conf_converter.py

class ConfConverter:
    def __init__(self, filename):
        self.filename = filename
        self.f = open(filename, "w")
    
    def __del__(self):
        self.f.close()
    
    def add_argument(self, name, type=None, default=None, required=None, choices=[], help="", nargs=None, action=None):
        f = self.f
        key = name.lstrip("--")
        f.write(f"{key}: ")
        if default is not None:
            if nargs == "+":
                vals = [str(v) for v in default]
                if type == str:
                    vals = ['"' + v + '"' for v in default]
                f.write("[" + ", ".join(vals) + "]")
            elif type == int or type == float:
                f.write(str(default))
            elif type == str:
                f.write(f"\"" + default + "\"")
            elif type == None:
                if default == True:
                    f.write("true")
                elif default == False:
                    f.write("false")
                else:
                    assert False
        if help:
            help = help.replace("\n", " ")
            help = " ".join(help.split())
            f.write(f" # {help}")
        f.write("\n")

    def parse_args(self):
        pass

config.yaml

root_dir:  # root directory of dataset
dataset_name: "blender" # which dataset to train/val
data_perturb: [] # what perturbation to add to data. Available choices: [], ["color"], ["occ"] or ["color", "occ"]
img_wh: [800, 800] # resolution (img_w, img_h) of the image
img_downscale: 1 # how much to downscale the images for phototourism dataset
use_cache: false # whether to use ray cache (make sure img_downscale is the same)
N_emb_xyz: 10 # number of xyz embedding frequencies
N_emb_dir: 4 # number of direction embedding frequencies
N_samples: 64 # number of coarse samples
N_importance: 128 # number of additional fine samples
use_disp: false # use disparity depth sampling
perturb: 1.0 # factor to perturb depth sampling points
noise_std: 1.0 # std dev of noise added to regularize sigma
N_vocab: 100 # number of vocabulary (number of images) in the dataset for nn.Embedding
encode_a: false # whether to encode appearance (NeRF-A)
N_a: 48 # number of embeddings for appearance
encode_t: false # whether to encode transient object (NeRF-U)
N_tau: 16 # number of embeddings for transient objects
beta_min: 0.1 # minimum color variance for each ray
optimise_cameras: false # optimise camera poses
cameras_lr: 0.001 # learning rate
cameras_lr_exp_gamma: 0.794 # learning rate
barf: false # BARF-style annealing
barf_start_epoch: 4 # epoch to end annealing
barf_end_epoch: 8 # epoch to end annealing
batch_size: 1024 # batch size
chunk: 32768 # chunk size to split the input to avoid OOM
num_epochs: 16 # number of training epochs
num_gpus: 1 # number of gpus
ckpt_path:  # pretrained checkpoint path to load
prefixes_to_ignore: ["loss"] # the prefixes to ignore in the checkpoint state dict
optimizer: "adam" # optimizer type
lr: 0.0005 # learning rate
lr_end: 0.0005 # learning rate
momentum: 0.9 # learning rate momentum
weight_decay: 0 # weight decay
lr_scheduler: "steplr" # scheduler type
warmup_multiplier: 1.0 # lr is multiplied by this factor after --warmup_epochs
warmup_epochs: 0 # Gradually warm-up(increasing) learning rate in optimizer
decay_step: [20] # scheduler decay step
decay_gamma: 0.1 # learning rate decay amount
poly_exp: 0.9 # exponent for polynomial learning rate decay
exp_name: "exp" # experiment name
refresh_every: 1 # print the progress bar every X steps

opt_conv.py

# import argparse
from conf_converter import ConfConverter


def get_opts():
    # parser = argparse.ArgumentParser()
    parser = ConvConverter("config.yaml")

    parser.add_argument('--root_dir', type=str, required=True,
                        help='root directory of dataset')
    parser.add_argument('--dataset_name', type=str, default='blender',
                        choices=['blender', 'phototourism', 'car'],
                        help='which dataset to train/val')
    # for blender
    parser.add_argument('--data_perturb', nargs="+", type=str, default=[],
                        help='''what perturbation to add to data.
                                Available choices: [], ["color"], ["occ"] or ["color", "occ"]
                             ''')
    parser.add_argument('--img_wh', nargs="+", type=int, default=[800, 800],
                        help='resolution (img_w, img_h) of the image')
    # for phototourism
    parser.add_argument('--img_downscale', type=int, default=1,
                        help='how much to downscale the images for phototourism dataset')
    parser.add_argument('--use_cache', default=False, action="store_true",
                        help='whether to use ray cache (make sure img_downscale is the same)')

