enable custom objects for keras model

pandas_ml_utils/model/models.py

@@ -17,7 +17,7 @@
 
 from pandas_ml_utils.summary.summary import Summary
 from pandas_ml_utils.model.features_and_labels.features_and_labels import FeaturesAndLabels
-from pandas_ml_utils.utils.functions import suitable_kwargs
+from pandas_ml_utils.utils.functions import suitable_kwargs, call_with_suitable_kwargs
 
 _log = logging.getLogger(__name__)
 
@@ -245,8 +245,17 @@ def __init__(self,
         else:
             self.is_tensorflow = False
 
+        # create keras model
         provider_args = suitable_kwargs(keras_compiled_model_provider, **kwargs)
-        self.keras_model = self._exec_within_session(keras_compiled_model_provider, **provider_args)
+        keras_model = self._exec_within_session(keras_compiled_model_provider, **provider_args)
+
+        # eventually compile keras model
+        if not keras_model.optimizer:
+            compile_args = suitable_kwargs(keras_model.compile, **kwargs)
+            self._exec_within_session(keras_model.compile, **compile_args)
+
+        # set all members
+        self.keras_model = keras_model
         self.epochs = epochs
         self.callbacks = callbacks
         self.history = None
@@ -333,9 +342,9 @@ def __setstate__(self, state):
             with self.graph.as_default():
                 self.session = tf.Session(graph=self.graph)
                 K.set_session(self.session)
-                self.keras_model = load_model(tmp_keras_file)
+                self.keras_model = load_model(tmp_keras_file, custom_objects=self.kwargs)
         else:
-            self.keras_model = load_model(tmp_keras_file)
+            self.keras_model = load_model(tmp_keras_file, custom_objects=self.kwargs)
 
         # clean up temp file
         with contextlib.suppress(OSError):

pandas_ml_utils/utils/functions.py

@@ -51,8 +51,8 @@ def suitable_kwargs(func, **kwargs):
 
 
 def call_with_suitable_kwargs(func, **kwargs):
-    args = suitable_kwargs(kwargs)
-    return func(args)
+    args = suitable_kwargs(func, **kwargs)
+    return func(**args)
 
 
 def call_callable_dyamic_args(func, *args):

test/unit_tests/model/test__save_load.py

@@ -68,3 +68,28 @@ def test_model_with_LazyDataFrame_copy(self):
         self.assertEqual(model.kwargs["ldf"].kwargs['foo'](None), 'bar')
         self.assertEqual(model2.kwargs["ldf"].kwargs['foo'](None), 'bar')
 
+    def test_save_load_keras_custom_loss(self):
+        """given"""
+        features_and_labels = pmu.FeaturesAndLabels(["a"], ["b"])
+        name = '/tmp/pandas-ml-utils-unittest-test_model_keras_custom_loss'
+
+        def my_custom_loss(x, y):
+            import keras.backend as K
+            return K.sum(x - y)
+
+        def keras_model_provider():
+            model = Sequential()
+            model.add(Dense(1, input_dim=1, activation='sigmoid'))
+            return model
+
+        """when"""
+        # FIXME my_custom_loss should have proper name
+        fit = df.fit(pmu.KerasModel(keras_model_provider, features_and_labels, optimizer='adam', loss=my_custom_loss, verbose=0, my_custom_loss=my_custom_loss))
+        fitted_model = fit.model
+
+        fit.save_model(name)
+        restored_model = pmu.Model.load(name)
+
+        """then"""
+        pd.testing.assert_frame_equal(df.predict(fitted_model), df.predict(restored_model))
+        pd.testing.assert_frame_equal(df.backtest(fitted_model).df, df.backtest(restored_model).df)