Merge pull request #93 from kundajelab/targetlayerwarningfix

Target layer warning fix

Author: AvantiShri

deeplift.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: deeplift
-Version: 0.6.9.2
+Version: 0.6.9.3
 Summary: DeepLIFT (Deep Learning Important FeaTures)
 Home-page: https://github.com/kundajelab/deeplift
 License: UNKNOWN

deeplift/__init__.py

@@ -1 +1 @@
-__version__ = '0.6.9.2'
+__version__ = '0.6.9.3'

deeplift/conversion/kerasapi_conversion.py

@@ -119,7 +119,9 @@ def conv2d_conversion(config,
             name=("preact_" if len(converted_activation) > 0
                         else "")+name,
             kernel=config[KerasKeys.weights][0],
-            bias=config[KerasKeys.weights][1],
+            bias=(config[KerasKeys.weights][1] if 
+                  len(config[KerasKeys.weights]) > 1
+                  else np.zeros(config[KerasKeys.weights][0].shape[-1])),
             strides=config[KerasKeys.strides],
             padding=config[KerasKeys.padding].upper(),
             data_format=config[KerasKeys.data_format],
@@ -150,7 +152,9 @@ def conv1d_conversion(config,
             name=("preact_" if len(converted_activation) > 0
                         else "")+name,
             kernel=config[KerasKeys.weights][0],
-            bias=config[KerasKeys.weights][1],
+            bias=(config[KerasKeys.weights][1] if 
+                  len(config[KerasKeys.weights]) > 1
+                  else np.zeros(config[KerasKeys.weights][0].shape[-1])),
             stride=config[KerasKeys.strides],
             padding=config[KerasKeys.padding].upper(),
             conv_mxts_mode=conv_mxts_mode)] 
@@ -177,7 +181,9 @@ def dense_conversion(config,
                   name=("preact_" if len(converted_activation) > 0
                         else "")+name, 
                   kernel=config[KerasKeys.weights][0],
-                  bias=config[KerasKeys.weights][1],
+                  bias=(config[KerasKeys.weights][1] if 
+                        len(config[KerasKeys.weights]) > 1 
+                        else np.zeros(config[KerasKeys.weights][0].shape[-1])),
                   verbose=verbose,
                   dense_mxts_mode=dense_mxts_mode)]
     to_return.extend(converted_activation)

deeplift/models.py

@@ -123,6 +123,7 @@ def get_target_contribs_of_input_with_filter_ref_func(
                 **kwargs)
 
     def _set_scoring_mode_for_target_layer(self, target_layer):
+        print("TARGET LAYER SET TO "+str(target_layer.get_name()))
         if (deeplift.util.is_type(target_layer,
                                   layers.Activation)):
             raise RuntimeError("You set the target layer to an"
@@ -131,55 +132,22 @@ def _set_scoring_mode_for_target_layer(self, target_layer):
                   +" to set the target layer to the layer *before*"
                   +" the activation layer instead? (recommended for "
                   +" classification)")
-        if (len(target_layer.get_output_layers())==0):
-            scoring_mode=ScoringMode.OneAndZeros
-        else:
-            assert len(target_layer.get_output_layers())==1,\
-                   "at most one output was expected for target layer "\
-                   +str(target_layer.get_name())+" but got: "+\
-                   str(target_layer.get_output_layers())
-            final_activation_layer = target_layer.get_output_layers()[0]
-            if (deeplift.util.is_type(final_activation_layer,
-                                      layers.Activation)==False):
-                raise RuntimeError("There is a layer after your target"
-                      +" layer but it is not an activation layer"
-                      +", which seems odd...if doing regression, make"
-                      +" sure to set the target layer to the last layer")
-            deeplift.util.assert_is_type(final_activation_layer,
-                                         layers.Activation,
-                                         "final_activation_layer")
-            final_activation_type = type(final_activation_layer).__name__
-
-            if (final_activation_type == "Sigmoid"):
-                scoring_mode=ScoringMode.OneAndZeros
-            elif (final_activation_type == "Softmax"):
-                #new_W, new_b =\
-                # deeplift.util.get_mean_normalised_softmax_weights(
-                #  target_layer.W, target_layer.b)
-                    #The weights need to be mean normalised before they are
-                    #passed in because build_fwd_pass_vars() has already
-                    #been called before this function is called,
-                    #because get_output_layers() (used in this function)
-                    #is updated during the build_fwd_pass_vars()
-                    #call - that is why I can't simply mean-normalise
-                    #the weights right here :-( (It is a pain and a
-                    #recipe for bugs to rebuild the forward pass
-                    #vars after they have already been built - in
-                    #particular for a model that branches because where
-                    #the branches unify you need really want them to be
-                    #using the same symbolic variables - no use having
-                    #needlessly complicated/redundant graphs and if a node
-                    #is common to two outputs, so should its symbolic vars
-                    #TODO: I should put in a 'reset_fwd_pass' function and use
-                    #it to invalidate the _built_fwd_pass_vars cache and recompile
-                #if (np.allclose(target_layer.W, new_W)==False):
-                #    print("Consider mean-normalising softmax layer")
-                #assert np.allclose(target_layer.b, new_b),\
-                #       "Please mean-normalise weights and biases of softmax layer"
+        scoring_mode=ScoringMode.OneAndZeros
+        if (len(target_layer.get_output_layers())>0):
+            if (len(target_layer.get_output_layers())>1):
+                print("WARNING: the target layer"
+                      +str(target_layer.get_name())
+                      +" has multiple output layers"
+                      +str(target_layer.get_output_layers()))
+            else: 
+                final_activation_layer = target_layer.get_output_layers()[0]
+                if (deeplift.util.is_type(final_activation_layer,
+                                          layers.Activation)==False):
+                    print("\n\nWARNING!!! There is a layer after your target"
+                          +" layer but it is not an activation layer"
+                          +", which is unusual; double check you have set"
+                          +" the target layer correctly.\n\n")
                 scoring_mode=ScoringMode.OneAndZeros
-            else:
-                raise RuntimeError("Unsupported final_activation_type: "
-                                   +final_activation_type)
         target_layer.set_scoring_mode(scoring_mode)    
 
     def save_to_yaml_only(self, file_name):

setup.py

@@ -8,7 +8,7 @@
 Implements the methods in "Learning Important Features Through Propagating Activation Differences" by Shrikumar, Greenside & Kundaje, as well as other commonly-used methods such as gradients, guided backprop and integrated gradients. See https://github.com/kundajelab/deeplift for documentation and FAQ.
           """,
           url='https://github.com/kundajelab/deeplift',
-          version='0.6.9.2',
+          version='0.6.9.3',
           packages=['deeplift',
                     'deeplift.layers', 'deeplift.visualization',
                     'deeplift.conversion'],