-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathoptions.py
118 lines (106 loc) · 10.1 KB
/
options.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
import configargparse
from munch import *
from pdb import set_trace as st
class BaseOptions():
def __init__(self):
self.parser = configargparse.ArgumentParser()
self.initialized = False
def initialize(self):
# Dataset options
dataset = self.parser.add_argument_group('dataset')
dataset.add_argument("--dataset_path", type=str, default='./datasets/FFHQ', help="path to the lmdb dataset")
# Experiment Options
experiment = self.parser.add_argument_group('experiment')
experiment.add_argument('--config', is_config_file=True, help='config file path')
experiment.add_argument("--expname", type=str, default='debug', help='experiment name')
experiment.add_argument("--ckpt", type=str, default='300000', help="path to the checkpoints to resume training")
experiment.add_argument("--continue_training", action="store_true", help="continue training the model")
# Training loop options
training = self.parser.add_argument_group('training')
training.add_argument("--checkpoints_dir", type=str, default='./checkpoints', help='checkpoints directory name')
training.add_argument("--trained_ckpt", type=str, default='', help='trained model path')
training.add_argument("--pretrain_render_path", type=str, default='', help='pretrain render model path')
training.add_argument("--iter", type=int, default=300000, help="total number of training iterations")
training.add_argument("--batch", type=int, default=4, help="batch sizes for each GPU. A single RTX2080 can fit batch=4, chunck=1 into memory.")
training.add_argument("--chunk", type=int, default=4, help='number of samples within a batch to processed in parallel, decrease if running out of memory')
training.add_argument("--val_n_sample", type=int, default=8, help="number of test samples generated during training")
training.add_argument("--d_reg_every", type=int, default=16, help="interval for applying r1 regularization to the StyleGAN generator")
training.add_argument("--g_reg_every", type=int, default=4, help="interval for applying path length regularization to the StyleGAN generator")
training.add_argument("--local_rank", type=int, default=0, help="local rank for distributed training")
training.add_argument("--mixing", type=float, default=0.9, help="probability of latent code mixing")
training.add_argument("--lr", type=float, default=0.002, help="learning rate")
training.add_argument("--Semantic_D_lambda", type=float, default=1, help="weight of semantic-discriminator loss")
training.add_argument("--r1_img", type=float, default=10, help="weight of the r1 regularization of img")
training.add_argument("--r1_seg", type=float, default=1000, help="weight of the r1 regularization of seg")
training.add_argument("--view_lambda", type=float, default=15, help="weight of the viewpoint regularization")
training.add_argument("--eikonal_lambda", type=float, default=0.1, help="weight of the eikonal regularization")
training.add_argument("--min_surf_lambda", type=float, default=0.001, help="weight of the minimal surface regularization")
training.add_argument("--min_surf_beta", type=float, default=100.0, help="weight of the minimal surface regularization")
training.add_argument("--path_regularize", type=float, default=2, help="weight of the path length regularization")
training.add_argument("--path_batch_shrink", type=int, default=2, help="batch size reducing factor for the path length regularization (reduce memory consumption)")
training.add_argument("--wandb", action="store_true", help="use weights and biases logging")
training.add_argument("--no_sphere_init", action="store_true", help="do not initialize the volume renderer with a sphere SDF")
# Inference Options
inference = self.parser.add_argument_group('inference')
inference.add_argument("--results_dir", type=str, default='./evaluations', help='results/evaluations directory name')
inference.add_argument("--semantics", type=str, default='', help='generate only some local semantics')
inference.add_argument("--truncation_ratio", type=float, default=0.7, help="truncation ratio, controls the diversity vs. quality tradeoff. Higher truncation ratio would generate more diverse results")
inference.add_argument("--truncation_mean", type=int, default=10000, help="number of vectors to calculate mean for the truncation")
inference.add_argument("--identities", type=int, default=16, help="number of identities to be generated")
inference.add_argument("--num_views_per_id", type=int, default=1, help="number of viewpoints generated per identity")
inference.add_argument("--no_surface_renderings", action="store_true", help="when true, only RGB outputs will be generated. otherwise, both RGB and depth videos/renderings will be generated. this cuts the processing time per video")
inference.add_argument("--fixed_camera_angles", action="store_true", help="when true, the generator will render indentities from a fixed set of camera angles.")
