Let us take train_RDDN_ShapeNet_airplane.yaml as an example.
name: RDDN_PointNetSeg_CD_shapenet_airplane_normcol_normpoint_c512 # experiment name
use_tb_logger: true # whether to use tensorboardX logger
gpu_id: [0] # GPU ID to use
type: RDNV1 # solver type, see more in code/solver/__init__.py
dataset:
train:
name: Raw ShapeNet Airplane Train # train set name
mode: raw_pair_shapenet_v0 # train dataset type, see more in code/data/__init__.py
root: /sailhome/jiangthu/jzy/projects/CycleConsistentDeformation/data/dataset_shapenet/ # path to full data
category: Airplane # data category
num_point: 2048 # number of points in sampled point cloud
phase: train # pahse: train | val | test
num_worker: 12 # number of threads for data loading
batch_size: 32 # input batch size
norm: True # whether to normalize input point cloud
val:
name: Raw ShapeNet Airplane Test # val set name
mode: raw_pair_shapenet_v0 # val dataset type, see more in code/data/__init__.py
root: /sailhome/jiangthu/jzy/projects/CycleConsistentDeformation/data/dataset_shapenet/ # path to full data
category: Airplane # data category
num_point: 2048 # number of points in sampled point cloud
phase: val # pahse: train | val | test
norm: True # whether to normalize input point cloud
model:
init: xavier # network parameter initialization method, currently only xavier is supported
model_type: RDDNV0 # network model type, see more in code/model/__init__.py
dict: # dictionary net options
version: 0 # dict net version, currently only 0 is supported
arch: PointNetSeg # dict net architecture: MLP | PointNetSeg
feature_dim: 1536 # dict net output column space dimension, should be 3X of coeff net out dimension
norm_column: True # whether to normalize dict columns
coeff: # coefficient net options
version: 0 # coeff net version, 0(output difference of dst and src coeff) | 1(compute coeff from the concatenation of src and dst feature)
arch: PointNetCls # coeff net architecture: PointNetCls for v0, PointNetMix for v1
out_dim: 512 # coeff net output dimention
train:
learning_rate: 0.0005 # learning rate
loss: # losses
fit_CD: # fitting loss
loss_type: CD # loss type: CD | EMD | l2
weight: 100 # loss weight
sym_CD: # symmetry loss
loss_type: CD # loss type: CD | EMD | l2
weight: 100 # loss weight
sym_axis: 2 # symmetry axis
lr_gamma: 0.5 # multiplicative factor of learning rate decay
lr_scheme: MultiStepLR # lr scheduler scheme, currently only MultiStepLR is supprted
lr_steps: [4.0e4, 8.0e4, 1.2e5] # lr decay milestones
niter: 1.5e5 # number of total iterations
save_freq: 5.0e3 # model saving frequency
val_freq: 5.0e3 # validation frequency
val_metric: loss_fit_CD # metric for best model
weight_decay: 0 # weight decay
logger:
print_freq: 500 # log frequency
num_save_image: 10 # number of images to save during validation
path:
root: ../ # project root