diffct 1.2.3__tar.gz → 1.2.5__tar.gz

diffct-1.2.5/.github/workflows/release.yml

@@ -0,0 +1,20 @@
+name: Create Release
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  release:
+    name: Create Release
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+      - name: Create Release
+        uses: softprops/action-gh-release@v2
+        with:
+          generate_release_notes: true

{diffct-1.2.3 → diffct-1.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffct
-Version: 1.2.3
+Version: 1.2.5
 Summary: A CUDA-based library for computed tomography (CT) projection and reconstruction with differentiable operators
 Project-URL: Homepage, https://github.com/sypsyp97/diffct
 Author-email: Yipeng Sun <yipeng.sun@fau.de>

{diffct-1.2.3 → diffct-1.2.5}/diffct/differentiable.py

@@ -19,6 +19,8 @@ _TPB_3D             = (8,  8,  8)
 # Trades numerical precision for performance in ray-tracing calculations
 # Safe for CT reconstruction where slight precision loss is acceptable for speed gains
 _FASTMATH_DECORATOR = cuda.jit(cache=True, fastmath=True)
+# Disable fastmath for backward kernels to ensure gradient correctness
+_NON_FASTMATH_DECORATOR = cuda.jit(cache=True, fastmath=False) 
 _INF                = _DTYPE(np.inf)
 _EPSILON            = _DTYPE(1e-6)
 # === Device Management Utilities ===
@@ -106,36 +108,9 @@ class TorchCUDABridge:
             raise ValueError("Tensor must be on CUDA device")
         return cuda.as_cuda_array(tensor.detach())
 
-    @staticmethod
-    def cuda_array_to_tensor(cuda_array, tensor_template):
-        """Convert a Numba CUDA array to a PyTorch tensor.
-
-        Wrap a Numba CUDA DeviceNDArray as a PyTorch tensor with matching device
-        and dtype from a template tensor, sharing underlying memory.
-
-        Parameters
-        ----------
-        cuda_array : numba.cuda.cudadrv.devicearray.DeviceNDArray
-            Numba CUDA array to wrap.
-        tensor_template : torch.Tensor
-            Template tensor specifying device and dtype.
-
-        Returns
-        -------
-        torch.Tensor
-            PyTorch tensor sharing data with the CUDA array on the template's
-            device and dtype.
-
-        Examples
-        --------
-        >>> arr = cuda.device_array((10,), dtype=np.float32)
-        >>> t = torch.zeros(10, device='cuda')
-        >>> new_t = TorchCUDABridge.cuda_array_to_tensor(arr, t)
-        """
-        return torch.as_tensor(cuda_array, device=tensor_template.device, dtype=tensor_template.dtype)
 
 # === GPU-aware Trigonometric Table Generation ===
-def _trig_tables(angles, dtype=_DTYPE):
+def _trig_tables(angles, dtype=_DTYPE, device=None):
     """Compute cosine and sine tables for input angles.
 
     Precompute cosine and sine values and return as torch tensors on the
@@ -163,7 +138,7 @@ def _trig_tables(angles, dtype=_DTYPE):
     device(type='cuda', index=0)
     """
     if isinstance(angles, torch.Tensor):
-        device = angles.device
+        device = angles.device if device is None else device
         cos = torch.cos(angles).to(dtype=dtype)
         sin = torch.sin(angles).to(dtype=dtype)
         return cos.to(device), sin.to(device)
@@ -182,7 +157,10 @@ def _trig_tables(angles, dtype=_DTYPE):
         angles_cpu = torch.tensor(angles, dtype=torch_dtype)
         cos_cpu = torch.cos(angles_cpu)
         sin_cpu = torch.sin(angles_cpu)
-        return cos_cpu, sin_cpu
+        if device is not None:
+            return cos_cpu.to(device), sin_cpu.to(device)
+        else:
+            return cos_cpu, sin_cpu
 
