Merge pull request #31 from hotosm/feature/yolo_v8_v2

Feature : Yolo v8 Seg Integration

.github/workflows/build.yml

@@ -11,10 +11,10 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v2
-      - name: Set up Python 3.8
+      - name: Set up Python 3.9
         uses: actions/setup-python@v1
         with:
-          python-version: 3.8
+          python-version: 3.9.20
 
       - name: Remove solaris
         run: sudo rm -rf ./docker ./weights
@@ -46,19 +46,20 @@ jobs:
       - name: Install tensorflow
         run: pip install tensorflow==2.9.2
 
-      - name : Install pytorch and ultralytics
-        run: pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113 ultralytics==8.1.6
+      - name: Install pytorch and ultralytics
+        run: pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --extra-index-url https://download.pytorch.org/whl/cu113 ultralytics==8.3.26
 
       - name: Install fair utilities
         run: pip install -e .
 
       - name: Run test ramp
         run: |
           pip uninstall -y gdal
-          pip install numpy
+          pip install numpy==1.26.4
           pip install GDAL==$(gdal-config --version) --global-option=build_ext --global-option="-I/usr/include/gdal"
-          python test_app.py
+          python test_ramp.py
 
-      - name : Run test yolo
-        run : |
-          python test_yolo.py
+      - name: Run test yolo
+        run: |
+          python test_yolo_v1.py
+          python test_yolo_v2.py

CHANGELOG.md

@@ -1,3 +1,26 @@
+## v3.0.0 (2024-11-15)
+
+### Feat
+
+- **yololib**: adds lib required for packaging yolo code
+
+### Fix
+
+- **bundlelib**: bundles lib itself with new pandas version
+
+## v2.0.0 (2024-11-15)
+
+### Feat
+
+- **yololib**: adds lib required for packaging yolo code
+- replace FastSAM with YOLO including training
+- add FastSAM inference
+
+### Fix
+
+- **postprocessing/utils**: resolve OAM-x-y-z.mask.tif
+- **predict**: support both .png and .tif in inference
+
 ## v1.3.0 (2024-10-04)
 
 ### Feat

README.md

@@ -6,6 +6,7 @@ Initially lib was developed during Open AI Challenge with [Omdeena](https://omde
 
 Installing all libraries could be pain so we suggest you to use docker , If you like to do it bare , You can follow `.github/build.yml`
 
+<!-- comment -->
 
 Clone repo
 

hot_fair_utilities/__init__.py

@@ -1,6 +1,7 @@
 from .georeferencing import georeference
-from .inference import predict, evaluate
+from .inference import evaluate, predict
 from .postprocessing import polygonize, vectorize
-from .preprocessing import preprocess
-from .training import train
+from .preprocessing import preprocess, yolo_v8_v1
+
+# from .training import ramp, yolo_v8_v1
 from .utils import bbox2tiles, tms2img

hot_fair_utilities/georeferencing.py

@@ -11,7 +11,7 @@
 from .utils import get_bounding_box
 
 
-def georeference(input_path: str, output_path: str, is_mask=False) -> None:
+def georeference(input_path: str, output_path: str, is_mask=False,epsg=3857) -> None:
     """Perform georeferencing and remove the fourth band from images (if any).
 
     CRS of the georeferenced images will be EPSG:3857 ('WGS 84 / Pseudo-Mercator').
@@ -38,7 +38,7 @@ def georeference(input_path: str, output_path: str, is_mask=False) -> None:
         out_file = f"{output_path}/{filename}.tif"
 
         # Get bounding box in EPSG:3857
-        x_min, y_min, x_max, y_max = get_bounding_box(filename)
+        x_min, y_min, x_max, y_max = get_bounding_box(filename,epsg=epsg)
 
