show more xai output

app.py

@@ -121,11 +121,14 @@ def explain_image(input_image,model_name):
     else:
         size = 600
     #saliency, integrated, smoothgrad,
-    rise,avg = explain(model,input_image,size = size, n_classes=n_classes)
+    exp_list = explain(model,input_image,size = size, n_classes=n_classes)
     #original =  saliency + integrated + smoothgrad 
     print('done')
-    rise1,rise2,rise3,rise4,rise5,avg = rise[0],rise[1],rise[2],rise[3],rise[4],avg[0]
-    return rise1,rise2,rise3,rise4,rise5,avg
+    sobol1,sobol2,sobol3,sobol4,sobol5 = exp_list[0],exp_list[1],exp_list[2],exp_list[3],exp_list[4]
+    rise1,rise2,rise3,rise4,rise5 = exp_list[5],exp_list[6],exp_list[7],exp_list[8],exp_list[9]
+    hsic1,hsic2,hsic3,hsic4,hsic5 = exp_list[10],exp_list[11],exp_list[12],exp_list[13],exp_list[14]
+    saliency1,saliency2,saliency3,saliency4,saliency5 = exp_list[15],exp_list[16],exp_list[17],exp_list[18],exp_list[19]
+    return sobol1,sobol2,sobol3,sobol4,sobol5,rise1,rise2,rise3,rise4,rise5,hsic1,hsic2,hsic3,hsic4,hsic5,saliency1,saliency2,saliency3,saliency4,saliency5
     
 #minimalist theme
 with gr.Blocks(theme='sudeepshouche/minimalist') as demo:
@@ -173,18 +176,42 @@ with gr.Blocks(theme='sudeepshouche/minimalist') as demo:
         
         with gr.Accordion("Explanations "):
             gr.Markdown("Computing Explanations from the model")
-            with gr.Row():
+            with gr.Column():
+                with gr.Row():
+
                 #original_input = gr.Image(label="Original Frame")
                 #saliency  = gr.Image(label="saliency")
                 #gradcam = gr.Image(label='integraged gradients')
                 #guided_gradcam = gr.Image(label='gradcam')
                 #guided_backprop = gr.Image(label='guided backprop')
-                rise1 = gr.Image(label = 'Rise1')
-                rise2 = gr.Image(label = 'Rise2')
-                rise3 = gr.Image(label = 'Rise3')
-                rise4 = gr.Image(label = 'Rise4')
-                rise5 = gr.Image(label = 'Rise5')
-                avg = gr.Image(label = 'Avg')
+                    sobol1 = gr.Image(label = 'Sobol1')
+                    sobol2= gr.Image(label = 'Sobol2')
+                    sobol3= gr.Image(label = 'Sobol3')
+                    sobol4= gr.Image(label = 'Sobol4')
+                    sobol5= gr.Image(label = 'Sobol5')
+                    
+                with gr.Row():
+                    rise1 = gr.Image(label = 'Rise1')
+                    rise2 = gr.Image(label = 'Rise2')
+                    rise3 = gr.Image(label = 'Rise3')
+                    rise4 = gr.Image(label = 'Rise4')
+                    rise5 = gr.Image(label = 'Rise5')
+                
+                with gr.Row():
+                    hsic1 = gr.Image(label = 'HSIC1')
+                    hsic2 = gr.Image(label = 'HSIC2')
+                    hsic3 = gr.Image(label = 'HSIC3')
+                    hsic4 = gr.Image(label = 'HSIC4')
+                    hsic5 = gr.Image(label = 'HSIC5')
+
+                with gr.Row():
+                    saliency1 = gr.Image(label = 'Saliency1')
+                    saliency2 = gr.Image(label = 'Saliency2')
+                    saliency3 = gr.Image(label = 'Saliency3')
+                    saliency4 = gr.Image(label = 'Saliency4')
+                    saliency5 = gr.Image(label = 'Saliency5')
+
+
             generate_explanations = gr.Button("Generate Explanations")
                 
         # with gr.Accordion('Closest Images'):
@@ -217,7 +244,7 @@ with gr.Blocks(theme='sudeepshouche/minimalist') as demo:
             
         #segment_button.click(segment_image, inputs=input_image, outputs=segmented_image)
         classify_image_button.click(classify_image, inputs=[input_image,model_name], outputs=class_predicted)
-        generate_explanations.click(explain_image, inputs=[input_image,model_name], outputs=[rise1,rise2,rise3,rise4,rise5,avg]) #
+        generate_explanations.click(explain_image, inputs=[input_image,model_name], outputs=[sobol1,sobol2,sobol3,sobol4,sobol5,rise1,rise2,rise3,rise4,rise5,hsic1,hsic2,hsic3,hsic4,hsic5,saliency1,saliency2,saliency3,saliency4,saliency5]) #
         #find_closest_btn.click(find_closest, inputs=[input_image,model_name], outputs=[label_closest_image_0,label_closest_image_1,label_closest_image_2,label_closest_image_3,label_closest_image_4,closest_image_0,closest_image_1,closest_image_2,closest_image_3,closest_image_4])
         def update_outputs(input_image,model_name):
             labels, images = find_closest(input_image,model_name)

