name: '' # Run name
debug: False # Debugging flag
arch:
seed: 42 # Random seed for Pytorch/Numpy initialization
min_epochs: 1 # Minimum number of epochs
max_epochs: 50 # Maximum number of epochs
checkpoint:
filepath: '' # Checkpoint filepath to save data
save_top_k: 5 # Number of best models to save
monitor: 'loss' # Metric to monitor for logging
monitor_index: 0 # Dataset index for the metric to monitor
mode: 'auto' # Automatically determine direction of improvement (increase or decrease)
s3_path: '' # s3 path for AWS model syncing
s3_frequency: 1 # How often to s3 sync
save:
folder: '' # Folder where data will be saved
viz: True # Flag for saving inverse depth map visualization
npz: True # Flag for saving numpy depth maps
wandb:
dry_run: True # Wandb dry-run (not logging)
name: '' # Wandb run name
project: os.environ.get("WANDB_PROJECT", "") # Wandb project
entity: os.environ.get("WANDB_ENTITY", "") # Wandb entity
tags: [] # Wandb tags
dir: '' # Wandb save folder
model:
name: '' # Training model
checkpoint_path: '' # Checkpoint path for model saving
optimizer:
name: 'Adam' # Optimizer name
depth:
lr: 0.0002 # Depth learning rate
weight_decay: 0.0 # Dept weight decay
pose:
lr: 0.0002 # Pose learning rate
weight_decay: 0.0 # Pose weight decay
scheduler:
name: 'StepLR' # Scheduler name
step_size: 10 # Scheduler step size
gamma: 0.5 # Scheduler gamma value
T_max: 20 # Scheduler maximum number of iterations
params:
crop: '' # Which crop should be used during evaluation
min_depth: 0.0 # Minimum depth value to evaluate
max_depth: 80.0 # Maximum depth value to evaluate
loss:
num_scales: 4 # Number of inverse depth scales to use
progressive_scaling: 0.0 # Training percentage to decay number of scales
flip_lr_prob: 0.5 # Probablity of horizontal flippping
rotation_mode: 'euler' # Rotation mode
upsample_depth_maps: True # Resize depth maps to highest resolution
ssim_loss_weight: 0.85 # SSIM loss weight
occ_reg_weight: 0.1 # Occlusion regularizer loss weight
smooth_loss_weight: 0.001 # Smoothness loss weight
C1: 1e-4 # SSIM parameter
C2: 9e-4 # SSIM parameter
photometric_reduce_op: 'min' # Method for photometric loss reducing
disp_norm: True # Inverse depth normalization
clip_loss: 0.0 # Clip loss threshold variance
padding_mode: 'zeros' # Photometric loss padding mode
automask_loss: True # Automasking to remove static pixels
supervised_method: 'sparse-l1' # Method for depth supervision
supervised_num_scales: 4 # Number of scales for supervised learning
supervised_loss_weight: 0.9 # Supervised loss weight
depth_net:
name: '' # Depth network name
checkpoint_path: '' # Depth checkpoint filepath
version: '' # Depth network version
dropout: 0.0 # Depth network dropout
pose_net:
name: '' # Pose network name
checkpoint_path: '' # Pose checkpoint filepath
version: '' # Pose network version
dropout: 0.0 # Pose network dropout
datasets:
augmentation:
image_shape: (192, 640) # Image shape
jittering: (0.2, 0.2, 0.2, 0.05) # Color jittering values
train:
batch_size: 8 # Training batch size
num_workers: 16 # Training number of workers
back_context: 1 # Training backward context
forward_context: 1 # Training forward context
dataset: [] # Training dataset
path: [] # Training data path
split: [] # Training split
depth_type: [''] # Training depth type
cameras: [] # Training cameras
repeat: [1] # Number of times training dataset is repeated per epoch
num_logs: 5 # Number of training images to log
validation:
batch_size: 1 # Validation batch size
num_workers: 8 # Validation number of workers
back_context: 0 # Validation backward context
forward_context: 0 # Validation forward contxt
dataset: [] # Validation dataset
path: [] # Validation data path
split: [] # Validation split
depth_type: [''] # Validation depth type
cameras: [] # Validation cameras
num_logs: 5 # Number of validation images to log
test:
batch_size: 1 # Test batch size
num_workers: 8 # Test number of workers
back_context: 0 # Test backward context
forward_context: 0 # Test forward context
dataset: [] # Test dataset
path: [] # Test data path
split: [] # Test split
depth_type: [''] # Test depth type
cameras: [] # Test cameras
num_logs: 5 # Number of test images to log