make classifier_params a property of cnn.BaseModule

this makes it easier to override the definition of how to retrieve .parameters() that should be treated as the classifier, without having to override the entire configure_optimizers() function

opensoundscape/ml/cnn.py

@@ -295,6 +295,14 @@ def training_step(self, samples, batch_idx):
 
     # def predict_step(self, batch): #runs forward() if we don't override default
 
+    @property
+    def classifier_params(self):
+        """return the parameters of the classifier layer of the network
+
+        override this method if the classifier parameters should be retrieved in a different way
+        """
+        return self.classifier.parameters()
+
     def configure_optimizers(
         self,
         reset_optimizer=False,
@@ -337,6 +345,7 @@ def configure_optimizers(
             dictionary with keys "optimizer" and "scheduler" containing the
             optimizer and learning rate scheduler objects to use during training
         """
+
         if reset_optimizer:
             self.optimizer = None
         if restart_scheduler:
@@ -354,28 +363,26 @@ def configure_optimizers(
         if self.optimizer_params["classifier_lr"] is not None:
             # customize the learning rate of the classifier layer
             try:
-                classifier_params = list(self.classifier.parameters())
-                # for some reason, I get tensor mismatch if I check whether
-                # parameters are in list. Instead, compare the objects' ids.
+                # Cannot check `param in param_list`. Instead, compare the objects' ids.
                 # see https://discuss.pytorch.org/t/confused-by-runtimeerror-when-checking-for-parameter-in-list/211308
-                classifier_param_ids = {id(p) for p in classifier_params}
+                classifier_param_ids = {id(p) for p in self.classifier_params}
+                # remove these parameters from their current group
+                for param_group in optimizer.param_groups:
+                    param_group["params"] = [
+                        p
+                        for p in param_group["params"]
+                        if id(p) not in classifier_param_ids
+                    ]
             except Exception as e:
                 raise ValueError(
                     "Could not access self.classifier.parameters(). "
                     "Make sure self.classifier propoerty returns a torch.nn.Module object."
                 ) from e
-            # remove these parameters from their current group
-            for param_group in optimizer.param_groups:
-                param_group["params"] = [
-                    p
-                    for p in param_group["params"]
-                    if id(p) not in classifier_param_ids
-                ]
 
             # add them to a new group with custom learning rate
             optimizer.add_param_group(
                 {
-                    "params": classifier_params,
+                    "params": self.classifier_params,
                     "lr": self.optimizer_params["classifier_lr"],
                 }
             )
@@ -2303,9 +2310,6 @@ def use_resample_loss(model, train_df):
     model.loss_fn = ResampleLoss(class_frequency)
 
 
-# TODO: implement `classifier_lr` key in self.optimizer_params and use in configure_optimizers
-
-
 @register_model_cls
 class InceptionV3(SpectrogramClassifier):
     """Child of SpectrogramClassifier class for InceptionV3 architecture"""