 
 # ############################################################################
@@ -191,14 +169,14 @@ def _trig_tables(angles, dtype=_DTYPE):
 
 def _validate_3d_memory_layout(tensor, expected_order='DHW'):
     """Validate 3D tensor memory layout to prevent coordinate system inconsistencies.
-    
+
     Parameters
     ----------
     tensor : torch.Tensor
         3D tensor to validate
     expected_order : str, optional
         Expected memory order ('DHW', 'VHW', etc.). Default is 'DHW'.
-        
+
     Raises
     ------
     ValueError
@@ -206,53 +184,50 @@ def _validate_3d_memory_layout(tensor, expected_order='DHW'):
     """
     if len(tensor.shape) != 3:
         raise ValueError(f"Expected 3D tensor, got {len(tensor.shape)}D")
-    
+
     # Check if tensor is contiguous to avoid memory duplication
     if not tensor.is_contiguous():
         raise ValueError(
             "Input tensor must be contiguous. Call .contiguous() before passing to "
             "cone beam functions to avoid memory duplication and ensure correct results."
         )
-    
-    # Validate expected memory order based on stride patterns
-    strides = tensor.stride()
-    
-    # Map expected orders to stride patterns
-    order_mapping = {
-        'DHW': (0, 1, 2),  # Depth, Height, Width
-        'VHW': (0, 1, 2),  # Views, Height, Width (for sinograms)
-        'WHD': (2, 1, 0),  # Width, Height, Depth (internal WHD format)
-    }
-    
-    if expected_order not in order_mapping:
-        raise ValueError(f"Unsupported expected_order: {expected_order}")
-    
-    expected_stride_order = order_mapping[expected_order]
-    
-    # Check if actual strides match expected order
-    sorted_strides = sorted(enumerate(strides), key=lambda x: x[1], reverse=True)
-    actual_order = tuple(idx for idx, _ in sorted_strides)
-    
-    if actual_order != expected_stride_order:
-        # Create appropriate error message based on context
-        if expected_order == 'VHW':
-            actual_str = f"({tensor.shape[0]}, {tensor.shape[1]}, {tensor.shape[2]})"
-            expected_str = "(Views, Height, Width)"
-            fix_str = "ensure your sinogram has shape (num_views, det_v, det_u)"
-        elif expected_order == 'DHW':
-            actual_str = f"({tensor.shape[0]}, {tensor.shape[1]}, {tensor.shape[2]})"
-            expected_str = "(Depth, Height, Width)"
-            fix_str = "ensure your volume has shape (D, H, W)"
-        else:
-            actual_str = str(tuple(tensor.shape))
-            expected_str = expected_order
-            fix_str = "check tensor dimensions"
-            
-        raise ValueError(
-            f"Memory layout mismatch: expected {expected_str} order, "
-            f"but tensor has shape {actual_str}. Please {fix_str} and ensure "
-            f"the tensor is contiguous (.contiguous()) before passing to the function."
-        )
+
+    # Only check memory order for DHW and VHW, not for internal WHD layout
+    if expected_order in ('DHW', 'VHW'):
+        strides = tensor.stride()
+        order_mapping = {
+            'DHW': (0, 1, 2),  # Depth, Height, Width
+            'VHW': (0, 1, 2),  # Views, Height, Width (for sinograms)
+        }
+        if expected_order not in order_mapping:
+            raise ValueError(f"Unsupported expected_order: {expected_order}")
+
+        expected_stride_order = order_mapping[expected_order]
+        # Check if actual strides match expected order
+        sorted_strides = sorted(enumerate(strides), key=lambda x: x[1], reverse=True)
+        actual_order = tuple(idx for idx, _ in sorted_strides)
+
+        if actual_order != expected_stride_order:
+            # Create appropriate error message based on context
+            if expected_order == 'VHW':
+                actual_str = f"({tensor.shape[0]}, {tensor.shape[1]}, {tensor.shape[2]})"
+                expected_str = "(Views, Height, Width)"
+                fix_str = "ensure your sinogram has shape (num_views, det_v, det_u)"
+            elif expected_order == 'DHW':
+                actual_str = f"({tensor.shape[0]}, {tensor.shape[1]}, {tensor.shape[2]})"
+                expected_str = "(Depth, Height, Width)"
+                fix_str = "ensure your volume has shape (D, H, W)"
+            else:
+                actual_str = str(tuple(tensor.shape))
+                expected_str = expected_order
+                fix_str = "check tensor dimensions"
+
+            raise ValueError(
+                f"Memory layout mismatch: expected {expected_str} order, "
+                f"but tensor has shape {actual_str}. Please {fix_str} and ensure "
+                f"the tensor is contiguous (.contiguous()) before passing to the function."
+            )
+    # For 'WHD' (internal layout), skip stride check entirely
 
