Merge pull request #101 from Oslandia/tanzania_datagen

Tanzania datagen

deeposlandia/datasets/tanzania.py

@@ -68,6 +68,76 @@ def __init__(self, img_size):
                        color=self.FOUNDATION_COLOR, is_evaluate=True)
 
 
+    def _generate_preprocessed_filenames(
+            self, image_filename, output_dir, x, y, suffix
+    ):
+        """Generate preprocessed image and label filenames on the file system,
+        starting from a raw image filename
+
+        Parameters
+        ----------
+        image_filename : str
+            Original image filename
+        output_dir : str
+            Output folder for preprocessed material
+        x : int
+            Extracted image west coordinates
+        y : int
+            Extracted image north coordinates
+        suffix : str
+            Preprocessed filename complement
+
+        Returns
+        -------
+        dict
+            Preprocessed image and corresponding label filenames
+        """
+        basename_decomp = os.path.splitext(os.path.basename(image_filename))
+        img_id_str = (str(self.image_size) + '_'
+                      + str(self.image_size) + '_'
+                      + str(x) + '_' + str(y) + "_" + suffix)
+        new_filename = basename_decomp[0] + '_' + img_id_str + ".png"
+        out_image_name = os.path.join(output_dir, 'images', new_filename)
+        out_label_name = out_image_name.replace("images", "labels")
+        return {"image": out_image_name, "labels": out_label_name}
+
+
+    def _serialize(
+            self, tile_image, labelled_image, label_dict,
+            image_filename, output_dir, x, y, suffix
+    ):
+        """Serialize a tiled image generated from an original high-resolution
+        raster as well as the labelled version of the tile
+
+        The method returns a dict that contains image-related file paths.
+
+        Parameters
+        ----------
+        tile_image : PIL.Image
+        labelled_image : PIL.Image
+        label_dict : dict
+        image_filename : str
+        output_dir : str
+        x : int
+        y : int
+
+        Returns
+        -------
+        dict
+            Information related to the serialized tile (file paths, encountered
+        labels)
+        """
+        dirs = self._generate_preprocessed_filenames(
+            image_filename, output_dir, x, y, suffix
+        )
+        tile_image.save(dirs["image"])
+        labelled_image.save(dirs["labels"])
+        return {"raw_filename": image_filename,
+                "image_filename": dirs["image"],
+                "label_filename": dirs["labels"],
+                "labels": label_dict}
+
+
     def _preprocess_tile(self, x, y, image_filename, output_dir,
                          raster, labels=None):
         """Preprocess one single tile built from `image_filename`, with respect
@@ -94,51 +164,24 @@ def _preprocess_tile(self, x, y, image_filename, output_dir,
             Key/values with the filenames and label ids
 
         """
-        basename_decomp = os.path.splitext(
-            os.path.basename(image_filename))
-        img_id_str = (str(self.image_size) + '_'
-                      + str(self.image_size) + '_'
-                      + str(x) + '_' + str(y))
-        new_in_filename = (basename_decomp[0] + '_'
-                           + img_id_str + ".png")
-        new_in_path = os.path.join(output_dir, 'images',
-                                   new_in_filename)
-        gdal.Translate(new_in_path, raster,
+        dirs = self._generate_preprocessed_filenames(
+            image_filename, output_dir, x, y
+        )
+        gdal.Translate(dirs["image"], raster,
                        format="PNG",
                        srcWin=[x, y, self.image_size, self.image_size])
-        if not labels is None:
-            raster_features = get_image_features(raster)
-            tile_items = extract_tile_items(raster_features, labels,
-                                            x, y,
-                                            self.image_size,
-                                            self.image_size,
-                                            tile_srid=32737)
-            out_labelname = (new_in_path
-                             .replace("images", "labels"))
-            mask = self.load_mask(tile_items, raster_features, x, y)
-            label_dict = utils.build_labels(mask,
-                                            range(self.get_nb_labels()),
-                                            "tanzania")
-            labelled_image = utils.build_image_from_config(mask,
-                                                           self.labels)
-            labelled_image.save(out_labelname)
-            return {"raw_filename": image_filename,
-                    "image_filename": new_in_path,
-                    "label_filename": out_labelname,
-                    "labels": label_dict}
-        else:
-            return {"raw_filename": image_filename,
-                    "image_filename": new_in_path}
+        return {"raw_filename": image_filename,
+                "image_filename": dirs["image"]}
 
