Added setting for LR reduction on plateau

pagesegmentation/lib/network.py

@@ -128,7 +128,7 @@ class Network:
 
     def train_dataset(self, train_data: Dataset, test__data: Dataset, output, epochs: int = 100,
                       early_stopping: bool = True, early_stopping_interval: int = 5, tensorboardlogs: bool = True,
-                      augmentation: bool = False,):
+                      augmentation: bool = False, reduce_lr_on_plateu=False):
         callbacks = []
         train_gen = self.create_dataset_inputs(train_data, augmentation)
         test_gen = self.create_dataset_inputs(test__data, data_augmentation=False)
@@ -173,6 +173,14 @@ class Network:
                                                              write_images=False)
                 callbacks.append(diagnose_cb)
                 callbacks.append(tensorboard)
+            if reduce_lr_on_plateu:
+                redurce_lr_plateau = tf.keras.callbacks.ReduceLROnPlateau(
+                    monitor='val_loss',
+                    factor=0.5,
+                    patience=early_stopping_interval / 2,
+                    min_lr=0.000001,
+                    verbose=1)
+                callbacks.append(redurce_lr_plateau)
 
         fg = self.model.fit(train_gen,
                             epochs=epochs,

pagesegmentation/lib/trainer.py

@@ -25,6 +25,7 @@ class TrainSettings(NamedTuple):
     compute_baseline: bool = False
     foreground_masks: bool = False
     tensorboard: bool = False
+    reduce_lr_on_plateu = True
 
 
 class Trainer:
@@ -56,7 +57,8 @@ class Trainer:
                                      True if self.settings.early_stopping_max_l_rate_drops != 0 else False,
                                      early_stopping_interval=self.settings.early_stopping_max_l_rate_drops,
                                      tensorboardlogs=self.settings.tensorboard,
-                                     augmentation=self.settings.data_augmentation)
+                                     augmentation=self.settings.data_augmentation, reduce_lr_on_plateu=
+                                     self.settings.reduce_lr_on_plateu)
 
     def eval(self) -> None:
         if len(self.settings.evaluation_data) > 0:
@@ -68,15 +70,15 @@ class Trainer:
 if __name__ == "__main__":
     from pagesegmentation.lib.dataset import DatasetLoader
     from pagesegmentation.scripts.generate_image_map import load_image_map_from_file
-    image_map = load_image_map_from_file('/home/alexander/Bilder/test_datenset/map.json/image_map.json')
+    image_map = load_image_map_from_file('/home/alexander/Bilder/datenset2/image_map.json')
     dataset_loader = DatasetLoader(6, color_map=image_map)
     print(dataset_loader.color_map)
     train_data = dataset_loader.load_data_from_json(
-        ['/home/alexander/Bilder/test_datenset/t.json'], "train")
+        ['/home/alexander/Bilder/datenset2/t.json'], "train")
     test_data = dataset_loader.load_data_from_json(
-        ['/home/alexander/Bilder/test_datenset/t.json'], "test")
+        ['/home/alexander/Bilder/datenset2/t.json'], "test")
     eval_data = dataset_loader.load_data_from_json(
-        ['/home/alexander/Bilder/test_datenset/t.json'], "eval")
+        ['/home/alexander/Bilder/datenset2/t.json'], "eval")
     settings = TrainSettings(
         n_iter=100,
         n_classes=len(dataset_loader.color_map),
@@ -84,14 +86,14 @@ if __name__ == "__main__":
         train_data=train_data,
         validation_data=test_data,
         display=10,
-        output='/home/alexander/Bilder/test_datenset/',
+        output='/home/alexander/Bilder/datenset2/',#'/home/alexander/Bilder/test_datenset/', #'/home/alexander/PycharmProjects/PageContent/pagecontent/demo/'
         threads=8,
         foreground_masks=False,
         data_augmentation=True,
         tensorboard=True,
         n_architecture='mobile_net',
         early_stopping_max_l_rate_drops=5,
-        load='/home/alexander/Bilder/test_datenset/best_model.hdf5'
+        load=None#'/home/alexander/Bilder/test_datenset/best_model.hdf5'
     )
     trainer = Trainer(settings)
     trainer.train()

pagesegmentation/scripts/train.py

@@ -38,12 +38,12 @@ def main():
                         help="Data used for early stopping"
                         )
     parser.add_argument("--eval", type=str, nargs="*", default=[])
-    parser.add_argument("--display", type=int, default=100,
-                        help="Display training progress each display iterations.")
     parser.add_argument("--foreground_masks", default=False, action="store_true",
                         help="keep only mask parts that are foreground in binary image")
     parser.add_argument("--tensorboard", type=str2bool, default=False,
-                        help="Generate tenlogs for use in tensorboard")
+                        help="Generate tensorboard logs")
+    parser.add_argument("--reduce_lr_on_plateu", type=str2bool, default=True,
+                        help="Reducing LR when on plateau")
     parser.add_argument("--color_map", type=str, required=True,
                         help="color_map to load")
     args = parser.parse_args()
@@ -83,6 +83,7 @@ def main():
         foreground_masks=args.foreground_masks,
         data_augmentation=args.data_augmentation,
         tensorboard=args.tensorboard,
+        reduce_lr_on_plateu=args.reduce_lr_on_plateu,
     )
     trainer = Trainer(settings)
     trainer.train()