ViewDesignEngine/python/vde/torch_sdf.py

"""
PyTorch integration for differentiable SDF operations.

Provides torch.autograd.Function wrappers that allow SDF
evaluation within PyTorch computation graphs.

Usage:
    import torch
    from vde.torch_sdf import SdfEvaluator
    
    evaluator = SdfEvaluator(node)
    x = torch.randn(100, 3, requires_grad=True)
    d = evaluator(x)  # SDF values
    loss = d.abs().mean()
    loss.backward()  # x.grad contains df/dx
"""
import torch

class SdfEvaluator(torch.autograd.Function):
    """Differentiable SDF evaluation for PyTorch.
    
    Forward: evaluate SDF at input points.
    Backward: compute gradient of SDF w.r.t. input points.
    
    Usage:
        evaluator = SdfEvaluator.apply
        d = evaluator(node, points)  # where node is an SdfNode
    """
    
    @staticmethod
    def forward(ctx, node, points):
        """Forward pass: evaluate SDF.
        
        Args:
            node: SdfNode (from C++ binding)
            points: torch.Tensor of shape (N, 3)
        Returns:
            distances: torch.Tensor of shape (N,)
        """
        import numpy as np
        from vde._vde import evaluate
        
        # Convert to numpy for C++ evaluation
        pts_np = points.detach().cpu().numpy().astype(np.float64)
        n = pts_np.shape[0]
        
        distances = np.zeros(n, dtype=np.float64)
        for i in range(n):
            p = pts_np[i]
            distances[i] = evaluate(node, p)
        
        ctx.save_for_backward(points)
        ctx.node = node  # Store node reference for backward
        
        return torch.from_numpy(distances).to(points.device).to(points.dtype)
    
    @staticmethod
    def backward(ctx, grad_output):
        """Backward pass: gradient of SDF w.r.t. input points.
        
        df/dx = gradient of SDF at each input point
        Chain rule: dL/dx = dL/d(sdf) * d(sdf)/dx
        """
        import numpy as np
        from vde._vde import evaluate
        
        points = ctx.saved_tensors[0]
        node = ctx.node
        
        pts_np = points.detach().cpu().numpy().astype(np.float64)
        n = pts_np.shape[0]
        
        gradients = np.zeros((n, 3), dtype=np.float64)
        h = 1e-6
        
        for i in range(n):
            p = pts_np[i]
            # Central finite differences
            for j in range(3):
                pp = p.copy()
                pm = p.copy()
                pp[j] += h
                pm[j] -= h
                gradients[i, j] = (evaluate(node, pp) - evaluate(node, pm)) / (2.0 * h)
        
        grad_input = torch.from_numpy(gradients).to(points.device).to(points.dtype)
        grad_node = None  # Don't backprop to SDF parameters (for now)
        
        return grad_node, grad_input * grad_output.unsqueeze(-1)


class SdfShapeFitter:
    """Fit SDF shapes to target point clouds with gradient descent.
    
    Usage:
        fitter = SdfShapeFitter(sphere(1.0))
        optimized = fitter.fit(target_points, lr=0.01, epochs=100)
    """
    
    def __init__(self, node):
        self.node = node
    
    def fit(self, target_points, lr=0.01, epochs=100, verbose=False):
        """Fit SDF shape to target point cloud.
        
        Args:
            target_points: torch.Tensor of shape (N, 3)
            lr: learning rate
            epochs: number of optimization steps
            verbose: print progress
        
        Returns:
            optimized node (SdfNode)
        """
        import numpy as np
        from vde._vde import evaluate
        
        pts = target_points.detach().cpu().numpy().astype(np.float64)
        n = pts.shape[0]
        h = 1e-6
        
        # Get current parameters
        params = self.node.params
        param_list = self._params_to_list(params)
        
        for epoch in range(epochs):
            # Forward: evaluate SDF at all points
            loss = 0.0
            grads = {k: 0.0 for k in self._param_names()}
            
            for i in range(n):
                p = pts[i]
                d = evaluate(self.node, p)
                loss += abs(d)
                
                # Compute numerical gradient w.r.t. parameters
                for name, idx in self._param_indices().items():
                    orig_val = param_list[idx]
                    param_list[idx] = orig_val + h
                    self._list_to_params(param_list, self.node.params)
                    dp = evaluate(self.node, p)
                    
                    param_list[idx] = orig_val - h
                    self._list_to_params(param_list, self.node.params)
                    dm = evaluate(self.node, p)
                    
                    param_list[idx] = orig_val
                    grads[name] += (abs(dp) - abs(dm)) / (2.0 * h)
            
            self._list_to_params(param_list, self.node.params)
            loss /= n
            for name in grads:
                grads[name] /= n
            
            # Gradient descent step
            for name, idx in self._param_indices().items():
                param_list[idx] -= lr * grads[name]
            
            self._list_to_params(param_list, self.node.params)
            
            if verbose and epoch % 10 == 0:
                print(f"Epoch {epoch}: loss = {loss:.6f}")
        
        return self.node
    
    def _param_names(self):
        return ['radius', 'extent_x', 'extent_y', 'extent_z', 'height']
    
    def _param_indices(self):
        return {
            'radius': 0,
            'extent_x': 1, 'extent_y': 2, 'extent_z': 3,
            'height': 4,
        }
    
    def _params_to_list(self, params):
        return [params.radius, params.extents.x, params.extents.y,
                params.extents.z, params.height]
    
    def _list_to_params(self, lst, params):
        params.radius = lst[0]
        params.extents.x = lst[1]
        params.extents.y = lst[2]
        params.extents.z = lst[3]
        params.height = lst[4]