debug

app.py

@@ -80,16 +80,14 @@ def generate_imgs(x: torch.Tensor,
     print(f"[Before inference] CUDA max allocated: {torch.cuda.max_memory_allocated() / 1024**2:.2f} MB")
     print(f"[Before inference] CUDA max reserved: {torch.cuda.max_memory_reserved() / 1024**2:.2f} MB")
 
-    if hasattr(physics.physics, 'imsize'):
-        physics.physics.imsize = x.shape[1:]
-    elif hasattr(physics.physics, 'img_size'):
-        physics.physics.img_size = x.shape[1:]
-    elif hasattr(physics.physics, 'tensor_size'):
+
+    if hasattr(physics.physics, 'tensor_size'):
         physics.physics.tensor_size = x.shape[1:]
 
     if physics.physics_generator is not None:  # we only change physic params but not noise levels
         if hasattr(physics.physics_generator, 'tensor_size'):
             physics.physics_generator.tensor_size = x.shape[1:]
+            physics.generator.tensor_size = x.shape[1:]
 
     ### Compute y
     with torch.no_grad():

factories.py

@@ -86,16 +86,16 @@ class PhysicsWithGenerator(torch.nn.Module):
         elif self.name == "Inpainting":
             sigma = 0.05
             split_ratio = 0.3
-            pixelwise = False
-            self.physics = dinv.physics.Inpainting(tensor_size=(256, 256), mask=split_ratio,
+            pixelwise = True
+            self.physics = dinv.physics.Inpainting(tensor_size=(3, 256, 256), mask=split_ratio,
                                                    noise_model=dinv.physics.GaussianNoise(sigma=sigma),
                                                    device=device_str)
             self.physics_generator = dinv.physics.generator.BernoulliSplittingMaskGenerator((3, 256, 256),
                                                                                               split_ratio=split_ratio, pixelwise=pixelwise,
-                                                                                              random_split_ratio=True, device=device_str)
+                                                                                              random_split_ratio=False, device=device_str)
             self.generator = dinv.physics.generator.BernoulliSplittingMaskGenerator((3, 256, 256),
                                                                                               split_ratio=split_ratio, pixelwise=pixelwise,
-                                                                                              random_split_ratio=True, device=device_str)
+                                                                                              random_split_ratio=False, device=device_str)
 
             self.saved_params = {"updatable_params": {},
                                  "updatable_params_converter": {"sigma": float},

requirements.txt

@@ -1,2 +1,2 @@
-git+https://github.com/deepinv/deepinv.git@fixes#egg=deepinv
+git+https://github.com/deepinv/deepinv.git
 timm