import numpy as np import drjit as dr import mitsuba as mi mi.set_variant("cuda_ad_rgb") from PDE3D.Coefficient import * from PDE3D.utils import * from PDE3D.BoundaryShape import * from PDE3D.Solver import * from python3D.optimization.sketch import * import time def compute_primal(wos : WosVariable, split : Split, bbox : mi.ScalarBoundingBox3f, resolution : list[int], spp : int, centered : bool, max_split_depth : int, conf_numbers : list[mi.UInt32], compute_std: bool, verbose : bool = False): points = create_bbox_points(bbox, resolution, spp, centered=centered) L, _ = wos.solve(points, split = split, max_depth_split = max_split_depth, conf_numbers=conf_numbers, all_inside = False, verbose = verbose) std = None if compute_std: _, tensor, var, _= create_volume_from_result(L, resolution, compute_std) std = np.sqrt(var) else: _, tensor = create_volume_from_result(L, resolution, compute_std) return L, tensor, std def compute_primals(wos : WosVariable, split : Split, seed : int, bbox : mi.ScalarBoundingBox3f, resolution : list[int], spp : int, centered : bool, max_split_depth : int, compute_std : bool, confs_iter : int = 8): num_conf = wos.input.shape.num_conf_d results = np.zeros([num_conf, resolution[0], resolution[1], resolution[2]]) results_std = np.zeros([num_conf, resolution[0], resolution[1], resolution[2]]) num_iter = int(np.ceil(num_conf / confs_iter)) initstate, initseq = tea(dr.arange(mi.UInt64, num_iter), mi.UInt64(seed)) pcg = mi.PCG32() pcg.seed(initstate, initseq) seeds = pcg.next_uint32_bounded(10000).numpy().squeeze() seeds = [seeds] if seeds.ndim == 0 else seeds for i, seed in enumerate(seeds): begin= i * confs_iter end = min(num_conf, (i + 1) * confs_iter) conf_numbers = [dr.opaque(mi.UInt32, j, shape = (1)) for j in range(begin, end)] wos.change_seed(seed) _, tensor, std = compute_primal(wos, split, bbox, resolution, spp, centered, max_split_depth, conf_numbers, compute_std) results[begin : end] = tensor.numpy() if std is not None: results_std[begin : end] = std return results, results_std def optimize(path : str, wos : WosVariable, wos_obj : WosVariable, objectives : mi.TensorXf, coeff_str : str, split : Split, bbox : list[list], resolution : list[int], spp : int, primal_spp : int, seed : int, conf_per_iter : int, max_split_depth : int, num_iter : int, learning_rate : float, post_process : callable, centered : bool, plot : bool = True, compute_std : bool = True, verbose : bool = False, max_range : list[float] = [0.1, 3], measure_time = False, vis_set = []): #set_matplotlib(9) num_conf = wos.input.shape.num_conf_d conf_per_iter = min(num_conf, conf_per_iter) # Compute the primals for getting the first loss values. primals, primals_std = compute_primals(wos, split, seed, bbox, resolution, primal_spp, centered, max_split_depth, compute_std, confs_iter = 16) losses = MSE_numpy(primals, objectives) coeff = wos.input.get_coefficient(coeff_str) coeff_obj = wos_obj.input.get_coefficient(coeff_str) plot_coeff(coeff, wos.input.shape, None, path, f"{coeff_str}-begin") plot_coeff(coeff, wos.input.shape, max_range, path, f"{coeff_str}-begin-scaled") # Create the optimizer opt = mi.ad.Adam(lr= learning_rate, params = wos.opt_params) wos.update(opt) loss_list = [] #loss_reg_list = [] # Record the objective optimization parameters if measure_time: primal_time = [] grad_time = [] vis_time = [] path_npy = os.path.join(path, "npy") np.save(os.path.join(path_npy, "objectives.npy"), objectives) np.save(os.path.join(path_npy, "primal", "primal-0.npy"), primals) if compute_std: