "/",     # ensure trailing slash
+        str(dst) + "/",  # ensure trailing slash
     ]
 
     print(f">>> Downloading '{remote_name}' from server...")

data/load_rid_data.py

@@ -1,7 +1,12 @@
-Few-shot learning for satellite image segmentation is crucial for urban planning and disaster response. Traditional deep learning requires thousands of labeled samples, but few-shot learning enables rapid adaptation to new geographic regions with minimal labeled data.
+## Memo
 
-**Policy Applications:**
-- Urban development monitoring
-- Disaster damage assessment
-- Property tax assessment
-- Solar panel installation planning
+This notebook offers an in-depth look at few-shot learning (FSL), an advanced deep learning technique that enables models to generalize effectively from only a small number of training examples. FSL becomes particularly valuable when traditional supervised learning is impractical. For instance, when annotation is expensive, requires specialized expertise, or when data is scarce.
+
+Few-shot learning is especially relevant in the context of public policy, where we often face limited access to comprehensive, well-curated datasets or lack the resources to collect large amounts of training data. Applications span a wide range of policy areas, including disaster classification (Lee et al., 2025), urban planning (Hu et al., 2022) and health policy, where FSL has supported COVID-19 detection (Jadon, 2021) and the diagnosis of rare genetic diseases (Alsentzer et al., 2025).
+
+Beyond classification tasks, FSL also performs well in <b>segmentation problems</b>, helping identify different types of buildings, vegetation such as forest cover (Puthumanaillam & Verma, 2023) and performing well in remote sensing imagery segmentation (Chen et al., 2022). All of these examples share a common challenge: <b>limited data</b>.
+
+Few-shot learning helps address this constraint, and we encounter similar situations within our own institution. A clear illustration is our Wildlife Trade Monitoring Project (WTMP), where labeled examples of rare, endangered species are extremely scarce and often inconsistent. This emphasizes just how valuable few-shot learning approaches can be.
+Today, a wide range of few-shot learning frameworks and architectures are available. Some of the most widely used include model-agnostic meta-learning (Finn, Abbeel & Levine, 2017), prototypical networks (Snell, Swersky & Zemel, 2017), and relation networks (Sung et al., 2018). Over the past few years, the field has continued to evolve, giving rise to more specialized and state-of-the-art techniques, such as SegPPD-FS for semantic segmentation (Ge et al., 2025) and CDCNet for classification tasks (Li et al., 2025).
+
+Because of its importance to public policy and its potential value for our organization, this notebook takes a deep dive into prototypical networks semantic segmentation of building rooftops. This is an essential step for many downstream tasks, such as assessing rooftop suitability for solar panel installation. By walking through this method, the notebook aims to help our team to better understand and apply this powerful technique in our day-to-day work.