closest_sample.py

@@ -35,7 +35,6 @@ def pca_distance(pca,sample,embedding):
     s = pca.transform(sample.reshape(1,-1))
     all = pca.transform(embedding[:,-1])
     distances = np.linalg.norm(all - s, axis=1)
-    #print(distances)
     return np.argsort(distances)[:5]
 
 def return_paths(argsorted,files):

explanations.py

@@ -2,8 +2,8 @@ import xplique
 import tensorflow as tf
 from xplique.attributions import (Saliency, GradientInput, IntegratedGradients, SmoothGrad, VarGrad,
                                   SquareGrad, GradCAM, Occlusion, Rise, GuidedBackprop,
-                                  GradCAMPP, Lime, KernelShap)
-
+                                  GradCAMPP, Lime, KernelShap,SobolAttributionMethod,HsicAttributionMethod)
+from xplique.attributions.global_sensitivity_analysis import LatinHypercube
 import numpy as np
 import matplotlib.pyplot as plt
 from inference_resnet import inference_resnet_finer, preprocess, _clever_crop
@@ -50,16 +50,19 @@ def explain(model, input_image,size=600, n_classes=171) :
     class_model = tf.keras.Model(model.input, model.output[1]) 
     
     explainers = [
-             #Saliency(class_model),
+        #Sobol, RISE, HSIC, Saliency
              #IntegratedGradients(class_model, steps=50, batch_size=BATCH_SIZE),
              #SmoothGrad(class_model, nb_samples=50, batch_size=BATCH_SIZE),
              #GradCAM(class_model),
-             Rise(class_model,nb_samples = 50, batch_size = BATCH_SIZE,grid_size=15,
-                 preservation_probability=0.5)
+             SobolAttributionMethod(class_model, grid_size=8, nb_design=32),
+             Rise(class_model,nb_samples = 5000, batch_size = BATCH_SIZE,grid_size=15,
+                 preservation_probability=0.5),
+             HsicAttributionMethod(class_model, 
+                                  grid_size=7, nb_design=1500,
+                                  sampler = LatinHypercube(binary=True)),
+             Saliency(class_model),
              #
     ]
-    explainer = Rise(class_model,nb_samples = 50, batch_size = BATCH_SIZE,grid_size=15,
-                 preservation_probability=0.5)
   
     cropped,repetitions = _clever_crop(input_image,(size,size))
     size_repetitions = int(size//(repetitions.numpy()+1))
@@ -70,30 +73,32 @@ def explain(model, input_image,size=600, n_classes=171) :
     #print(top_5_indices)
     X = np.expand_dims(X, 0)
     explanations = []
-    for i,Y in enumerate(top_5_indices):
-        Y = tf.one_hot([Y], n_classes)
-        print(f'{i}/{len(top_5_indices)}')
-        phi = np.abs(explainer(X, Y))[0]
-        if len(phi.shape) == 3:
-            phi = np.mean(phi, -1)
-        show(X[0][:,size_repetitions:2*size_repetitions,:])
-        show(phi[:,size_repetitions:2*size_repetitions], p=1, alpha=0.4)
-        plt.savefig(f'phi_{i}.png')
-        explanations.append(f'phi_{i}.png')
-    avg=[]
-    for i,Y in enumerate(top_5_indices):
-        Y = tf.one_hot([Y], n_classes)
-        print(f'{i}/{len(top_5_indices)}')
-        phi = np.abs(explainer(X, Y))[0]
-        if len(phi.shape) == 3:
-            phi = np.mean(phi, -1)
-        show(X[0][:,size_repetitions:2*size_repetitions,:])
-    show(phi[:,size_repetitions:2*size_repetitions], p=1, alpha=0.4)
-    plt.savefig(f'phi_6.png')
-    avg.append(f'phi_6.png')
+    for e,explainer in enumerate(explainers):
+        print(f'{e}/{len(explainers)}')
+        for i,Y in enumerate(top_5_indices):
+            Y = tf.one_hot([Y], n_classes)
+            print(f'{i}/{len(top_5_indices)}')
+            phi = np.abs(explainer(X, Y))[0]
+            if len(phi.shape) == 3:
+                phi = np.mean(phi, -1)
+            show(X[0][:,size_repetitions:2*size_repetitions,:])
+            show(phi[:,size_repetitions:2*size_repetitions], p=1, alpha=0.4)
+            plt.savefig(f'phi_{e}{i}.png')
+            explanations.append(f'phi_{e}{i}.png')
+    # avg=[]
+    # for i,Y in enumerate(top_5_indices):
+    #     Y = tf.one_hot([Y], n_classes)
+    #     print(f'{i}/{len(top_5_indices)}')
+    #     phi = np.abs(explainer(X, Y))[0]
+    #     if len(phi.shape) == 3:
+    #         phi = np.mean(phi, -1)
+    #     show(X[0][:,size_repetitions:2*size_repetitions,:])
+    # show(phi[:,size_repetitions:2*size_repetitions], p=1, alpha=0.4)
+    # plt.savefig(f'phi_6.png')
+    # avg.append(f'phi_6.png')
 
     print('Done')
     if len(explanations)==1:
         explanations = explanations[0]
-    
-    return explanations,avg
+    # return explanations,avg
+    return explanations