|
- import os
- import numpy as np
- import torch.nn
- from tqdm.auto import tqdm
- from opt_sem_replica import config_parser
- from opt_sem_kitti_360 import kitti_config_parser
-
- import json, random
- from renderer import *
- from utils.utils import *
- from torch.utils.tensorboard import SummaryWriter
- import datetime
-
- from datasets import dataset_dict
- import sys
-
- from models.tensorRFSemVMSplit import TensorSemVMSplit
-
-
- device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # input from TensorBase ?
-
- sem_renderer = OctreeRender_trilinear_fast_with_sem
-
-
- class SimpleSampler:
- def __init__(self, total, batch):
- self.total = total
- self.batch = batch
- self.curr = total
- self.ids = None
-
- def nextids(self):
- self.curr += self.batch
- if self.curr + self.batch > self.total:
- self.ids = torch.LongTensor(np.random.permutation(self.total))
- self.curr = 0
- return self.ids[self.curr:self.curr + self.batch]
-
-
- @torch.no_grad()
- def export_mesh(args):
- ckpt = torch.load(args.ckpt, map_location=device)
- kwargs = ckpt['kwargs']
- kwargs.update({'device': device})
- tensorf = eval(args.model_name)(**kwargs)
- tensorf.load(ckpt)
-
- alpha, _ = tensorf.getDenseAlpha()
- convert_sdf_samples_to_ply(alpha.cpu(), f'{args.ckpt[:-3]}.ply', bbox=tensorf.aabb.cpu(), level=0.005)
-
-
- def tps_reconstruction(args):
- # office_0
- start_recon_case = 161
- end_recon_case = 170
-
- # room_0
- # start_recon_case = 125
- # end_recon_case = 134
-
- base_datadir = args.datadir
- for case in range(start_recon_case, end_recon_case+1):
- args.datadir = os.path.join(base_datadir, "case_"+str(case))
- # init dataset
- dataset = dataset_dict[args.dataset_name]
-
- if args.dataset_name == "kitti360":
- train_dataset = dataset(args.datadir, split='train', start=args.start, end=args.end, near=args.near, far=args.far, downsample=args.downsample_train, is_stack=False)
- test_dataset = dataset(args.datadir, split='test', start=args.start, end=args.end, near=args.near, far=args.far, test_ids=args.test_ids, downsample=args.downsample_train, is_stack=True)
- train_dataset.remap_sem_gt_label(train_dataset.sem_samples["sem_img"],
- test_dataset.sem_samples["sem_img"])
- test_dataset.remap_sem_gt_label(train_dataset.sem_samples["sem_img"],
- test_dataset.sem_samples["sem_img"])
- train_dataset.set_label_colour_map()
- test_dataset.set_label_colour_map()
- else:
- train_dataset = dataset(args.datadir, near=args.near, far=args.far,
- scene_bbox_stretch=args.scene_bbox_stretch,
- split='train', downsample=args.downsample_train, is_stack=False,
- use_tps_dataset=True)
-
- if args.dataset_name == "replica" or "replica_dmnerf":
- train_dataset.remap_sem_gt_label(load_map=True, ins2label_path=args.ins2label_path)
- train_dataset.set_label_colour_map(sem_info_path=args.sem_info_path,
- label2color_path=args.label2color_path)
-
- white_bg = train_dataset.white_bg
- near_far = train_dataset.near_far
- ndc_ray = args.ndc_ray
-
- # init resolution ?
- upsamp_list = args.upsamp_list
- update_AlphaMask_list = args.update_AlphaMask_list
-
- n_lamb_sigma = args.n_lamb_sigma
- n_lamb_sh = args.n_lamb_sh
- n_lamb_sem = args.n_lamb_sem
-
- if args.add_timestamp:
- logfolder = f'{args.datadir}/{args.expname}{datetime.datetime.now().strftime("-%Y%m%d-%H%M%S")}'
- else:
- logfolder = f'{args.datadir}/{args.expname}'
-
- # init log file
- os.makedirs(logfolder, exist_ok=True)
- os.makedirs(f'{logfolder}/imgs_vis', exist_ok=True)
- os.makedirs(f'{logfolder}/imgs_rgba', exist_ok=True)
- os.makedirs(f'{logfolder}/rgba', exist_ok=True)
- os.makedirs(f'{logfolder}/sem', exist_ok=True) # ?
