0.30.5

EarlyStopping refined

Author: bartzbeielstein

pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "spotpython"
-version = "0.30.4"
+version = "0.30.5"
 authors = [
   { name="T. Bartz-Beielstein", email="tbb@bartzundbartz.de" }
 ]

src/spotpython/light/trainmodel.py

@@ -185,7 +185,18 @@ def train_model(config: dict, fun_control: dict, timestamp: bool = True) -> floa
     #   This allows accessing the latest checkpoint in a deterministic manner.
     #   Default: None.
     config_id = generate_config_id_with_timestamp(config=config, timestamp=timestamp)
-    callbacks = [EarlyStopping(monitor="val_loss", patience=config["patience"], mode="min", strict=False, verbose=False)]
+    callbacks = [
+        EarlyStopping(
+            monitor="val_loss",
+            patience=config["patience"],
+            divergence_threshold=fun_control["divergence_threshold"],
+            check_finite=fun_control["check_finite"],
+            stopping_threshold=fun_control["stopping_threshold"],
+            mode="min",
+            strict=False,
+            verbose=False,
+        )
+    ]
     if not timestamp:
         # add ModelCheckpoint only if timestamp is False
         dirpath = os.path.join(fun_control["CHECKPOINT_PATH"], config_id)
@@ -500,7 +511,18 @@ def train_model_xai(config: dict, fun_control: dict, timestamp: bool = True) ->
     #   This allows accessing the latest checkpoint in a deterministic manner.
     #   Default: None.
     config_id = generate_config_id_with_timestamp(config=config, timestamp=timestamp)
-    callbacks = [EarlyStopping(monitor="val_loss", patience=config["patience"], mode="min", strict=False, verbose=False)]
+    callbacks = [
+        EarlyStopping(
+            monitor="val_loss",
+            patience=config["patience"],
+            divergence_threshold=fun_control["divergence_threshold"],
+            check_finite=fun_control["check_finite"],
+            stopping_threshold=fun_control["stopping_threshold"],
+            mode="min",
+            strict=False,
+            verbose=False,
+        )
+    ]
     if not timestamp:
         # add ModelCheckpoint only if timestamp is False
         dirpath = os.path.join(fun_control["CHECKPOINT_PATH"], config_id)

src/spotpython/spot/spot.py

@@ -1936,7 +1936,9 @@ def infill(self, x) -> float:
         else:
             return self.surrogate.predict(X)
 
-    def plot_progress(self, show=True, log_x=False, log_y=False, filename="plot.png", style=["ko", "k", "ro-"], dpi=300, tkagg=False) -> None:
+    def plot_progress(
+        self, show=True, log_x=False, log_y=False, filename="plot.png", style=["ko", "k", "ro-"], dpi=300, tkagg=False, title="Objective function value over iterations", y_label="y"
+    ) -> None:
         """Plot the progress of the hyperparameter tuning (optimization).
 
         Args:
@@ -1951,6 +1953,10 @@ def plot_progress(self, show=True, log_x=False, log_y=False, filename="plot.png"
             style (list):
                 Style of the plot. Default: ['k', 'ro-'], i.e., the initial points are plotted as a black line
                 and the subsequent points as red dots connected by a line.
+            title (str):
+                Title of the plot. Default: "Objective function value over iterations".
+            y_label (str):
+                Label for the y-axis. Default: "y".
 
         Returns:
             None
@@ -2012,6 +2018,10 @@ def plot_progress(self, show=True, log_x=False, log_y=False, filename="plot.png"
             ax.set_xscale("log")
         if log_y:
             ax.set_yscale("log")
+        # add a grid
+        ax.grid(True, which="both", linestyle="--", linewidth=0.5)
+        ax.set_ylabel(y_label)
+        ax.set_title(title)
         if filename is not None:
             pylab.savefig(filename, dpi=dpi, bbox_inches="tight")
         if show:

src/spotpython/utils/init.py

@@ -27,12 +27,14 @@ def fun_control_init(
     PREFIX=None,
     TENSORBOARD_CLEAN=False,
     accelerator="auto",
+    check_finite=True,
     collate_fn_name=None,
     converters=None,
     core_model=None,
     core_model_name=None,
     data=None,
     data_full_train=None,
+    divergence_threshold=None,
     hacky=False,  # !TODO: Documentation
     data_val=None,
     data_dir="./data",
@@ -90,6 +92,7 @@ def fun_control_init(
     shuffle_val=False,
     shuffle_test=False,
     sigma=0.0,
+    stopping_threshold=None,
     strategy="auto",
     surrogate=None,
     target_column=None,
@@ -129,6 +132,9 @@ def fun_control_init(
             The accelerator to be used by the Lighting Trainer.
             It can be either "auto", "dp", "ddp", "ddp2", "ddp_spawn", "ddp_cpu", "gpu", "tpu".
             Default is "auto".
+        check_finite (bool):
+            When set True, stops training when the monitor becomes NaN or infinite.
+            Default is True.
         collate_fn_name (str):
             The name of the collate function. Default is None.
         converters (dict):
@@ -164,6 +170,9 @@ def fun_control_init(
             Default is 1. Can be "auto" or an integer.
         design (object):
             The experimental design object. Default is None.
+        divergence_threshold (float):
+            Stop training as soon as the monitored quantity becomes worse than this threshold.
+            Default is None.
         enable_progress_bar (bool):
             Whether to enable the progress bar or not.
         eval (str):
@@ -284,6 +293,9 @@ def fun_control_init(
             Whether the test data were shuffled or not. Default is False.
         surrogate (object):
             The surrogate model object. Default is None.
+        stopping_threshold (float):
+            Stop training immediately once the monitored quantity reaches this threshold.
+            Default is None.
         strategy (str):
             The strategy to use. Default is "auto".
         target_column (str):
@@ -355,13 +367,15 @@ def fun_control_init(
                 '_L_out': 11,
                 '_L_cond': None,
                 'accelerator': "auto",
+                'check_finite': True,
                 'core_model': None,
                 'core_model_name': None,
                 'data': None,
                 'data_dir': './data',
                 'db_dict_name': None,
                 'device': None,
                 'devices': "auto",
+                'divergence_threshold': None,
                 'enable_progress_bar': False,
                 'eval': None,
                 'horizon': 7,
@@ -391,6 +405,7 @@ def fun_control_init(
                 'show_batch_interval': 1000000,
                 'shuffle': None,
                 'sigma': 0.0,
+                'stopping_threshold': None,
                 'target_column': None,
                 'target_type': None,
                 'train': None,
@@ -425,6 +440,7 @@ def fun_control_init(
         "_L_cond": _L_cond,
         "_torchmetric": _torchmetric,
         "accelerator": accelerator,
+        "check_finite": check_finite,
         "collate_fn_name": collate_fn_name,
         "converters": converters,
         "core_model": core_model,
@@ -433,6 +449,7 @@ def fun_control_init(
         "data": data,
         "data_dir": data_dir,
         "data_full_train": data_full_train,
+        "divergence_threshold": divergence_threshold,
         "hacky": hacky,
         "data_module": data_module,
         "data_set": data_set,
@@ -497,6 +514,7 @@ def fun_control_init(
         "shuffle_val": shuffle_val,
         "shuffle_test": shuffle_test,
         "sigma": sigma,
+        "stopping_threshold": stopping_threshold,
         "spot_tensorboard_path": spot_tensorboard_path,
         "strategy": strategy,
         "target_column": target_column,