 
-    def _preprocess(self, image_filename, output_dir, labelling):
+    def _preprocess_for_inference(self, image_filename, output_dir):
         """Resize/crop then save the training & label images
 
         Parameters
         ----------
         image_filename : str
             Full path towards the image on the disk
-        datadir : str
+        output_dir : str
             Output path where preprocessed image must be saved
 
         Returns
@@ -150,29 +193,121 @@ def _preprocess(self, image_filename, output_dir, labelling):
         raw_img_width = raster.RasterXSize
         raw_img_height = raster.RasterYSize
         result_dicts = []
-        logger.info("Raw image size: %s, %s" % (raw_img_width, raw_img_height))
         logger.info("Image filename: %s" % image_filename)
-
-        labels = None
-        if labelling:
-            label_filename = (image_filename
-                              .replace("images", "labels")
-                              .replace(".tif", ".geojson"))
-            labels = gpd.read_file(label_filename)
-            labels = labels.loc[~labels.geometry.isna(), ["condition", "geometry"]]
-            none_mask = [lc is None for lc in labels.condition]
-            labels.loc[none_mask, "condition"] = "Complete"
+        logger.info("Raw image size: %s, %s" % (raw_img_width, raw_img_height))
 
         for x in range(0, raw_img_width, self.image_size):
             for y in range(0, raw_img_height, self.image_size):
                 tile_results = self._preprocess_tile(x, y, image_filename,
-                                                     output_dir,
-                                                     raster, labels)
+                                                     output_dir, raster)
                 result_dicts.append(tile_results)
         del raster
         return result_dicts
 