 
 def _grid_2d(n1, n2, tpb=_TPB_2D):
@@ -392,7 +367,7 @@ def _parallel_2d_forward_kernel(
     cos_a = d_cos[iang]  # Precomputed cosine of projection angle
     sin_a = d_sin[iang]  # Precomputed sine of projection angle
     # Normalize all physical distances to voxel units
-    u     = (idet - (n_det - 1) * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
+    u     = (idet - n_det * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
 
     # Define ray direction and starting point for parallel beam geometry
     # Ray direction is perpendicular to detector array (cos_a, sin_a)
@@ -471,8 +446,8 @@ def _parallel_2d_forward_kernel(
                 dx, dy = mid_x - ix0, mid_y - iy0  # Fractional parts: distance from base voxel center [0,1]
                 
                 # Clamp indices to stay in-bounds during interpolation
-                ix0 = min(ix0, Nx - 2)
-                iy0 = min(iy0, Ny - 2)
+                ix0 = max(0, min(ix0, Nx - 2))
+                iy0 = max(0, min(iy0, Ny - 2))
                 
                 # === BILINEAR INTERPOLATION WEIGHT CALCULATION ===
                 # Mathematical basis: Bilinear interpolation formula f(x,y) = Σ f(xi,yi) * wi(x,y)
@@ -501,7 +476,7 @@ def _parallel_2d_forward_kernel(
     
     d_sino[iang, idet] = accum
 
-@_FASTMATH_DECORATOR
+@_NON_FASTMATH_DECORATOR
 def _parallel_2d_backward_kernel(
     d_sino, n_ang, n_det,
     d_image, Nx, Ny,
@@ -554,7 +529,7 @@ def _parallel_2d_backward_kernel(
     cos_a = d_cos[iang]         # Precomputed cosine of projection angle
     sin_a = d_sin[iang]         # Precomputed sine of projection angle
     # Normalize all physical distances to voxel units
-    u     = (idet - (n_det - 1) * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
+    u     = (idet - n_det * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
 
     # Define ray direction and starting point for parallel beam geometry
     dir_x, dir_y = cos_a, sin_a
@@ -599,8 +574,8 @@ def _parallel_2d_backward_kernel(
                 dx, dy = mid_x - ix0, mid_y - iy0
                 
                 # Clamp indices to stay in-bounds during interpolation
-                ix0 = min(ix0, Nx - 2)
-                iy0 = min(iy0, Ny - 2)
+                ix0 = max(0, min(ix0, Nx - 2))
+                iy0 = max(0, min(iy0, Ny - 2))
                 
                 # === ATOMIC BACKPROJECTION WITH BILINEAR WEIGHTS ===
                 # Distribute contribution weighted by segment length and interpolation weights
@@ -686,7 +661,7 @@ def _fan_2d_forward_kernel(
     cos_a = d_cos[iang]  # Precomputed cosine of projection angle
     sin_a = d_sin[iang]  # Precomputed sine of projection angle
     # Normalize all physical distances to voxel units
-    u     = (idet - (n_det - 1) * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
+    u     = (idet - n_det * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
     sid_v = sid / voxel_spacing  # Source-to-isocenter distance in voxel units
     sdd_v = sdd / voxel_spacing  # Source-to-detector distance in voxel units
 