- summary_writer = SummaryWriter(logfolder)
-
- # init parameters
- # tensorVM, renderer = init_parameters(args, train_dataset.scene_bbox.to(device), reso_list[0])
- aabb = train_dataset.scene_bbox.to(device) # adaptive scene_bbox world coordinates, (2, 3)
- reso_cur = N_to_reso(args.N_voxel_init, aabb) # N_voxel_grids of each dimension
- nSamples = min(args.nSamples, cal_n_samples(reso_cur, args.step_ratio))
-
- # build the model
- if args.ckpt is not None:
- ckpt = torch.load(args.ckpt, map_location=device)
- kwargs = ckpt['kwargs']
- kwargs.update({'device': device})
- sem_tensorf = eval(args.model_name)(**kwargs)
- sem_tensorf.load(ckpt)
- else:
- sem_tensorf = eval(args.model_name)(aabb, reso_cur, device,
- density_n_comp=n_lamb_sigma, appearance_n_comp=n_lamb_sh, sem_n_comp=n_lamb_sem,
- app_dim=args.data_dim_color, sem_dim=train_dataset.num_valid_semantic_class,
- near_far=near_far,
- shadingMode=args.shadingMode, alphaMask_thres=args.alpha_mask_thre,
- density_shift=args.density_shift, distance_scale=args.distance_scale,
- pos_pe=args.pos_pe, view_pe=args.view_pe, fea_pe=args.fea_pe,
- featureC=args.featureC, step_ratio=args.step_ratio,
- fea2denseAct=args.fea2denseAct)
-
- grad_vars = sem_tensorf.get_optparam_groups(args.lr_init, args.lr_basis)
- if args.lr_decay_iters > 0:
- lr_factor = args.lr_decay_target_ratio ** (1 / args.lr_decay_iters)
- else:
- args.lr_decay_iters = args.n_iters
- lr_factor = args.lr_decay_target_ratio ** (1 / args.n_iters)
-
- print("lr decay", args.lr_decay_target_ratio, args.lr_decay_iters)
-
- optimizer = torch.optim.Adam(grad_vars, betas=(0.9, 0.99))
-
- # linear in logrithmic space
- N_voxel_list = (torch.round(torch.exp(
- torch.linspace(np.log(args.N_voxel_init), np.log(args.N_voxel_final), len(upsamp_list) + 1))).long()).tolist()[
- 1:]
-
- torch.cuda.empty_cache()
- PSNRs, PSNRs_test = [], [0]
-
- allrays, allrgbs, allsems = train_dataset.all_rays, train_dataset.all_rgbs, train_dataset.sem_samples["sem_remap"]
- allsems = allsems.reshape(allrgbs.shape[0], 1) # (num_imgs*h*w, 1)
- allsems = torch.tensor(allsems)
-
- if not args.ndc_ray:
- allrays, allrgbs, allsems = sem_tensorf.filtering_rays(allrays, allrgbs, allsems, bbox_only=True)
- trainingSampler = SimpleSampler(allrays.shape[0], args.batch_size)
-
- Ortho_reg_weight = args.Ortho_weight
- print("initial Ortho_reg_weight", Ortho_reg_weight)
-
- L1_reg_weight = args.L1_weight_inital
- print("initial L1_reg_weight", L1_reg_weight)
- TV_weight_density, TV_weight_app, TV_weight_sem = args.TV_weight_density, args.TV_weight_app, args.TV_weight_sem
- tvreg = TVLoss()
- print(f"initial TV_weight density: {TV_weight_density} appearance: {TV_weight_app} semantic: {TV_weight_sem}")
-
- pbar = tqdm(range(args.n_iters), miniters=args.progress_refresh_rate, file=sys.stdout)
- # use values on each pixel, not each image
- for iteration in pbar:
- ray_idx = trainingSampler.nextids() # Tensor, ()
- rays_train, rgb_train, sem_train = allrays[ray_idx], allrgbs[ray_idx].to(device), allsems[ray_idx].to(device)
-
- # rgb_map, alphas_map, sem_map, depth_map, weights, uncertainty
- rgb_map, alphas_map, sem_map, depth_map, weights, uncertainty = sem_renderer(rays_train, sem_tensorf,
- chunk=args.batch_size,
- N_samples=nSamples, white_bg=white_bg,
- ndc_ray=ndc_ray,
- device=device,
- is_train=True)
-
- rgb_loss = torch.mean((rgb_map - rgb_train) ** 2)
-
- sem_loss_weight = 4e-2
- sem_crossentropy_loss = train_dataset.get_sem_loss(sem_map, sem_train)
- sem_loss = sem_crossentropy_loss * sem_loss_weight
-
- # loss
- total_loss = rgb_loss + sem_loss
- if Ortho_reg_weight > 0:
- loss_reg = sem_tensorf.vector_comp_diffs()
- total_loss += Ortho_reg_weight * loss_reg # ?