 
+    def _preprocess_for_training(self, image_filename, output_dir, nb_images):
+        """Resize/crop then save the training & label images
+
+        Parameters
+        ----------
+        image_filename : str
+            Full path towards the image on the disk
+        output_dir : str
+            Output path where preprocessed image must be saved
+
+        Returns
+        -------
+        dict
+            Key/values with the filenames and label ids
+        """
+        raster = gdal.Open(image_filename)
+        raw_img_width = raster.RasterXSize
+        raw_img_height = raster.RasterYSize
+        image_data = raster.ReadAsArray()
+        image_data = np.swapaxes(image_data, 0, 2)
+        result_dicts = []
+        logger.info("Image filename: %s" % image_filename)
+        logger.info("Raw image size: %s, %s" % (raw_img_width, raw_img_height))
+
+        label_filename = (image_filename
+                          .replace("images", "labels")
+                          .replace(".tif", ".geojson"))
+        labels = gpd.read_file(label_filename)
+        labels = labels.loc[~labels.geometry.isna(), ["condition", "geometry"]]
+        none_mask = [lc is None for lc in labels.condition]
+        labels.loc[none_mask, "condition"] = "Complete"
+
+        nb_attempts = 0
+        image_counter = 0
+        empty_image_counter = 0
+        while image_counter < nb_images and nb_attempts < 2 * nb_images:
+            # randomly pick an image
+            x = np.random.randint(0, raw_img_width - self.image_size)
+            y = np.random.randint(0, raw_img_height - self.image_size)
+
+            tile_data = image_data[x:(x+self.image_size),
+                                   y:(y+self.image_size)]
+            tile_image = Image.fromarray(tile_data)
+            raster_features = get_image_features(raster)
+            tile_items = extract_tile_items(raster_features, labels,
+                                            x, y,
+                                            self.image_size,
+                                            self.image_size,
+                                            tile_srid=32737)
+            mask = self.load_mask(tile_items, raster_features, x, y)
+            label_dict = utils.build_labels(mask,
+                                            range(self.get_nb_labels()),
+                                            "tanzania")
+            labelled_image = utils.build_image_from_config(mask, self.labels)
+            if len(tile_items) > 0:
+                tiled_results = self._serialize(
+                    tile_image, labelled_image, label_dict,
+                    image_filename, output_dir, x, y, "nw"
+                )
+                result_dicts.append(tiled_results)
+                image_counter += 1
+                tile_image_ne = tile_image.transpose(Image.FLIP_LEFT_RIGHT)
+                labelled_image_ne = labelled_image.transpose(Image.FLIP_LEFT_RIGHT)
+                tiled_results_ne = self._serialize(
+                    tile_image_ne, labelled_image_ne, label_dict,
+                    image_filename, output_dir, x, y, "ne"
+                )
+                result_dicts.append(tiled_results_ne)
+                image_counter += 1
+                tile_image_sw = tile_image.transpose(Image.FLIP_TOP_BOTTOM)
+                labelled_image_sw = labelled_image.transpose(Image.FLIP_TOP_BOTTOM)
+                tiled_results_sw = self._serialize(
+                    tile_image_sw, labelled_image_sw, label_dict,
+                    image_filename, output_dir, x, y, "sw"
+                )
+                result_dicts.append(tiled_results_sw)
+                image_counter += 1
+                tile_image_se = tile_image_sw.transpose(Image.FLIP_LEFT_RIGHT)
+                labelled_image_se = labelled_image_sw.transpose(Image.FLIP_LEFT_RIGHT)
+                tiled_results_se = self._serialize(
+                    tile_image_se, labelled_image_se, label_dict,
+                    image_filename, output_dir, x, y, "se"
+                )
+                result_dicts.append(tiled_results_se)
+                image_counter += 1
+                del tile_image_se, tile_image_sw, tile_image_ne
+                del labelled_image_se, labelled_image_sw, labelled_image_ne
+            else:
+                if empty_image_counter < 0.1 * nb_images:
+                    tiled_results = self._serialize(
+                        tile_image, labelled_image, label_dict,
+                        image_filename, output_dir, x, y, "nw"
+                        )
+                    result_dicts.append(tiled_results)
+                    image_counter += 1
+                    empty_image_counter += 1
+            nb_attempts += 1
+        del raster
+        logger.info("Generate %s images after %s attempts."
+                    % (image_counter, nb_attempts))
+        return result_dicts
+
+
     def populate(self, output_dir, input_dir, nb_images=None,
                  aggregate=False, labelling=True):
         """ Populate the dataset with images contained into `datadir` directory
@@ -195,15 +330,24 @@ class method genericity
         image_list = os.listdir(os.path.join(input_dir, "images"))
         image_list_longname = [os.path.join(input_dir, "images", l)
                                for l in image_list
-                               if not l.startswith('.')][:nb_images]
+                               if not l.startswith('.')]
+        nb_image_files = len(image_list_longname)
 