np.save(os.path.join(path_npy, "primal", "std-0.npy"), primal_std) for key in opt.keys(): np.save(os.path.join(path_npy, "tensor", f"{key}-0.npy"), opt[key].numpy().squeeze()) if wos_obj is not None: np.save(os.path.join(path_npy, f"objective-tensor.npy"), wos_obj.input.get_coefficient(coeff_str).tensor.numpy().squeeze()) initstate, initseq = tea(dr.arange(mi.UInt32, conf_per_iter), mi.UInt64(seed)) sampler = mi.PCG32() sampler.seed(initstate, initseq) print("Optimization Started!") # Begin optimization for i in range(num_iter): if len(vis_set) > 0: if measure_time: dr.sync_thread() t0 = time.time() wos.change_seed(i+1) _, primal_vis, _ = compute_primal(wos, split, bbox, resolution, primal_spp, centered, max_split_depth, vis_set, compute_std=compute_std, verbose = verbose) np.save(os.path.join(path_npy, "primal", f"primalvis-{i}.npy"), primal_vis.numpy().squeeze()) if measure_time: dr.sync_thread() t1 = time.time() t_vis = t1 - t0 #print(f"Primal time: {t_primal}") vis_time.append(t_vis) seed_iter = np.random.randint(0, 2**15) wos.change_seed(seed_iter) confs_iter = np.random.choice(range(num_conf), conf_per_iter, replace = False) confs_opaque = [dr.opaque(mi.UInt32, j, shape = (1)) for j in confs_iter.tolist()] obj_res_iter = objectives[confs_iter, :] if measure_time: dr.sync_thread() t0 = time.time() L, primal, primal_std = compute_primal(wos, split, bbox, resolution, primal_spp, centered, max_split_depth, confs_opaque, compute_std=compute_std, verbose = verbose) dr.eval(primal) primal_np = primal.numpy() primals[confs_iter] = primal_np losses[confs_iter] = MSE_numpy(primal_np, obj_res_iter) if compute_std: primals_std[confs_iter] = primal_std if measure_time: dr.sync_thread() t1 = time.time() t_primal = t1 - t0 #print(f"Primal time: {t_primal}") primal_time.append(t_primal) obj_opaque = mi.TensorXf(obj_res_iter.tolist()) dr.make_opaque(obj_opaque) loss_grad = compute_loss_grad_vol(primal, obj_opaque) dL = compute_dL_vol(loss_grad, spp) points = create_bbox_points(bbox, resolution, spp, seed = seed, centered = centered) with dr.isolate_grad(): _ = wos.solve_grad(points_in = points, split = split, dL = dL, max_depth_split = max_split_depth, conf_numbers=confs_opaque, verbose = verbose) coeff = wos.input.get_coefficient(coeff_str) grad_np = dr.grad(coeff.tensor).numpy() opt.step() post_process(opt) wos.update(opt) if measure_time: dr.sync_thread() t2 = time.time() t_grad = t2 - t1 #print(f"Grad time: {t_grad}") grad_time.append(t_grad) np.save(os.path.join(path, "npy", "grad", f"grad-{i+1}.npy"), np.array(grad_np).squeeze()) np.save(os.path.join(path, "npy", "primal", f"primal-{i+1}.npy"), np.array(primals).squeeze()) if compute_std: np.save(os.path.join(path, "npy", "primal", f"std-{i+1}.npy"), np.array(primals_std).squeeze()) for key in opt.keys(): np.save(os.path.join(path, "npy", "tensor", f"{key}-{i+1}.npy"), opt[key].numpy().squeeze()) loss_list.append(np.array(losses)) print(f"Iteration {i} is finished. Loss = {np.array(losses).sum()}") np.save(os.path.join(path, "npy", "losses.npy"), np.array(loss_list)) if measure_time: np.save(os.path.join(path, "npy", "primal_time.npy"), np.array(primal_time)) np.save(os.path.join(path, "npy", "grad_time.npy"), np.array(grad_time)) np.save(os.path.join(path, "npy", "vis_time.npy"), np.array(vis_time)) print("Optimization Ended!") plot_summary(loss_list, path, log=False) plot_summary(loss_list, path, log=True) plot_coeff(coeff, wos.input.shape, max_range, path, f"{coeff_str}-end-scaled") plot_coeff(coeff, wos.input.shape, None, path, f"{coeff_str}-end") return wos