Add some loss weighting and provisioning for early stopping

supirfactor_dynamical/_io/_args.py

@@ -5,6 +5,7 @@
     'time_dependent_decay',
     'decay_k',
     'decay_epoch_delay',
+    'decay_loss_weight',
     'n_genes',
     'hidden_layer_width',
     'n_peaks',

supirfactor_dynamical/models/biophysical_model.py

@@ -41,6 +41,7 @@ class SupirFactorBiophysical(
     separately_optimize_decay_model = False
     decay_epoch_delay = 0
     decay_k = 20
+    decay_loss_weight = None
 
     @property
     def has_decay(self):
@@ -52,6 +53,7 @@ def __init__(
         decay_model=None,
         decay_epoch_delay=None,
         decay_k=20,
+        decay_loss_weight=None,
         separately_optimize_decay_model=False,
         use_prior_weights=False,
         input_dropout_rate=0.5,
@@ -104,6 +106,7 @@ def __init__(
 
         self.decay_epoch_delay = decay_epoch_delay
         self.separately_optimize_decay_model = separately_optimize_decay_model
+        self.decay_loss_weight = decay_loss_weight
 
         if decay_model is False:
             output_activation = None
@@ -433,32 +436,17 @@ def _training_step(
 
             decay_loss = 0.
 
-        loss = self._training_step_joint(
-                epoch_num,
-                train_x,
-                optimizer[2] if self._decay_optimize(epoch_num) else
-                optimizer[0],
-                loss_function
-            )
-
-        return loss, decay_loss
-
-    def _training_step_joint(
-        self,
-        epoch_num,
-        train_x,
-        optimizer,
-        loss_function
-    ):
-
-        return super()._training_step(
+        loss = super()._training_step(
             epoch_num,
             train_x,
-            optimizer,
+            optimizer[2] if self._decay_optimize(epoch_num) else
+            optimizer[0],
             loss_function,
             input_x=self._slice_data_and_forward(train_x)
         )
 
+        return loss, decay_loss
+
     def _training_step_decay(
         self,
         epoch_num,
@@ -494,7 +482,8 @@ def _training_step_decay(
             optimizer,
             loss_function,
             input_x=neg,
-            target_x=_compare_x
+            target_x=_compare_x,
+            loss_weight=self.decay_loss_weight
         )
 
     def _calculate_all_losses(

supirfactor_dynamical/training/train_decoders.py

@@ -191,6 +191,10 @@ def train_decoder_submodels(
         pbar.set_description(f"[{epoch_num} n={np.sum(_batch_n)}]")
 
         if post_epoch_hook is not None:
-            post_epoch_hook(model_ref)
+
+            # If the hook returns True
+            # it's early stopping time
+            if post_epoch_hook(model_ref) is True:
+                break
 
     return model

supirfactor_dynamical/training/train_simple_decoders.py

@@ -205,6 +205,10 @@ def train_simple_multidecoder(
         pbar.set_description(f"[{epoch_num} n={np.sum(_batch_n)}]")
 
         if post_epoch_hook is not None:
-            post_epoch_hook(model_ref)
+
+            # If the hook returns True
+            # it's early stopping time
+            if post_epoch_hook(model_ref) is True:
+                break
 
     return model

supirfactor_dynamical/training/train_simple_models.py

@@ -139,6 +139,10 @@ def train_simple_model(
         pbar.set_description(f"[{epoch_num} n={np.sum(_batch_n)}]")
 
         if post_epoch_hook is not None:
-            post_epoch_hook(model_ref)
+
+            # If the hook returns True
+            # it's early stopping time
+            if post_epoch_hook(model_ref) is True:
+                break
 
     return model

supirfactor_dynamical/training/train_standard_loop.py

@@ -132,7 +132,11 @@ def train_model(
         pbar.set_description(f"[{epoch_num} n={np.sum(_batch_n)}]")
 
         if post_epoch_hook is not None:
-            post_epoch_hook(model_ref)
+
+            # If the hook returns True
+            # it's early stopping time
+            if post_epoch_hook(model_ref) is True:
+                break
 
     model_ref.eval()