@@ -757,8 +732,8 @@ def _fan_2d_forward_kernel(
                 dx, dy = mid_x - ix0, mid_y - iy0
                 
                 # Clamp indices to stay in-bounds during interpolation
-                ix0 = min(ix0, Nx - 2)
-                iy0 = min(iy0, Ny - 2)
+                ix0 = max(0, min(ix0, Nx - 2))
+                iy0 = max(0, min(iy0, Ny - 2))
                 
                 # Bilinear interpolation (identical to parallel beam)
                 val = (
@@ -781,7 +756,7 @@ def _fan_2d_forward_kernel(
     
     d_sino[iang, idet] = accum
 
-@_FASTMATH_DECORATOR
+@_NON_FASTMATH_DECORATOR
 def _fan_2d_backward_kernel(
     d_sino, n_ang, n_det,
     d_image, Nx, Ny,
@@ -839,7 +814,7 @@ def _fan_2d_backward_kernel(
     cos_a = d_cos[iang]         # Precomputed cosine of projection angle
     sin_a = d_sin[iang]         # Precomputed sine of projection angle
     # Normalize all physical distances to voxel units
-    u     = (idet - (n_det - 1) * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
+    u     = (idet - n_det * 0.5) * det_spacing / voxel_spacing  # Detector coordinate in voxel units
     sid_v = sid / voxel_spacing  # Source-to-isocenter distance in voxel units
     sdd_v = sdd / voxel_spacing  # Source-to-detector distance in voxel units
 
@@ -901,8 +876,8 @@ def _fan_2d_backward_kernel(
                 dx, dy = mid_x - ix0, mid_y - iy0
                 
                 # Clamp indices to stay in-bounds during interpolation
-                ix0 = min(ix0, Nx - 2)
-                iy0 = min(iy0, Ny - 2)
+                ix0 = max(0, min(ix0, Nx - 2))
+                iy0 = max(0, min(iy0, Ny - 2))
                 
                 # === ATOMIC BACKPROJECTION WITH BILINEAR WEIGHTS ===
                 # Distribute contribution weighted by segment length and interpolation weights
@@ -995,8 +970,8 @@ def _cone_3d_forward_kernel(
     # === 3D CONE BEAM GEOMETRY SETUP ===
     cos_a, sin_a = d_cos[iview], d_sin[iview]  # Projection angle trigonometry
     # Normalize all physical distances to voxel units
-    u     = (iu - (n_u - 1) * 0.5) * du / voxel_spacing  # Detector u-coordinate in voxel units
-    v     = (iv - (n_v - 1) * 0.5) * dv / voxel_spacing  # Detector v-coordinate in voxel units
+    u     = (iu - n_u * 0.5) * du / voxel_spacing  # Detector u-coordinate in voxel units
+    v     = (iv - n_v * 0.5) * dv / voxel_spacing  # Detector v-coordinate in voxel units
     sid_v = sid / voxel_spacing  # Source-to-isocenter distance in voxel units
     sdd_v = sdd / voxel_spacing  # Source-to-detector distance in voxel units
 
@@ -1091,9 +1066,9 @@ def _cone_3d_forward_kernel(
                 dx, dy, dz = mid_x - ix0, mid_y - iy0, mid_z - iz0  # Fractional parts: distance from base voxel center [0,1]
                 
                 # Clamp indices to stay in-bounds during interpolation
-                ix0 = min(ix0, Nx - 2)
-                iy0 = min(iy0, Ny - 2)
-                iz0 = min(iz0, Nz - 2)
+                ix0 = max(0, min(ix0, Nx - 2))
+                iy0 = max(0, min(iy0, Ny - 2))
+                iz0 = max(0, min(iz0, Nz - 2))
                 
