fix training smoothing bug (#145)

src/bmi/estimators/neural/_training_log.py

@@ -22,17 +22,22 @@ def __init__(
         Args:
             max_n_steps: maximum number of training steps allowed
             early_stopping: whether early stopping is turned on
-            train_smooth_factor: TODO(Frederic, Pawel): Add description.
+            train_smooth_factor: fraction of the training history length used
+               as the smoothing window for convergence checks. E.g. when
+               `train_smooth_factor=0.1` and the current training history
+               has length 1000, averages over 100 steps will be used. Max 0.5.
             verbose: whether to print information during the training
             enable_tqdm: whether to use tqdm's progress bar during training
             history_in_additional_information: whether the generated additional
               information should contain training history (evaluated loss on
               training and test populations). We recommend keeping this flag
               turned on.
         """
+        assert train_smooth_factor <= 0.5, "train_smooth_factor can be at most 0.5"
+
         self.max_n_steps = max_n_steps
         self.early_stopping = early_stopping
-        self.train_smooth_window = int(max_n_steps * train_smooth_factor)
+        self.train_smooth_factor = train_smooth_factor
         self.verbose = verbose
 
         self._train_history_in_additional_information = train_history_in_additional_information
@@ -114,40 +119,46 @@ def detect_warnings(self):  # noqa: C901
             if self.verbose:
                 print("WARNING: Early stopping enabled but max_n_steps reached.")
 
-        # analyze training
+        # get train MI history
         train_mi = jnp.array([mi for _step, mi in self._mi_train_history])
-        w = self.train_smooth_window
-        cs = jnp.cumsum(train_mi)
-        # TODO(Pawel, Frederic): If training smooth window is too
-        #   long we will have an error that subtraction between (n,)
-        #   and (0,) arrays cannot be performed.
+        w = int(self.train_smooth_factor * len(train_mi))
+
+        # check if training long enough to compute diagnostics
+        if w < 1:
+            self._additional_information["training_too_short_for_diagnostics"] = True
+            if self.verbose:
+                print("WARNING: Training too short to compute diagnostics.")
+            return
+
+        # compute smoothed mi
+        cs = jnp.cumsum(jnp.concatenate([jnp.zeros(1), train_mi]))
         train_mi_smooth = (cs[w:] - cs[:-w]) / w
 
-        if len(train_mi_smooth) > 0:
-            train_mi_smooth_max = float(train_mi_smooth.max())
-            train_mi_smooth_fin = float(train_mi_smooth[-1])
-            if train_mi_smooth_max > 1.05 * train_mi_smooth_fin:
-                self._additional_information["max_training_mi_decreased"] = True
-                if self.verbose:
-                    print(
-                        f"WARNING: Smoothed training MI fell compared to highest value: "
-                        f"max={train_mi_smooth_max:.3f} vs "
-                        f"final={train_mi_smooth_fin:.3f}"
-                    )
-
-        w = self.train_smooth_window
-        if len(train_mi_smooth) >= w:
-            train_mi_smooth_fin = float(train_mi_smooth[-1])
-            train_mi_smooth_prv = float(train_mi_smooth[-w])
-            if train_mi_smooth_fin > 1.05 * train_mi_smooth_prv:
-                self._additional_information["training_mi_still_increasing"] = True
-                if self.verbose:
-                    print(
-                        f"WARNING: Smoothed raining MI was still "
-                        f"increasing when training stopped: "
-                        f"final={train_mi_smooth_fin:.3f} vs "
-                        f"{w} step(s) ago={train_mi_smooth_prv:.3f}"
-                    )
+        # n + 1 - w >= w + 1 since w <= int(0.5 * n)
+        assert len(train_mi_smooth) >= w + 1
+
+        train_mi_smooth_max = float(train_mi_smooth.max())
+        train_mi_smooth_fin = float(train_mi_smooth[-1])
+        if train_mi_smooth_max > 1.05 * train_mi_smooth_fin:
+            self._additional_information["max_training_mi_decreased"] = True
+            if self.verbose:
+                print(
+                    f"WARNING: Smoothed training MI fell compared to highest value: "
+                    f"max={train_mi_smooth_max:.3f} vs "
+                    f"final={train_mi_smooth_fin:.3f}"
+                )
+
+        train_mi_smooth_fin = float(train_mi_smooth[-1])
+        train_mi_smooth_prv = float(train_mi_smooth[-w])
+        if train_mi_smooth_fin > 1.05 * train_mi_smooth_prv:
+            self._additional_information["training_mi_still_increasing"] = True
+            if self.verbose:
+                print(
+                    f"WARNING: Smoothed raining MI was still "
+                    f"increasing when training stopped: "
+                    f"final={train_mi_smooth_fin:.3f} vs "
+                    f"{w} step(s) ago={train_mi_smooth_prv:.3f}"
+                )
 
     def _tqdm_init(self):
         self._tqdm = tqdm.tqdm(