Remove unnecessary CPU-GPU data transfers and update the examples accordingly.

Author: sypsyp97

diffct/differentiable.py

@@ -58,8 +58,8 @@ def main():
     isocenter_distance = 500.0
 
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-    image_torch = torch.tensor(phantom, device=device, requires_grad=True)
-    angles_torch = torch.tensor(angles_np, device=device)
+    image_torch = torch.tensor(phantom, device=device, dtype=torch.float32, requires_grad=True)
+    angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
 
     sinogram = FanProjectorFunction.apply(image_torch, angles_torch, num_detectors,
                                           detector_spacing, source_distance, isocenter_distance)
@@ -73,7 +73,7 @@ def main():
 
     # Apply weights before filtering
     sino_weighted = sinogram * weights
-    sinogram_filt = ramp_filter(sino_weighted).detach().requires_grad_(True).contiguous()
+    sinogram_filt = ramp_filter(sino_weighted)
 
     reconstruction = FanBackprojectorFunction.apply(sinogram_filt, angles_torch,
                                                     detector_spacing, Nx, Ny,

examples/fbp_fan.py

@@ -55,13 +55,13 @@ def main():
     detector_spacing = 1.0
 
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-    image_torch = torch.tensor(phantom, device=device, requires_grad=True)
-    angles_torch = torch.tensor(angles_np, device=device, requires_grad=False)
+    image_torch = torch.tensor(phantom, device=device, dtype=torch.float32, requires_grad=True)
+    angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
 
     sinogram = ParallelProjectorFunction.apply(image_torch, angles_torch,
                                                num_detectors, detector_spacing)
 
-    sinogram_filt = ramp_filter(sinogram).detach().requires_grad_(True).contiguous() 
+    sinogram_filt = ramp_filter(sinogram)
 
     reconstruction = ParallelBackprojectorFunction.apply(sinogram_filt, angles_torch,
                                                          detector_spacing, Nx, Ny)

examples/fbp_parallel.py

@@ -84,8 +84,8 @@ def main():
     isocenter_distance = 600.0
 
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-    phantom_torch = torch.tensor(phantom_cpu, device=device, requires_grad=True)
-    angles_torch = torch.tensor(angles_np, device=device)
+    phantom_torch = torch.tensor(phantom_cpu, device=device, dtype=torch.float32, requires_grad=True)
+    angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
 
     sinogram = ConeProjectorFunction.apply(phantom_torch, angles_torch,
                                            det_u, det_v, du, dv,
@@ -106,7 +106,7 @@ def main():
 
     # Apply weights and then filter
     sino_weighted = sinogram * weights
-    sinogram_filt = ramp_filter_3d(sino_weighted).detach().requires_grad_(True).contiguous()
+    sinogram_filt = ramp_filter_3d(sino_weighted)
 
     reconstruction = ConeBackprojectorFunction.apply(sinogram_filt, angles_torch, Nx, Ny, Nz,
                                                     du, dv, source_distance, isocenter_distance)

examples/fdk_cone.py

@@ -123,20 +123,21 @@ def main():
     isocenter_distance = 400.0
 
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-    phantom_torch = torch.tensor(phantom_cpu, device=device)
+    phantom_torch = torch.tensor(phantom_cpu, device=device, dtype=torch.float32)
 
     # Generate the "real" sinogram
-    real_sinogram = ConeProjectorFunction.apply(phantom_torch, angles_np,
+    angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
+    real_sinogram = ConeProjectorFunction.apply(phantom_torch, angles_torch,
                                                det_u, det_v, du, dv,
                                                source_distance, isocenter_distance)
 
     pipeline_instance = Pipeline(lr=1e-1, 
-                                 volume_shape=(Nz,Ny,Nx), 
-                                 angles=angles_np, 
-                                 det_u=det_u, det_v=det_v, 
-                                 du=du, dv=dv, 
-                                 source_distance=source_distance, 
-                                 isocenter_distance=isocenter_distance, 
+                                 volume_shape=(Nz,Ny,Nx),
+                                 angles=angles_torch,
+                                 det_u=det_u, det_v=det_v,
+                                 du=du, dv=dv,
+                                 source_distance=source_distance,
+                                 isocenter_distance=isocenter_distance,
                                  device=device, epoches=1000)
 
     ini_guess = torch.zeros_like(phantom_torch)

examples/iterative_reco_cone.py

@@ -93,13 +93,13 @@ def main():
     angles_np = np.linspace(0, 2 * math.pi, num_views, endpoint=False).astype(np.float32)
 
     num_detectors = 256
-    detector_spacing = 0.75
+    detector_spacing = 1.0
     source_distance = 600.0
     isocenter_distance = 400.0
 
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-    phantom_torch = torch.tensor(phantom_cpu, device=device)
-    angles_torch = torch.tensor(angles_np, device=device)
+    phantom_torch = torch.tensor(phantom_cpu, device=device, dtype=torch.float32)
+    angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
 
     # Generate the "real" sinogram
     real_sinogram = FanProjectorFunction.apply(phantom_torch, angles_torch,

examples/iterative_reco_fan.py

@@ -84,11 +84,11 @@ def main():
     angles_np = np.linspace(0, 2 * math.pi, num_views, endpoint=False).astype(np.float32)
 
     num_detectors = 256
-    detector_spacing = 0.75
+    detector_spacing = 0.5
 
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-    phantom_torch = torch.tensor(phantom_cpu, device=device)
-    angles_torch = torch.tensor(angles_np, device=device)
+    phantom_torch = torch.tensor(phantom_cpu, device=device, dtype=torch.float32)
+    angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
 
     # Generate the "real" sinogram
     real_sinogram = ParallelProjectorFunction.apply(phantom_torch, angles_torch,

examples/iterative_reco_parallel.py

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "diffct"
-version = "1.1.7"
+version = "1.2.0"
 description = "A CUDA-based library for computed tomography (CT) projection and reconstruction with differentiable operators"
 readme = "README.md"
 authors = [