                 # === TRILINEAR INTERPOLATION WEIGHT CALCULATION ===
                 # Mathematical basis: Trilinear interpolation formula f(x,y,z) = Σ f(xi,yi,zi) * wi(x,y,z)
@@ -1131,7 +1106,7 @@ def _cone_3d_forward_kernel(
     
     d_sino[iview, iu, iv] = accum
 
-@_FASTMATH_DECORATOR
+@_NON_FASTMATH_DECORATOR
 def _cone_3d_backward_kernel(
     d_sino, n_views, n_u, n_v,
     d_vol, Nx, Ny, Nz,
@@ -1196,8 +1171,8 @@ def _cone_3d_backward_kernel(
     g = d_sino[iview, iu, iv]  # Sinogram value to backproject along this ray
     cos_a, sin_a = d_cos[iview], d_sin[iview]  # Projection angle trigonometry
     # Normalize all physical distances to voxel units
-    u     = (iu - (n_u - 1) * 0.5) * du / voxel_spacing  # Detector u-coordinate in voxel units
-    v     = (iv - (n_v - 1) * 0.5) * dv / voxel_spacing  # Detector v-coordinate in voxel units
+    u     = (iu - n_u * 0.5) * du / voxel_spacing  # Detector u-coordinate in voxel units
+    v     = (iv - n_v * 0.5) * dv / voxel_spacing  # Detector v-coordinate in voxel units
     sid_v = sid / voxel_spacing  # Source-to-isocenter distance in voxel units
     sdd_v = sdd / voxel_spacing  # Source-to-detector distance in voxel units
 
@@ -1281,9 +1256,9 @@ def _cone_3d_backward_kernel(
                 dx, dy, dz = mid_x - ix0, mid_y - iy0, mid_z - iz0  # Fractional parts for 3D weights
                 
                 # Clamp indices to stay in-bounds during interpolation
-                ix0 = min(ix0, Nx - 2)
-                iy0 = min(iy0, Ny - 2)
-                iz0 = min(iz0, Nz - 2)
+                ix0 = max(0, min(ix0, Nx - 2))
+                iy0 = max(0, min(iy0, Ny - 2))
+                iz0 = max(0, min(iz0, Nz - 2))
                 
                 # === ATOMIC BACKPROJECTION WITH TRILINEAR WEIGHTS ===
                 # Distribute contribution weighted by segment length and interpolation weights
@@ -1408,7 +1383,7 @@ class ParallelProjectorFunction(torch.autograd.Function):
         sinogram = torch.zeros((n_angles, num_detectors), dtype=image.dtype, device=device)
 
         # Prepare trigonometric tables on the correct device
-        d_cos, d_sin = _trig_tables(angles, dtype=image.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=image.dtype, device=device)
 
         # Get Numba CUDA array views for kernel
         d_image = TorchCUDABridge.tensor_to_cuda_array(image)
@@ -1441,7 +1416,7 @@ class ParallelProjectorFunction(torch.autograd.Function):
 
         n_angles = angles.shape[0]
         grad_image = torch.zeros((Ny, Nx), dtype=grad_sinogram.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=grad_sinogram.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=grad_sinogram.dtype, device=device)
 
         d_grad_sino = TorchCUDABridge.tensor_to_cuda_array(grad_sinogram)
         d_img_grad = TorchCUDABridge.tensor_to_cuda_array(grad_image)
@@ -1541,7 +1516,7 @@ class ParallelBackprojectorFunction(torch.autograd.Function):
         reco = torch.zeros((Ny, Nx), dtype=sinogram.dtype, device=device)
 
         # Prepare trigonometric tables on the correct device
-        d_cos, d_sin = _trig_tables(angles, dtype=sinogram.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=sinogram.dtype, device=device)
 