         logger.info("Getting %s images to preprocess..."
-                    % len(image_list_longname))
+                    % nb_image_files)
         logger.info(image_list_longname)
-        with Pool() as p:
-            self.image_info = p.starmap(self._preprocess,
-                                        [(x, output_dir, labelling)
-                                         for x in image_list_longname])
+        if labelling:
+            nb_tile_per_image = int(nb_images/nb_image_files)
+            with Pool(processes=3) as p:
+                self.image_info = p.starmap(self._preprocess_for_training,
+                                            [(x, output_dir, nb_tile_per_image)
+                                             for x in image_list_longname])
+        else:
+            with Pool(processes=3) as p:
+                self.image_info = p.starmap(self._preprocess_for_inference,
+                                            [(x, output_dir)
+                                             for x in image_list_longname])
+
         self.image_info = [item for sublist in self.image_info
                            for item in sublist]
         logger.info("Saved %s images in the preprocessed dataset."
@@ -242,36 +386,45 @@ def load_mask(self, buildings, raster_features, min_x, min_y):
         if buildings.shape[0] == 0:
             return mask
         for idx, row in buildings.iterrows():
-            points = self.extract_points_from_polygon(row["geometry"],
-                                                      raster_features)
-            points[:, 0] -= min_x
-            points[:, 1] -= min_y
+            points = extract_points_from_polygon(row["geometry"],
+                                                 raster_features,
+                                                 min_x, min_y)
             label_id = [label["id"] for label in self.labels
                         if label["name"] == row["condition"].lower()][0]
             mask = cv2.fillPoly(mask, [points], label_id)
         return mask
 
 
-    def extract_points_from_polygon(self, p, features):
-        """Extract points from a polygon
+def extract_points_from_polygon(p, features, min_x, min_y):
+    """Extract points from a polygon
 
-        Parameters
-        ----------
-        p : shapely.geometry.Polygon
-            Polygon to detail
-        features : dict
-            Geographical features associated to the image
-        Returns
-        -------
-        np.array
-            Polygon vertices
+    Parameters
+    ----------
+    p : shapely.geometry.Polygon
+        Polygon to detail
+    features : dict
+        Geographical features associated to the image
+    min_x : int
+        Minimal x-coordinate (west)
+    min_y : int
+        Minimal y-coordinate (north)
+    Returns
+    -------
+    np.array
+        Polygon vertices
 
-        """
-        raw_xs, raw_ys = p.exterior.xy
-        xs = get_x_pixel(raw_xs, features["east"], features["west"], features["width"])
-        ys = get_y_pixel(raw_ys, features["south"], features["north"], features["height"])
-        points = np.array([[x, y] for x, y in zip(xs, ys)], dtype=np.int32)
-        return points
+    """
+    raw_xs, raw_ys = p.exterior.xy
+    xs = get_x_pixel(
+        raw_xs, features["east"], features["west"], features["width"]
+    )
+    ys = get_y_pixel(
+        raw_ys, features["south"], features["north"], features["height"]
+    )
+    points = np.array([[y, x] for x, y in zip(xs, ys)], dtype=np.int32)
+    points[:, 0] -= min_y
+    points[:, 1] -= min_x
+    return points
 
 
 def get_x_pixel(coord, east, west, width):

deeposlandia/paramoptim.py

@@ -292,8 +292,7 @@ def run_model(train_generator, validation_generator, dl_model, output_folder,
         save_weights_only=False,
         mode='auto', period=1)
     terminate_on_nan = callbacks.TerminateOnNaN()
-    earlystop = callbacks.EarlyStopping(monitor='val_acc',
-                                        min_delta=0.001,
+    earlystop = callbacks.EarlyStopping(monitor='val_loss',
                                         patience=10,
                                         verbose=1,
                                         mode='max')

deeposlandia/train.py

@@ -241,8 +241,7 @@ def add_training_arguments(parser):
         save_weights_only=False,
         mode='auto', period=1)
     terminate_on_nan = callbacks.TerminateOnNaN()
-    earlystop = callbacks.EarlyStopping(monitor='val_acc',
-                                        min_delta=0.001,
+    earlystop = callbacks.EarlyStopping(monitor='val_loss',
                                         patience=10,
                                         verbose=1,
                                         mode='max')

deeposlandia/webapp/main.py

@@ -112,7 +112,7 @@ def recover_image_info(dataset, filename):
     elif dataset == "aerial":
         size_aggregation = "250_full"
     elif dataset == "tanzania":
-        size_aggregation = "384_full"
+        size_aggregation = "512_full"
     elif dataset == "shapes":
         size_aggregation = "64_full"
     else: