We use python files as configs, incorporate modular and inheritance design into our config system, which is convenient to conduct various experiments.
You can find all the provided configs under $PROJECT/configs.
We can logically divide the configuration file into components:
- training
- model
- data
The fllowing content explain these configuration components one by one.
-
training training configurations contains all paramters to control model training, include optimizer, hooks, runner and soon on.
import os from datetime import datetime method = 'nerf' # which nerf method # optimizer setting optimizer = dict(type='Adam', lr=5e-4, betas=(0.9, 0.999)) optimizer_config = dict(grad_clip=None) max_iters = 20000 # train for how many iters lr_config = dict(policy='step', step=500 * 1000, gamma=0.1, by_epoch=False) # learning rate and decay checkpoint_config = dict(interval=5000, by_epoch=False) # when to save checkpoint log_level = 'INFO' log_config = dict(interval=5000, by_epoch=False, hooks=[dict(type='TextLoggerHook')]) workflow = [('train', 5000), ('val', 1)] # loop: train 5000 iters, validate 1 iter # hooks # 'params' are numeric type value, 'variables' are variables in local environment train_hooks = [ dict(type='SetValPipelineHook', params=dict(), variables=dict(valset='valset')), dict(type='ValidateHook', params=dict(save_folder='visualizations/validation')), dict(type='SaveSpiralHook', params=dict(save_folder='visualizations/spiral')), dict(type='PassIterHook', params=dict()), # 将当前iter数告诉dataset dict(type='OccupationHook', params=dict()), # no need for open-source vision ] test_hooks = [ dict(type='SetValPipelineHook', params=dict(), variables=dict(valset='testset')), dict(type='TestHook', params=dict()), ] # runner train_runner = dict(type='NerfTrainRunner') test_runner = dict(type='NerfTestRunner') # runtime settings num_gpus = 1 distributed = (num_gpus > 1) # whether to use ddp work_dir = './work_dirs/nerfsv3/nerf_#DATANAME#_base01/' # where to save ckpt, images, video, logs timestamp = datetime.now().strftime('%d-%b-%H-%M') # to make sure different log-files each train # some shared params by model and data, to avoid define twice dataset_type = 'blender' no_batching = True # only take random rays from 1 image at a time no_ndc = True white_bkgd = True # set to render synthetic data on a white bkgd (always use for dvoxels) is_perturb = True # set to 0. for no jitter, 1. for jitter use_viewdirs = True # use full 5D input instead of 3D N_rand_per_sampler = 1024 * 4 # how many N_rand in get_item() function lindisp = False # sampling linearly in disparity rather than depth N_samples = 64 # number of coarse samples per ray # resume_from = os.path.join(work_dir, 'latest.pth') # load_from = os.path.join(work_dir, 'latest.pth')
-
model define network structure, a network is usually composed of embedder, mlp and render.
model = dict( type='NerfNetwork', # network class name cfg=dict( phase='train', # 'train' or 'test' N_importance=128, # number of additional fine samples per ray is_perturb=is_perturb, # see above chunk=1024 * 32, # mainly work for val, to avoid oom bs_data='rays_o', # the data's shape indicates the real batch-size, this's also the num of rays ), mlp=dict( # coarse mlp model type='NerfMLP', # mlp class name skips=[4], netdepth=8, # layers in network netwidth=256, # channels per layer netchunk=1024 * 32, # to avoid oom output_ch=5, # 5 if cfg.N_importance>0 else 4 use_viewdirs=use_viewdirs, embedder=dict( type='BaseEmbedder', # embedder class name i_embed=0, # set 0 for default positional encoding, -1 for none multires=10, # log2 of max freq for positional encoding (3D location) multires_dirs=4, # this is 'multires_views' in origin codes, log2 of max freq for positional encoding (2D direction) ), ), mlp_fine=dict( # fine model type='NerfMLP', skips=[4], netdepth=8, netwidth=256, netchunk=1024 * 32, output_ch=5, use_viewdirs=use_viewdirs, embedder=dict( type='BaseEmbedder', i_embed=0, multires=10, multires_dirs=4, ), ), render=dict( type='NerfRender', # render cloass name white_bkgd=white_bkgd, # see above raw_noise_std=0, # std dev of noise added to regularize sigma_a output, 1e0 recommended ), )
-
data define network structure, a network is usually composed of embedder, mlp and render.
basedata_cfg = dict( dataset_type=dataset_type, datadir='data/nerf_synthetic/#DATANAME#', half_res=True, # load blender synthetic data at 400x400 instead of 800x800 testskip= 8, # will load 1/N images from test/val sets, useful for large datasets like deepvoxels white_bkgd=white_bkgd, is_batching=False, # True for blender, False for llff mode='train', ) traindata_cfg = basedata_cfg.copy() valdata_cfg = basedata_cfg.copy() testdata_cfg = basedata_cfg.copy() traindata_cfg.update(dict()) valdata_cfg.update(dict(mode='val')) testdata_cfg.update(dict(mode='test', testskip=0)) train_pipeline = [ dict(type='Sample'), dict(type='DeleteUseless', keys=['images', 'poses', 'i_data', 'idx']), dict(type='ToTensor', keys=['pose', 'target_s']), dict(type='GetRays'), dict(type='SelectRays', sel_n=N_rand_per_sampler, precrop_iters=500, precrop_frac=0.5), # in the first 500 iter, select rays inside center of image dict(type='GetViewdirs', enable=use_viewdirs), dict(type='ToNDC', enable=(not no_ndc)), dict(type='GetBounds'), dict(type='GetZvals', lindisp=lindisp, N_samples=N_samples), # N_samples: number of coarse samples per ray dict(type='PerturbZvals', enable=is_perturb), dict(type='GetPts'), dict(type='DeleteUseless', keys=['pose', 'iter_n']), ] test_pipeline = [ dict(type='ToTensor', keys=['pose']), dict(type='GetRays'), dict(type='FlattenRays'), dict(type='GetViewdirs', enable=use_viewdirs), dict(type='ToNDC', enable=(not no_ndc)), dict(type='GetBounds'), dict(type='GetZvals', lindisp=lindisp, N_samples=N_samples), dict(type='PerturbZvals', enable=False), # do not perturb when test dict(type='GetPts'), dict(type='DeleteUseless', keys=['pose']), ] data = dict( train_loader=dict(batch_size=1, num_workers=4), train=dict( type='SceneBaseDataset', cfg=traindata_cfg, pipeline=train_pipeline, ), val_loader=dict(batch_size=1, num_workers=0), val=dict( type='SceneBaseDataset', cfg=valdata_cfg, pipeline=test_pipeline, ), test_loader=dict(batch_size=1, num_workers=0), test=dict( type='SceneBaseDataset', cfg=testdata_cfg, pipeline=test_pipeline, # same pipeline as validation ), )