added hierarchical manipulation

backend/disentangle_concepts.py

@@ -51,7 +51,7 @@ def get_separation_space(type_bin, annotations, df, samples=100, method='LR', C=
         return clf.coef_ / np.linalg.norm(clf.coef_), imp_features, imp_nodes, np.round(clf.score(x_val, y_val),2)
 
 
-def regenerate_images(model, z, decision_boundary, min_epsilon=-3, max_epsilon=3, count=5, latent_space='Z'):
+def regenerate_images(model, z, decision_boundary, min_epsilon=-3, max_epsilon=3, count=5, latent_space='Z', layers=None):
     """
     The regenerate_images function takes a model, z, and decision_boundary as input.  It then
     constructs an inverse rotation/translation matrix and passes it to the generator.  The generator
@@ -69,13 +69,12 @@ def regenerate_images(model, z, decision_boundary, min_epsilon=-3, max_epsilon=3
     """
     device = torch.device('cpu')
     G = model.to(device) # type: ignore
-
+    
     if False:
         decision_boundary = z - (np.dot(z, decision_boundary.T) / np.dot(decision_boundary, decision_boundary.T)) * decision_boundary
     # Labels.
     label = torch.zeros([1, G.c_dim], device=device)
-    
-  
+
     z = torch.from_numpy(z.copy()).to(device)
     decision_boundary = torch.from_numpy(decision_boundary.copy()).to(device)
         
@@ -84,25 +83,26 @@ def regenerate_images(model, z, decision_boundary, min_epsilon=-3, max_epsilon=3
     # Generate images.
     for _, lambda_ in enumerate(lambdas):
         z_0 = z + lambda_ * decision_boundary
-        # Construct an inverse rotation/translation matrix and pass to the generator.  The
-        # generator expects this matrix as an inverse to avoid potentially failing numerical
-        # operations in the network.
-        #if hasattr(G.synthesis, 'input'):
-            #m = make_transform(translate, rotate)
-            #m = np.linalg.inv(m)
-            #G.synthesis.input.transform.copy_(torch.from_numpy(m))
         if latent_space == 'Z':
-            img = G(z_0, label, truncation_psi=0.7, noise_mode='const')
+            W_0 = G.mapping(z_0, label, truncation_psi=1)
+            W = G.mapping(z, label, truncation_psi=1)
+        else:
+            W_0 = z_0.expand((14, -1)).unsqueeze(0)
+            W = z.expand((14, -1)).unsqueeze(0)
         
+        if layers:
+            W_f = W.copy()
+            W_f[layers] = W_0[layers]
+            img = G.synthesis(W_f, noise_mode='const')
         else:
-            W = z_0.expand((14, -1)).unsqueeze(0)
-            img = G.synthesis(W, noise_mode='const')
-            
+            img = G.synthesis(W_0, noise_mode='const')
+                
         img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
         images.append(PIL.Image.fromarray(img[0].cpu().numpy(), 'RGB'))
-
+            
     return images, lambdas
 
+
 def generate_joint_effect(model, z, decision_boundaries, min_epsilon=-3, max_epsilon=3, count=5, latent_space='Z'):
     decision_boundary_joint = np.sum(decision_boundaries, axis=0)
     print(decision_boundary_joint.shape)
@@ -172,6 +172,7 @@ def get_concepts_vectors(concepts, annotations, df, samples=100, method='LR', C=
     return vectors, nodes_in_common, performances
 
 
+
 def get_verification_score(concept, decision_boundary, model, annotations, samples=100, latent_space='Z'):
     import open_clip
     import os

pages/1_Disentanglement.py

@@ -141,8 +141,14 @@ with input_col_3:
     with st.form('Variate along the disentangled concept'):
         st.write('**Set range of change**')
         chosen_epsilon_input = st.empty()
-        epsilon = chosen_epsilon_input.number_input('Epsilon:', min_value=1, step=1)
-        epsilon_button = st.form_submit_button('Choose the defined epsilon')
+        epsilon = chosen_epsilon_input.number_input('Lambda:', min_value=1, step=1)
+        st.write('**Select hierarchical levels to manipulate**')
+        layers = st.selectbox('Layers:', tuple(range(14)))
+        if len(layers) == 0:
+            layers = None
+        print(layers)
+        epsilon_button = st.form_submit_button('Choose the defined lambda and layers')
+        
 
 # ---------------------------- DISPLAY COL 2 ROW 1 ------------------------------
 
@@ -165,7 +171,7 @@ with smoothgrad_col_3:
     smooth_head_3.write(f'Base image')
 
 
-images, lambdas = regenerate_images(model, original_image_vec, separation_vector, min_epsilon=-(int(epsilon)), max_epsilon=int(epsilon), latent_space=st.session_state.space_id)
+images, lambdas = regenerate_images(model, original_image_vec, separation_vector, min_epsilon=-(int(epsilon)), max_epsilon=int(epsilon), latent_space=st.session_state.space_id, layers=layers)
 
 with smoothgrad_col_1:
     st.image(images[0])