         # Use one band for masks and the first three bands for images
         bands = [1] if is_mask else [1, 2, 3]
@@ -51,7 +51,7 @@ def georeference(input_path: str, output_path: str, is_mask=False) -> None:
             format="GTiff",
             bandList=bands,
             outputBounds=[x_min, y_max, x_max, y_min],
-            outputSRS="EPSG:3857",
+            outputSRS=f"EPSG:{epsg}",
         )
         # Close dataset
         _ = None

hot_fair_utilities/inference/evaluate.py

@@ -1,12 +1,21 @@
 # Patched from ramp-code.scripts.calculate_accuracy.iou created for ramp project by [email protected]
 
+# Standard library imports
 from pathlib import Path
+
+# Third party imports
 import geopandas as gpd
 
-from ramp.utils.eval_utils import get_iou_accuracy_metrics
+try:
+    # Third party imports
+    from ramp.utils.eval_utils import get_iou_accuracy_metrics
+except ImportError:
+    print("Ramp eval metrics are not available, Possibly ramp is not installed")
 
 
-def evaluate(test_path, truth_path, filter_area_m2=None, iou_threshold=0.5, verbose=False):
+def evaluate(
+    test_path, truth_path, filter_area_m2=None, iou_threshold=0.5, verbose=False
+):
     """
     Calculate precision/recall/F1-score based on intersection-over-union accuracy evaluation protocol defined by RAMP.
 
@@ -29,9 +38,9 @@ def evaluate(test_path, truth_path, filter_area_m2=None, iou_threshold=0.5, verb
     truth_df, test_df = gpd.read_file(str(truth_path)), gpd.read_file(str(test_path))
     metrics = get_iou_accuracy_metrics(test_df, truth_df, filter_area_m2, iou_threshold)
 
-    n_detections = metrics['n_detections']
+    n_detections = metrics["n_detections"]
     n_truth = metrics["n_truth"]
-    n_truepos = metrics['true_pos']
+    n_truepos = metrics["true_pos"]
     n_falsepos = n_detections - n_truepos
     n_falseneg = n_truth - n_truepos
     agg_precision = n_truepos / n_detections

hot_fair_utilities/inference/predict.py

@@ -12,15 +12,19 @@
 
 from ..georeferencing import georeference
 from ..utils import remove_files
-from .utils import open_images, save_mask, initialize_model
+from .utils import initialize_model, open_images, save_mask
 
 BATCH_SIZE = 8
 IMAGE_SIZE = 256
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
 
 def predict(
-    checkpoint_path: str, input_path: str, prediction_path: str, confidence: float = 0.5, remove_images=True
+    checkpoint_path: str,
+    input_path: str,
+    prediction_path: str,
+    confidence: float = 0.5,
+    remove_images=True,
 ) -> None:
     """Predict building footprints for aerial images given a model checkpoint.
 
@@ -46,7 +50,7 @@ def predict(
     """
     start = time.time()
     print(f"Using : {checkpoint_path}")
-    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     model = initialize_model(checkpoint_path, device=device)
     print(f"It took {round(time.time()-start)} sec to load model")
     start = time.time()
@@ -74,14 +78,24 @@ def predict(
                 )
     elif isinstance(model, YOLO):
         for idx in range(0, len(image_paths), BATCH_SIZE):
-            batch = image_paths[idx:idx + BATCH_SIZE]
-            for i, r in enumerate(model(batch, stream=True, conf=confidence, verbose=False)):
-                if r.masks is None:
-                    preds = np.zeros((IMAGE_SIZE, IMAGE_SIZE,), dtype=np.float32)
+            batch = image_paths[idx : idx + BATCH_SIZE]
+            for i, r in enumerate(
+                model.predict(batch, conf=confidence, imgsz=IMAGE_SIZE, verbose=False)
+            ):
+                if hasattr(r, "masks") and r.masks is not None:
+                    preds = (
+                        r.masks.data.max(dim=0)[0].detach().cpu().numpy()
+                    )  # Combine masks and convert to numpy
+
                 else:
-                    preds = r.masks.data.max(dim=0)[0]  # dim=0 means to take only footprint
-                    preds = torch.where(preds > confidence, torch.tensor(1), torch.tensor(0))
-                    preds = preds.detach().cpu().numpy()
+                    preds = np.zeros(
+                        (
+                            IMAGE_SIZE,
+                            IMAGE_SIZE,
+                        ),
+                        dtype=np.float32,
+                    )  # Default if no masks
+
                 save_mask(preds, str(f"{prediction_path}/{Path(batch[i]).stem}.png"))
     else:
         raise RuntimeError("Loaded model is not supported")