         # Get Numba CUDA array views for kernel
         d_sino = TorchCUDABridge.tensor_to_cuda_array(sinogram)
@@ -1578,7 +1553,7 @@ class ParallelBackprojectorFunction(torch.autograd.Function):
         grad_sino = torch.zeros((n_ang, n_det), dtype=grad_output.dtype, device=device)
 
         # Prepare trigonometric tables on the correct device
-        d_cos, d_sin = _trig_tables(angles, dtype=grad_output.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=grad_output.dtype, device=device)
 
         # Get Numba CUDA array views for kernel
         d_grad_out = TorchCUDABridge.tensor_to_cuda_array(grad_output)
@@ -1679,7 +1654,7 @@ class FanProjectorFunction(torch.autograd.Function):
         n_ang = angles.shape[0]
 
         sinogram = torch.zeros((n_ang, num_detectors), dtype=image.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=image.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=image.dtype, device=device)
 
         d_image = TorchCUDABridge.tensor_to_cuda_array(image)
         d_sino = TorchCUDABridge.tensor_to_cuda_array(sinogram)
@@ -1713,7 +1688,7 @@ class FanProjectorFunction(torch.autograd.Function):
 
         n_ang = angles.shape[0]
         grad_img = torch.zeros((Ny, Nx), dtype=grad_sinogram.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=grad_sinogram.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=grad_sinogram.dtype, device=device)
 
         d_grad_sino = TorchCUDABridge.tensor_to_cuda_array(grad_sinogram)
         d_img_grad = TorchCUDABridge.tensor_to_cuda_array(grad_img)
@@ -1817,7 +1792,7 @@ class FanBackprojectorFunction(torch.autograd.Function):
         Ny, Nx = H, W
     
         reco = torch.zeros((Ny, Nx), dtype=sinogram.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=sinogram.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=sinogram.dtype, device=device)
 
         d_sino = TorchCUDABridge.tensor_to_cuda_array(sinogram)
         d_reco = TorchCUDABridge.tensor_to_cuda_array(reco)
@@ -1851,7 +1826,7 @@ class FanBackprojectorFunction(torch.autograd.Function):
         Ny, Nx = grad_output.shape
 
         grad_sino = torch.zeros((n_ang, n_det), dtype=grad_output.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=grad_output.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=grad_output.dtype, device=device)
 
         d_grad_out = TorchCUDABridge.tensor_to_cuda_array(grad_output)
         d_sino_grad = TorchCUDABridge.tensor_to_cuda_array(grad_sino)
@@ -1959,7 +1934,7 @@ class ConeProjectorFunction(torch.autograd.Function):
         _validate_3d_memory_layout(volume, expected_order='DHW')
 
         sino = torch.zeros((n_views, det_u, det_v), dtype=volume.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=volume.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=volume.dtype, device=device)
 
         volume_perm = volume.permute(2, 1, 0).contiguous()
         d_vol = TorchCUDABridge.tensor_to_cuda_array(volume_perm)
@@ -1997,7 +1972,7 @@ class ConeProjectorFunction(torch.autograd.Function):
         n_views = angles.shape[0]
 
         grad_vol_perm = torch.zeros((W, H, D), dtype=grad_sinogram.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=grad_sinogram.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=grad_sinogram.dtype, device=device)
 
         d_grad_sino = TorchCUDABridge.tensor_to_cuda_array(grad_sinogram)
         d_vol_grad = TorchCUDABridge.tensor_to_cuda_array(grad_vol_perm)
@@ -2116,7 +2091,7 @@ class ConeBackprojectorFunction(torch.autograd.Function):
         _validate_3d_memory_layout(sinogram, expected_order='VHW')
 
         vol_perm = torch.zeros((W, H, D), dtype=sinogram.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=sinogram.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=sinogram.dtype, device=device)
 
         d_sino = TorchCUDABridge.tensor_to_cuda_array(sinogram)
         d_reco = TorchCUDABridge.tensor_to_cuda_array(vol_perm)
@@ -2153,7 +2128,7 @@ class ConeBackprojectorFunction(torch.autograd.Function):
         n_views = angles.shape[0]
 