inference.add_argument("--azim_video", action="store_true", help="when true, the camera trajectory will travel along the azimuth direction. Otherwise, the camera will travel along an ellipsoid trajectory.")
inference.add_argument("--flag", type=str, default='', help="set output name")
# Generator options
model = self.parser.add_argument_group('model')
model.add_argument("--size", type=int, default=256, help="image sizes for the model")
model.add_argument("--style_dim", type=int, default=256, help="number of style input dimensions")
model.add_argument("--channel_multiplier", type=int, default=2, help="channel multiplier factor for the StyleGAN decoder. config-f = 2, else = 1")
model.add_argument("--n_mlp", type=int, default=8, help="number of mlp layers in stylegan's mapping network")
model.add_argument("--lr_mapping", type=float, default=0.01, help='learning rate reduction for mapping network MLP layers')
model.add_argument("--renderer_spatial_output_dim", type=int, default=64, help='spatial resolution of the StyleGAN decoder inputs')
model.add_argument("--project_noise", action='store_true', help='when true, use geometry-aware noise projection to reduce flickering effects (see supplementary section C.1 in the paper). warning: processing time significantly increases with this flag to ~20 minutes per video.')
model.add_argument("--n_wcode", type=int, default=9, help="the number of semantic mask")
# Camera options
camera = self.parser.add_argument_group('camera')
camera.add_argument("--uniform", action="store_true", help="when true, the camera position is sampled from uniform distribution. Gaussian distribution is the default")
camera.add_argument("--azim", type=float, default=0.3, help="camera azimuth angle std/range in Radians")
camera.add_argument("--elev", type=float, default=0.15, help="camera elevation angle std/range in Radians")
camera.add_argument("--fov", type=float, default=6, help="camera field of view half angle in Degrees")
camera.add_argument("--dist_radius", type=float, default=0.12, help="radius of points sampling distance from the origin. determines the near and far fields")
# Volume Renderer options
rendering = self.parser.add_argument_group('rendering')
# MLP model parameters
rendering.add_argument("--n_render", type=int, default=9, help="the number of local semantic 3d generator")
rendering.add_argument("--depth", type=int, default=3, help='layers in network')
rendering.add_argument("--width", type=int, default=128, help='channels per layer')
# Volume representation options
rendering.add_argument("--no_sdf", action='store_true', help='By default, the raw MLP outputs represent an underline signed distance field (SDF). When true, the MLP outputs represent the traditional NeRF density field.')
rendering.add_argument("--no_z_normalize", action='store_true', help='By default, the model normalizes input coordinates such that the z coordinate is in [-1,1]. When true that feature is disabled.')
rendering.add_argument("--static_viewdirs", action='store_true', help='when true, use static viewing direction input to the MLP')
# Ray intergration options
rendering.add_argument("--N_samples", type=int, default=24, help='number of samples per ray')
rendering.add_argument("--no_offset_sampling", action='store_true', help='when true, use random stratified sampling when rendering the volume, otherwise offset sampling is used. (See Equation (3) in Sec. 3.2 of the paper)')
rendering.add_argument("--perturb", type=float, default=1., help='set to 0. for no jitter, 1. for jitter')
rendering.add_argument("--raw_noise_std", type=float, default=0., help='std dev of noise added to regularize sigma_a output, 1e0 recommended')
rendering.add_argument("--force_background", action='store_true', help='force the last depth sample to act as background in case of a transparent ray')
# Set volume renderer outputs
rendering.add_argument("--return_xyz", action='store_true', help='when true, the volume renderer also returns the xyz point could of the surface. This point cloud is used to produce depth map renderings')
rendering.add_argument("--return_sdf", action='store_true', help='when true, the volume renderer also returns the SDF network outputs for each location in the volume')
self.initialized = True
def parse(self):
self.opt = Munch()
if not self.initialized:
self.initialize()
try:
args = self.parser.parse_args()
except:
args = self.parser.parse_args(args=[])
for group in self.parser._action_groups[2:]:
title = group.title
self.opt[title] = Munch()
for action in group._group_actions:
dest = action.dest
self.opt[title][dest] = args.__getattribute__(dest)
return self.opt