hot_fair_utilities/postprocessing/building_footprint.py

@@ -1,5 +1,7 @@
+# Standard library imports
 import collections
 
+# Third party imports
 from geopandas import GeoSeries
 from shapely.geometry import MultiPolygon, Polygon
 from shapely.ops import unary_union
@@ -75,4 +77,4 @@ def extract(self, tile, mask):
                 self.features.append(feature)
 
     def save(self, out):
-        GeoSeries(self.features).set_crs(CRS).to_file(out)
+        GeoSeries(self.features).set_crs(CRS).to_file(out, driver="GeoJSON")

hot_fair_utilities/postprocessing/merge_polygons.py

@@ -3,6 +3,7 @@
 from shapely.geometry import MultiPolygon, Polygon
 from shapely.validation import make_valid
 from tqdm import tqdm
+import os
 
 from .utils import UndirectedGraph, make_index, project, union
 
@@ -19,7 +20,7 @@ def merge_polygons(polygons_path, new_polygons_path, distance_threshold):
       new_polygons_path: Path to GeoJSON file where the merged polygons will be saved
       distance_threshold: Minimum distance to define adjacent polygons, in meters
     """
-    gdf = read_file(polygons_path)
+    gdf = read_file(os.path.relpath(polygons_path))
     shapes = list(gdf["geometry"])
 
     graph = UndirectedGraph()
@@ -71,4 +72,4 @@ def unbuffered(shape):
             features.append(feature)
 
     gs = GeoSeries(features).set_crs(SOURCE_CRS)
-    gs.simplify(TOLERANCE).to_file(new_polygons_path)
+    gs.simplify(TOLERANCE).to_file(new_polygons_path, driver="GeoJSON")

hot_fair_utilities/postprocessing/vectorize.py

@@ -15,7 +15,9 @@
 AREA_THRESHOLD = 5
 
 
-def vectorize(input_path: str, output_path: str , tolerance: float = 0.5, area_threshold: float = 5) -> None:
+def vectorize(
+    input_path: str, output_path: str, tolerance: float = 0.5, area_threshold: float = 5
+) -> None:
     """Polygonize raster tiles from the input path.
 
     Note that as input, we are expecting GeoTIF images with EPSG:3857 as
@@ -52,15 +54,14 @@ def vectorize(input_path: str, output_path: str , tolerance: float = 0.5, area_t
     polygons = [
         Polygon(poly.exterior.coords)
         for poly in polygons
-        if poly.area != max_area
-        and poly.area / median_area > area_threshold
+        if poly.area != max_area and poly.area / median_area > area_threshold
     ]
 
     gs = gpd.GeoSeries(polygons, crs=kwargs["crs"]).simplify(tolerance)
     gs = remove_overlapping_polygons(gs)
     if gs.empty:
         raise ValueError("No Features Found")
-    gs.to_crs("EPSG:4326").to_file(output_path)
+    gs.to_crs("EPSG:4326").to_file(output_path, driver="GeoJSON")
 
 
 def remove_overlapping_polygons(gs: gpd.GeoSeries) -> gpd.GeoSeries:
@@ -79,4 +80,4 @@ def remove_overlapping_polygons(gs: gpd.GeoSeries) -> gpd.GeoSeries:
                 else:
                     to_remove.add(j)
 
-    return gs.drop(list(to_remove))
+    return gs.drop(list(to_remove))

hot_fair_utilities/preprocessing/__init__.py

@@ -1 +1,2 @@
 from .preprocess import preprocess
+from .yolo_v8_v1 import yolo_format

hot_fair_utilities/preprocessing/clip_labels.py

@@ -3,6 +3,7 @@
 from glob import glob
 from pathlib import Path
 
+# Third party imports
 # Third-party imports
 import geopandas
 from osgeo import gdal
@@ -13,7 +14,12 @@
 
 
 def clip_labels(
-    input_path: str, output_path: str, rasterize=False, rasterize_options=None
+    input_path: str,
+    output_path: str,
+    rasterize=False,
+    rasterize_options=None,
+    all_geojson_file=None,
+    epsg=3857,
 ) -> None:
     """Clip and rasterize the GeoJSON labels for each aerial image.
 
