dtypes weird stuff

backend/disentangle_concepts.py

@@ -84,14 +84,16 @@ def regenerate_images(model, z, decision_boundary, min_epsilon=-3, max_epsilon=3
     for _, lambda_ in enumerate(lambdas):
         z_0 = z + lambda_ * decision_boundary
         if latent_space == 'Z':
-            W_0 = G.mapping(z_0, label, truncation_psi=1)
-            W = G.mapping(z, label, truncation_psi=1)
+            W_0 = G.mapping(z_0, label, truncation_psi=1).to(torch.float32)
+            W = G.mapping(z, label, truncation_psi=1).to(torch.float32)
+            print(W.dtype)
         else:
-            W_0 = z_0.expand((14, -1)).unsqueeze(0)
-            W = z.expand((14, -1)).unsqueeze(0)
+            W_0 = z_0.expand((14, -1)).unsqueeze(0).to(torch.float32)
+            W = z.expand((14, -1)).unsqueeze(0).to(torch.float32)
+            print(W.dtype)
         
         if layers:
-            W_f = torch.empty_like(W).copy_(W)
+            W_f = torch.empty_like(W).copy_(W).to(torch.float32)
             W_f[:, layers, :] = W_0[:, layers, :]
             img = G.synthesis(W_f, noise_mode='const')
         else:
@@ -125,7 +127,7 @@ def generate_original_image(z, model, latent_space='Z'):
     label = torch.zeros([1, G.c_dim], device=device)
     if latent_space == 'Z':
         z = torch.from_numpy(z.copy()).to(device)
-        img = G(z, label, truncation_psi=0.7, noise_mode='const')
+        img = G(z, label, truncation_psi=1, noise_mode='const')
     else:
         W = torch.from_numpy(np.repeat(z, 14, axis=0).reshape(1, 14, z.shape[1]).copy()).to(device)
         print(W.shape)