- summary_writer.add_scalar('train/reg', loss_reg.detach().item(), global_step=iteration)
- if L1_reg_weight > 0:
- loss_reg_L1 = sem_tensorf.density_L1()
- total_loss += L1_reg_weight * loss_reg_L1
- summary_writer.add_scalar('train/reg_l1', loss_reg_L1.detach().item(), global_step=iteration)
-
- if TV_weight_density > 0:
- TV_weight_density *= lr_factor
- loss_tv = sem_tensorf.TV_loss_density(tvreg) * TV_weight_density
- total_loss = total_loss + loss_tv
- summary_writer.add_scalar('train/reg_tv_density', loss_tv.detach().item(), global_step=iteration)
- if TV_weight_app > 0:
- TV_weight_app *= lr_factor
- loss_tv = sem_tensorf.TV_loss_app(tvreg) * TV_weight_app
- total_loss = total_loss + loss_tv
- summary_writer.add_scalar('train/reg_tv_app', loss_tv.detach().item(), global_step=iteration)
- if TV_weight_sem > 0:
- TV_weight_sem *= lr_factor
- loss_tv = sem_tensorf.TV_loss_sem(tvreg) * TV_weight_sem
- total_loss = total_loss + loss_tv
- summary_writer.add_scalar('train/reg_tv_sem', loss_tv.detach().item(), global_step=iteration)
-
- optimizer.zero_grad()
- total_loss.backward()
- optimizer.step()
-
- rgb_loss = rgb_loss.detach().item()
- sem_loss = sem_loss.detach().item()
-
- PSNRs.append(-10.0 * np.log(rgb_loss) / np.log(10.0))
- summary_writer.add_scalar('train/PSNR', PSNRs[-1], global_step=iteration)
- summary_writer.add_scalar('train/mse', rgb_loss, global_step=iteration)
- summary_writer.add_scalar('train/sem_CE', sem_loss, global_step=iteration)
-
- for param_group in optimizer.param_groups:
- param_group['lr'] = param_group['lr'] * lr_factor
-
- if iteration == (args.n_iters-1):
- final_train_PSNR = np.mean(PSNRs)
- if final_train_PSNR < args.PSNR_thresh:
- print(f"bad case {case}")
-
- # Print the current values of the losses.
- if iteration % args.progress_refresh_rate == 0:
- pbar.set_description(
- f'Iteration {iteration:05d}:'
- + f' train_psnr = {float(np.mean(PSNRs)):.2f}'
- + f' test_psnr = {float(np.mean(PSNRs_test)):.2f}'
- + f' mse = {rgb_loss:.6f}'
- + f' sem_CE = {sem_loss:.6f}'
- )
- PSNRs = []
-
- if iteration in update_AlphaMask_list:
- if reso_cur[0] * reso_cur[1] * reso_cur[2] < 256 ** 3: # update volume resolution
- reso_mask = reso_cur
- new_aabb = sem_tensorf.updateAlphaMask(tuple(reso_mask))
- if iteration == update_AlphaMask_list[0]:
- sem_tensorf.shrink(new_aabb)
- # tensorVM.alphaMask = None
- L1_reg_weight = args.L1_weight_rest
- print("continuing L1_reg_weight", L1_reg_weight)
-
- if not args.ndc_ray and iteration == update_AlphaMask_list[1]:
- # filter rays outside the bbox
- allrays, allrgbs, allsems = sem_tensorf.filtering_rays(allrays, allrgbs, allsems)
- trainingSampler = SimpleSampler(allrgbs.shape[0], args.batch_size)
-
- if iteration in upsamp_list:
- n_voxels = N_voxel_list.pop(0)
- reso_cur = N_to_reso(n_voxels, sem_tensorf.aabb)
- nSamples = min(args.nSamples, cal_n_samples(reso_cur, args.step_ratio))
- sem_tensorf.upsample_volume_grid(reso_cur)
-
- if args.lr_upsample_reset:
- print("reset lr to initial")
- lr_scale = 1 # 0.1 ** (iteration / args.n_iters)
- else:
- lr_scale = args.lr_decay_target_ratio ** (iteration / args.n_iters)
- grad_vars = sem_tensorf.get_optparam_groups(args.lr_init * lr_scale, args.lr_basis * lr_scale)
- optimizer = torch.optim.Adam(grad_vars, betas=(0.9, 0.99))
-
- sem_tensorf.save(f'{logfolder}/{args.expname}.pth')
-
- if args.render_train:
- os.makedirs(f'{logfolder}/imgs_train_all', exist_ok=True)
- train_dataset = dataset(args.datadir, split='train', downsample=args.downsample_train, is_stack=True, tps_recon=True)
- PSNRs_test = evaluation_with_sem(train_dataset, sem_tensorf, args, sem_renderer, f'{logfolder}/imgs_train_all/',
- N_vis=-1, N_samples=-1, white_bg=white_bg, ndc_ray=ndc_ray, device=device,
- chunk_size=args.batch_size)
- print(f'======> {args.expname} train all psnr: {np.mean(PSNRs_test)} <========================')
-
- if args.render_test:
- os.makedirs(f'{logfolder}/imgs_test_all', exist_ok=True)
- PSNRs_test = evaluation_with_sem(test_dataset, sem_tensorf, args, sem_renderer, f'{logfolder}/imgs_test_all/',
- N_vis=-1, N_samples=-1, white_bg=white_bg, ndc_ray=ndc_ray, device=device,
- chunk_size=args.batch_size)
- summary_writer.add_scalar('test/psnr_all', np.mean(PSNRs_test), global_step=iteration)
- print(f'======> {args.expname} test all psnr: {np.mean(PSNRs_test)} <========================')
-
- print(f'======> reconstruction case {case} done<========================')
-
-
- if __name__ == '__main__':
- torch.set_default_dtype(torch.float32)
- torch.manual_seed(20211202)
- np.random.seed(20211202)
-
- args = config_parser()
- # args = kitti_config_parser()
- # print(args)
-
- if args.export_mesh:
- export_mesh(args)
-
- tps_reconstruction(args)
|