@@ -71,24 +77,29 @@ def clip_labels(
         glob(f"{input_path}/*.png"), desc=f"Clipping labels for {Path(input_path).stem}"
     ):
         filename = Path(path).stem
-        geojson_file_all_labels = f"{output_path}/labels_epsg3857.geojson"
+        if all_geojson_file:
+            geojson_file_all_labels = all_geojson_file
+        else:
+            geojson_file_all_labels = f"{output_path}/labels_epsg3857.geojson"
         clipped_geojson_file = f"{output_geojson_path}/{filename}.geojson"
 
         # Bounding box as a tuple
-        x_min, y_min, x_max, y_max = get_bounding_box(filename)
+        x_min, y_min, x_max, y_max = get_bounding_box(filename, epsg=epsg)
         # Bounding box as a polygon
         bounding_box_polygon = box(x_min, y_min, x_max, y_max)
 
         # Read all labels into a GeoDataFrame, clip it and
         # write to GeoJSON
-        gdf_all_labels = geopandas.read_file(geojson_file_all_labels)
+        gdf_all_labels = geopandas.read_file(os.path.relpath(geojson_file_all_labels))
         gdf_clipped = gdf_all_labels.clip(bounding_box_polygon)
         if len(gdf_clipped) > 0:
-            gdf_clipped.to_file(clipped_geojson_file)
+            gdf_clipped.to_file(clipped_geojson_file, driver="GeoJSON")
         else:
             schema = {"geometry": "Polygon", "properties": {"id": "int"}}
-            crs = "EPSG:3857"
-            gdf_clipped.to_file(clipped_geojson_file, schema=schema, crs=crs)
+            crs = f"EPSG:{epsg}"
+            gdf_clipped.to_file(
+                clipped_geojson_file, schema=schema, crs=crs, driver="GeoJSON"
+            )
 
         # Rasterizing
         if rasterize:

hot_fair_utilities/preprocessing/fix_labels.py

@@ -1,7 +1,7 @@
 # Third party imports
 import geopandas
 from shapely.validation import explain_validity, make_valid
-
+import os
 
 def remove_self_intersection(row):
     """Fix self-intersections in the polygons.
@@ -29,8 +29,9 @@ def fix_labels(input_path: str, output_path: str) -> None:
         input_path: Path to the GeoJSON file where the input data are stored.
         output_path: Path to the GeoJSON file where the output data will go.
     """
-    gdf = geopandas.read_file(input_path)
+    gdf = geopandas.read_file(os.path.relpath(input_path))
+    # print(gdf)
     if gdf.empty:
         raise ValueError("Error: gdf is empty, No Labels found : Check your labels")
     gdf["geometry"] = gdf.apply(remove_self_intersection, axis=1)
-    gdf.to_file(output_path)
+    gdf.to_file(output_path, driver="GeoJSON")

hot_fair_utilities/preprocessing/multimasks_from_polygons.py

@@ -2,7 +2,7 @@
 
 # Standard library imports
 from pathlib import Path
-
+import os
 # Third party imports
 import geopandas as gpd
 import rasterio as rio
@@ -88,7 +88,7 @@ def multimasks_from_polygons(
         # workaround for bug in solaris
         mask_shape, mask_transform = get_rasterio_shape_and_transform(chip_path)
 
-        gdf = gpd.read_file(json_path)
+        gdf = gpd.read_file(os.path.relpath(json_path))
 
         # remove empty and null geometries
         gdf = gdf[~gdf["geometry"].isna()]