    # original NeRF parameters
    parser.add_argument('--N_emb_xyz', type=int, default=10,
                        help='number of xyz embedding frequencies')
    parser.add_argument('--N_emb_dir', type=int, default=4,
                        help='number of direction embedding frequencies')
    parser.add_argument('--N_samples', type=int, default=64,
                        help='number of coarse samples')
    parser.add_argument('--N_importance', type=int, default=128,
                        help='number of additional fine samples')
    parser.add_argument('--use_disp', default=False, action="store_true",
                        help='use disparity depth sampling')
    parser.add_argument('--perturb', type=float, default=1.0,
                        help='factor to perturb depth sampling points')
    parser.add_argument('--noise_std', type=float, default=1.0,
                        help='std dev of noise added to regularize sigma')

    # NeRF-W parameters
    parser.add_argument('--N_vocab', type=int, default=100,
                        help='''number of vocabulary (number of images) 
                                in the dataset for nn.Embedding''')
    parser.add_argument('--encode_a', default=False, action="store_true",
                        help='whether to encode appearance (NeRF-A)')
    parser.add_argument('--N_a', type=int, default=48,
                        help='number of embeddings for appearance')
    parser.add_argument('--encode_t', default=False, action="store_true",
                        help='whether to encode transient object (NeRF-U)')
    parser.add_argument('--N_tau', type=int, default=16,
                        help='number of embeddings for transient objects')
    parser.add_argument('--beta_min', type=float, default=0.1,
                        help='minimum color variance for each ray')

    # BARF parameters
    parser.add_argument('--optimise_cameras', default=False, action="store_true",
                        help='optimise camera poses')
    parser.add_argument('--cameras_lr', type=float, default=1e-3,
                        help='learning rate')
    parser.add_argument('--cameras_lr_exp_gamma', type=float, default=0.794,
                        help='learning rate')
    parser.add_argument('--barf', default=False, action="store_true",
                        help='BARF-style annealing')
    parser.add_argument('--barf_start_epoch', type=int, default=4,
                        help='epoch to end annealing')
    parser.add_argument('--barf_end_epoch', type=int, default=8,
                        help='epoch to end annealing')

    parser.add_argument('--batch_size', type=int, default=1024,
                        help='batch size')
    parser.add_argument('--chunk', type=int, default=32*1024,
                        help='chunk size to split the input to avoid OOM')
    parser.add_argument('--num_epochs', type=int, default=16,
                        help='number of training epochs')
    parser.add_argument('--num_gpus', type=int, default=1,
                        help='number of gpus')

    parser.add_argument('--ckpt_path', type=str, default=None,
                        help='pretrained checkpoint path to load')
    parser.add_argument('--prefixes_to_ignore', nargs='+', type=str, default=['loss'],
                        help='the prefixes to ignore in the checkpoint state dict')

    parser.add_argument('--optimizer', type=str, default='adam',
                        help='optimizer type',
                        choices=['sgd', 'adam', 'radam', 'ranger'])
    parser.add_argument('--lr', type=float, default=5e-4,
                        help='learning rate')
    parser.add_argument('--lr_end', type=float, default=5e-4,
                        help='learning rate')
    parser.add_argument('--momentum', type=float, default=0.9,
                        help='learning rate momentum')
    parser.add_argument('--weight_decay', type=float, default=0,
                        help='weight decay')
    parser.add_argument('--lr_scheduler', type=str, default='steplr',
                        help='scheduler type',
                        choices=['steplr', 'cosine', 'poly', 'exp_interval'])
    #### params for warmup, only applied when optimizer == 'sgd' or 'adam'
    parser.add_argument('--warmup_multiplier', type=float, default=1.0,
                        help='lr is multiplied by this factor after --warmup_epochs')
    parser.add_argument('--warmup_epochs', type=int, default=0,
                        help='Gradually warm-up(increasing) learning rate in optimizer')
    ###########################
    #### params for steplr ####
    parser.add_argument('--decay_step', nargs='+', type=int, default=[20],
                        help='scheduler decay step')
    parser.add_argument('--decay_gamma', type=float, default=0.1,
                        help='learning rate decay amount')
    ###########################
    #### params for poly ####
    parser.add_argument('--poly_exp', type=float, default=0.9,
                        help='exponent for polynomial learning rate decay')
    ###########################

    parser.add_argument('--exp_name', type=str, default='exp',
                        help='experiment name')
    parser.add_argument('--refresh_every', type=int, default=1,
                        help='print the progress bar every X steps')

    return parser.parse_args()


get_opts()