import torch from torch.autograd import grad def hessian(y, x): ''' hessian of y wrt x y: shape (meta_batch_size, num_observations, channels) x: shape (meta_batch_size, num_observations, 2) ''' meta_batch_size, num_observations = y.shape[:2] grad_y = torch.ones_like(y[..., 0]).to(y.device) h = torch.zeros(meta_batch_size, num_observations, y.shape[-1], x.shape[-1], x.shape[-1]).to(y.device) for i in range(y.shape[-1]): # calculate dydx over batches for each feature value of y dydx = grad(y[..., i], x, grad_y, create_graph=True)[0] # calculate hessian on y for each x value for j in range(x.shape[-1]): h[..., i, j, :] = grad(dydx[..., j], x, grad_y, create_graph=True)[0][..., :] status = 0 if torch.any(torch.isnan(h)): status = -1 return h, status def laplace(y, x): grad = gradient(y, x) return divergence(grad, x) def divergence(y, x): div = 0. for i in range(y.shape[-1]): div += grad(y[..., i], x, torch.ones_like(y[..., i]), create_graph=True)[0][..., i:i+1] return div def gradient(y, x, grad_outputs=None): if grad_outputs is None: grad_outputs = torch.ones_like(y) grad = torch.autograd.grad(y, [x], grad_outputs=grad_outputs, create_graph=True)[0] return grad def jacobian(y, x): ''' jacobian of y wrt x ''' meta_batch_size, num_observations = y.shape[:2] jac = torch.zeros(meta_batch_size, num_observations, y.shape[-1], x.shape[-1]).to(y.device) # (meta_batch_size*num_points, 2, 2) for i in range(y.shape[-1]): # calculate dydx over batches for each feature value of y y_flat = y[...,i].view(-1, 1) jac[:, :, i, :] = grad(y_flat, x, torch.ones_like(y_flat), create_graph=True)[0] status = 0 if torch.any(torch.isnan(jac)): status = -1 return jac, status def get_gradient(net, x): y = net(x) input_val = torch.ones_like(y) x_grad = grad(y, x, input_val, create_graph=True)[0] return x_grad def get_gradient_sdf(net, x): y = net(x) y = torch.abs(y) input_val = torch.ones_like(y) x_grad = grad(y, x, input_val, create_graph=True)[0] return x_grad def get_batch_jacobian(net, x0, x1, x, noutputs): # x = x.unsqueeze(1) # b, 1 ,in_dim # x: b, 3, num_points n, np = x.size(0), x.size(2) x = x.repeat(1, 1, noutputs) # b, 3, num_points * out_dim # x.requires_grad_(True) y = net(x0, x1, x) # b, out_dim, num_points * out_dim y = y.view(n, noutputs, noutputs, np) input_val = torch.eye(noutputs).view(1, noutputs, noutputs, 1).repeat(n, 1, 1, np).cuda() # y.backward(input_val) x_grad = grad(y, x, input_val, create_graph=True)[0] return x_grad.view(n, 3, noutputs, np).transpose(1, 2) def get_batch_jacobian_netG(net, netG, x0, x1, x2, x, noutputs): # x = x.unsqueeze(1) # b, 1 ,in_dim # x: b, 3, num_points n, np = x.size(0), x.size(2) x = x.repeat(1, 1, noutputs) # b, 3, num_points * out_dim # x.requires_grad_(True) y = net(netG, x0, x1, x2, x) # b, out_dim, num_points * out_dim y = y.view(n, noutputs, noutputs, np) input_val = torch.eye(noutputs).view(1, noutputs, noutputs, 1).repeat(n, 1, 1, np).cuda() # y.backward(input_val) x_grad = grad(y, x, input_val, create_graph=True)[0] return x_grad.view(n, 3, noutputs, np).transpose(1, 2)