allow multi model with different kwargs per target

pandas_ml_utils/model/models.py

@@ -6,7 +6,7 @@
 import tempfile
 import uuid
 from copy import deepcopy
-from typing import List, Callable, TYPE_CHECKING, Tuple
+from typing import List, Callable, TYPE_CHECKING, Tuple, Dict, Any
 
 import dill as pickle
 import numpy as np
@@ -257,8 +257,11 @@ def __init__(self,
             for i in range(1, len(keras_model)):
                 if hasattr(keras_model[i], "__name__"):
                     self.custom_objects[keras_model[i].__name__] = keras_model[i]
+                else:
+                    raise ValueError("keras custom object must have a __name__")
 
             keras_model = keras_model[0]
+            _log.warning(f"keras model using custom objects {self.custom_objects}")
 
         # eventually compile keras model
         if not keras_model.optimizer:
@@ -378,7 +381,7 @@ def __call__(self, *args, **kwargs):
                                self.summary_provider,
                                self.epochs,
                                deepcopy(self.callbacks),
-                               **deepcopy(self.kwargs), **kwargs)
+                               **{**deepcopy(self.kwargs), **kwargs})
 
         # copy weights before return
         new_model.set_weights(self.get_weights())
@@ -390,19 +393,26 @@ class MultiModel(Model):
     def __init__(self,
                  model_provider: Model,
                  summary_provider: Callable[[pd.DataFrame], Summary] = Summary,
-                 loss_alpha: float = 0.5):
+                 loss_alpha: float = 0.5,
+                 target_kwargs: Dict[str, Any] = None):
         super().__init__(model_provider.features_and_labels, summary_provider, **model_provider.kwargs)
 
         if isinstance(model_provider, MultiModel):
             raise ValueError("Nesting Multi Models is not supported, you might use a flat structure of all your models")
 
         self.model_provider = model_provider
-        self.models = {target: model_provider() for target in self.features_and_labels.labels.keys()}
+        self.target_kwargs = target_kwargs
         self.loss_alpha = loss_alpha
 
+        if target_kwargs:
+            self.models = {target: model_provider(**kwargs) for target, kwargs in target_kwargs.items()}
+        else:
+            self.models = {target: model_provider() for target in self.features_and_labels.labels.keys()}
+
     def fit(self, x, y, x_val, y_val, df_index_train, df_index_test) -> float:
         losses = []
         pos = 0
+        # FIXME use the features and labels of each individual model
         for target, labels in self.features_and_labels.labels.items():
             index = range(pos, pos + len(labels))
             target_y = y[:,index]
@@ -438,10 +448,10 @@ def predict_as_column_matrix(target):
                               axis=1)
 
     def __call__(self, *args, **kwargs):
-        new_multi_model = MultiModel(self.model_provider, self.summary_provider, self.loss_alpha, **self.kwargs)
+        new_multi_model = MultiModel(self.model_provider, self.summary_provider, self.loss_alpha, self.target_kwargs)
 
         if kwargs:
-            new_multi_model.models = {target: self.model_provider(**kwargs) for target in self.features_and_labels.get_goals().keys()}
+            raise ValueError("kwargs on cloning multi model ist currently not supported!")
 
         return new_multi_model