Updated Sampler

Author: parthnatekar

BioExp/clusters/concept.py

@@ -170,7 +170,6 @@ def flow_based_identifier(self, concept_info,
         exp_features = layers.Conv2D(1,1, name='Expectation')(features)
         model = Model(inputs = self.model.input, outputs=exp_features)
 
-  
         # newmodel = Sequential()
         # newmodel.add(model)
         # newmodel.add()
@@ -180,22 +179,21 @@ def flow_based_identifier(self, concept_info,
         #    newmodel.layers[ii].set_weights(self.model.layers[ii].get_weights())
         model.layers[-1].set_weights((np.ones((1, 1, len(total_filters), 1)), np.ones(1)))
 
-
         grad = singlelayercam(model, test_img, 
                         nclasses = 1,
                         name  = concept_info['concept_name'], 
                         st_layer_idx = -1, 
-                        end_layer_idx = 1,
+                        end_layer_idx = -3,
                         threshold = 0.5)
 
         if base_grad == True:
             print ("[INFO: BioExp Concept Identification] Base Concept in layer {}".format(layer_name))
         else:
             print ("[INFO: BioExp Concept Identification] Identified Concept {} in layer {}".format(concept_info['concept_name'], layer_name))
 
-        del model, newmodel
+        del model
 
-        return grad[0]
+        return grad
 
     def flow_based_identifier_no_conv(self, concept_info, save_path, test_img, base_grad=False):
 
@@ -226,7 +224,7 @@ def flow_based_identifier_no_conv(self, concept_info, save_path, test_img, base_
                         save_path = save_path, 
                         name  = concept_info['concept_name'], 
                         st_layer_idx = -1, 
-                        end_layer_idx = -2,
+                        end_layer_idx = -3,
                         threshold = 0.5)
         print ("[INFO: BioExp Concept Identification] Identified Concept {} in layer {}".format(concept_info['concept_name'], layer_name))
 
@@ -237,12 +235,13 @@ def flow_based_identifier_no_conv(self, concept_info, save_path, test_img, base_
 
     def _gaussian_sampler_(self, data, size, ax=-1):
         shape = np.mean(data, ax).shape + (size,)
-        return lambda: np.std(data, -1)[..., None] * np.random.randn(*list(shape)) + np.mean(data, -1)[..., None] # np.random.normal(loc=np.mean(data, axis=ax), scale=np.std(data, axis=ax), size=size)
+        return lambda: np.std(data, -1)[..., None] * np.random.randn(*list(shape)) + np.mean(data, -1)[..., None] 
+        # return lambda: np.random.normal(np.mean(data, -1)[..., None], np.std(data, -1)[..., None], size = size)
 
     def _uniform_sampler_(self, data, size, ax=-1):
         shape = np.mean(data, ax).shape + (size,)
-        print(np.quantile(data, 0.1, axis=-1)[..., None].shape)
-        return lambda: np.random.uniform(np.quantile(data, 0, axis=-1)[..., None], np.quantile(data, 1, axis=-1)[..., None], size = size)
+        return lambda: np.quantile(data, 0.1, axis=-1)[..., None] * np.random.rand(*list(shape)) + np.quantile(data, 0.9, axis=-1)[..., None]
+        # return lambda: np.random.uniform(np.quantile(data, 0.1, axis=-1)[..., None], np.quantile(data, 0.9, axis=-1)[..., None], size = size)
 
     def concept_distribution(self, concept_info):
         """
@@ -329,25 +328,24 @@ def concept_robustness(self, concept_info,
             except: pass
 
             self.model.layers[node_idx].set_weights(occluded_weights)
-            model = Model(inputs = self.model.input, outputs=self.model.get_layer(concept_info['layer_name']).output)
-  
-            newmodel = Sequential()
-            newmodel.add(model)
-            newmodel.add(layers.Conv2D(1,1))
+            features = self.model.get_layer(concept_info['layer_name']).output
+            exp_features = layers.Conv2D(1,1, name='Expectation')(features)
+            model = Model(inputs = self.model.input, outputs=exp_features)
 
             # for ii in range(len(self.model.layers)):
             #    newmodel.layers[ii].set_weights(self.model.layers[ii].get_weights())
-            newmodel.layers[-1].set_weights((np.ones((1, 1, len(total_filters), 1)), np.ones(1)))
+            model.layers[-1].set_weights((np.ones((1, 1, len(total_filters), 1)), np.ones(1)))
 
-            nclass_grad = singlelayercam(newmodel, test_img, 
+            nclass_grad = singlelayercam(model, test_img, 
                         nclasses = 1, 
                         name  = concept_info['concept_name'], 
                         st_layer_idx = -1, 
-                        end_layer_idx = -2,
+                        end_layer_idx = -3,
                         threshold = 0.5)
-            gradlist.append(nclass_grad[0])
 
-            del model, newmodel
+            gradlist.append(nclass_grad)
+
+            del model
 
         # try:
         #     del bias_sampler

BioExp/spatial/flow.py

@@ -18,6 +18,7 @@ def singlelayercam(model, img,
         nclasses = 2,
         save_path = None,
         name = None,
+        filter_indices=[0],
         end_layer_idx = 3,
         st_layer_idx = -1,
         threshold = 0.5,
@@ -34,11 +35,11 @@ def singlelayercam(model, img,
         plt.figure(figsize=(10*nclasses, 10))
         gs = gridspec.GridSpec(1, nclasses)
         gs.update(wspace=0.025, hspace=0.05)
-    
+
     nclass_grad = []
-    for i  in range(nclasses):
+    for i  in range(len(filter_indices)):
         grads_ = visualize_cam(model, st_layer_idx, filter_indices=i, penultimate_layer_idx = end_layer_idx,  
-                    seed_input = img[None, ...], backprop_modifier = modifier)
+                    seed_input = img[None, ...], backprop_modifier = None)
         if save_path is not None:
             ax = plt.subplot(gs[i])
             im = ax.imshow(np.squeeze(img), vmin=0, vmax=1)
@@ -49,14 +50,17 @@ def singlelayercam(model, img,
             ax.tick_params(bottom='off', top='off', labelbottom='off' )
 
         nclass_grad.append(grads_)
-        
+
+    if len(np.array(nclass_grad).shape) == 3:
+        nclass_grad = np.mean(np.array(nclass_grad), axis=0)  
+
     if save_path is not None:
         divider = make_axes_locatable(ax)
         cax = divider.append_axes("right", size="5%", pad=0.2)
         # cb = plt.colorbar(im, ax=ax, cax=cax )
         os.makedirs(save_path, exist_ok = True)
         plt.savefig(os.path.join(save_path, name +'.png'), bbox_inches='tight')
-    return np.array(nclass_grad)
+    return np.squeeze(np.array(nclass_grad))
 
 
 def cam(model, img, gt=None,