         grad_sino = torch.zeros((n_views, n_u, n_v), dtype=grad_output.dtype, device=device)
-        d_cos, d_sin = _trig_tables(angles, dtype=grad_output.dtype)
+        d_cos, d_sin = _trig_tables(angles, dtype=grad_output.dtype, device=device)
 
         grad_output_perm = grad_output.permute(2, 1, 0).contiguous()
         d_grad_out = TorchCUDABridge.tensor_to_cuda_array(grad_output_perm)

{diffct-1.2.3 → diffct-1.2.5}/examples/fbp_fan.py

@@ -2,6 +2,7 @@ import math
 import numpy as np
 import torch
 import matplotlib.pyplot as plt
+import torch.nn.functional as F
 from diffct.differentiable import FanProjectorFunction, FanBackprojectorFunction
 
 
@@ -76,9 +77,9 @@ def main():
     sino_weighted = sinogram * weights
     sinogram_filt = ramp_filter(sino_weighted)
 
-    reconstruction = FanBackprojectorFunction.apply(sinogram_filt, angles_torch,
+    reconstruction = F.relu(FanBackprojectorFunction.apply(sinogram_filt, angles_torch,
                                                     detector_spacing, Ny, Nx,
-                                                    sdd, sid, voxel_spacing)
+                                                    sdd, sid, voxel_spacing)) # ReLU to ensure non-negativity
     
     # --- FBP normalization ---
     # The backprojection is a sum over all angles. To approximate the integral,

{diffct-1.2.3 → diffct-1.2.5}/examples/fbp_parallel.py

@@ -1,6 +1,7 @@
 import numpy as np
 import torch
 import matplotlib.pyplot as plt
+import torch.nn.functional as F
 from diffct.differentiable import ParallelProjectorFunction, ParallelBackprojectorFunction
 
 
@@ -64,8 +65,8 @@ def main():
     
     sinogram_filt = ramp_filter(sinogram)
 
-    reconstruction = ParallelBackprojectorFunction.apply(sinogram_filt, angles_torch,
-                                                         detector_spacing, Ny, Nx, voxel_spacing)
+    reconstruction = F.relu(ParallelBackprojectorFunction.apply(sinogram_filt, angles_torch,
+                                                         detector_spacing, Ny, Nx, voxel_spacing)) # ReLU to ensure non-negativity
     
     # --- FBP normalization ---
     # The backprojection is a sum over all angles. To approximate the integral,

{diffct-1.2.3 → diffct-1.2.5}/examples/fdk_cone.py

@@ -2,6 +2,7 @@ import math
 import numpy as np
 import torch
 import matplotlib.pyplot as plt
+import torch.nn.functional as F
 from diffct.differentiable import ConeProjectorFunction, ConeBackprojectorFunction
 
 
@@ -110,8 +111,8 @@ def main():
     sino_weighted = sinogram * weights
     sinogram_filt = ramp_filter_3d(sino_weighted).contiguous()
 
-    reconstruction = ConeBackprojectorFunction.apply(sinogram_filt, angles_torch, Nz, Ny, Nx,
-                                                    du, dv, sdd, sid, voxel_spacing)
+    reconstruction = F.relu(ConeBackprojectorFunction.apply(sinogram_filt, angles_torch, Nz, Ny, Nx,
+                                                    du, dv, sdd, sid, voxel_spacing)) # ReLU to ensure non-negativity
     
     # --- FDK normalization ---
     # The backprojection is a sum over all angles. To approximate the integral,

{diffct-1.2.3 → diffct-1.2.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "diffct"
-version = "1.2.3"
+version = "1.2.5"
 description = "A CUDA-based library for computed tomography (CT) projection and reconstruction with differentiable operators"
 readme = "README.md"
 authors = [