Added config and hyperparameter files

config/config_datasets.yaml

@@ -0,0 +1,65 @@
+# General ------------------------------------
+## Deployment ------------
+horizontal_field_of_view: [ -179.9, 179.9 ] # Always this for rotating LiDAR sensors
+
+## Preprocessing ---------
+### Neighborhood estimation
+min_num_points_in_neighborhood_to_determine_point_class: 10 # neighborhood below this threshold --> no normal computed
+epsilon_plane: 0.01 # 1e-2, this is the epsilon value when eigenvalues are considered to describe a plane
+epsilon_line: 0.01 # 1e-2, -"- to describe a line
+epsilon_range: 0.5 # Epsilon when points are removed from patch used for normal estimation
+
+# KITTI --------------------------------------
+kitti: # must be the same as dataset name specified in deployment_options.yaml
+  ## Deployment --------
+  ### Training/testing configuration
+  training_identifiers: [ 0 , 1, 2, 3, 4, 5, 6, 7, 8 ]
+  testing_identifiers: [ 9, 10 ]
+  ### Sensor specs
+  vertical_field_of_view: [ -24.5, 2.0 ]  # taken from the datasheet, in degrees
+  vertical_cells: 64 # Number of LiDAR rings
+  horizontal_cells: 720 # Can be defined, should not be larger than horizontal resolution of LiDAR
+  ### Pre-processed file
+  preprocessed_path: "./datasets/kitti/preprocessed/sequences"
+  ### Ground-truth poses
+  pose_data_path: "./datasets/kitti/data_odometry_poses/dataset/poses"
+
+  ## Pre-processing ----
+  ### Dataset type, currently supported: "kitti" and "rosbag"
+  dataset_type: "kitti"
+  ### Specific configurations needed for pre-processing
+  horizontal_cells_preprocessing: 2250 # Calculated from 10Hz frequency of LiDAR
+  ### Preprocessing hyperparameters
+  neighborhood_side_length: [ 7, 11 ] # Patch dimensions used for normal estimation
+  ### Source files (before pre-processing)
+  data_path: "./datasets/kitti/data_odometry_velodyne/dataset/sequences" # Path to dataset
+  rosbag_path: "./datasets/kitti/rosbags/" # Only needed when converted to rosbag
+
+# DARPA --------------------------------------
+darpa:
+  ## Deployment --------
+  ### Training/testing configuration
+  training_identifiers: [ 0 ]
+  testing_identifiers: [ 0 ]
+  ### Sensor specs
+  vertical_field_of_view: [ -22.5, 22.5 ]
+  vertical_cells: 64
+  horizontal_cells: 512
+  ### Pre-processed file
+  preprocessed_path: "./datasets/darpa/preprocessed/sequences"
+  ### Ground-truth poses
+  pose_data_path: null
+
+  ## Pre-processing ----
+  ### Dataset type, currently supported: "kitti" and "rosbag"
+  dataset_type: "rosbag"
+  topic: "/sherman/lidar_points"
+  ### Specific configurations needed for pre-processing
+  horizontal_cells_preprocessing: 512
+  ### Preprocessing hyperparameters
+  neighborhood_side_length: [ 7, 11 ]
+  ### Source files (before pre-processing)
+  data_path: "./datasets/darpa"
+
+# Custom dataset
+custom: # etc.

config/deployment_options.yaml

@@ -0,0 +1,21 @@
+# Dataset
+## Selection
+datasets: ["kitti"] # Can contain multiple datasets, e.g. ["darpa", "kitti"]
+mode: "training" # Valid options: "training" and "testing", will also be set by training/testing scripts
+experiment: "trainings_experiments_1"
+
+# System information
+device: "cuda" # "cuda", "cpu", if multiple GPUs: select cuda device via CUDA_VISIBLE_DEVICES=...
+store_dataset_in_RAM: False # If false, each scan and normal file will be read from disk
+                            # For KITTI requires approximately 50GB RAM/SWAP
+num_dataloader_workers: 0 # default: 0
+
+## Comfort options
+unsupervised_at_start: False # if False, it starts to train supervised in the first epoch (Identity matching)
+inference_only: True # During testing, it does not compute any losses, during training: no impact
+log_normals: True
+visualize_images: True
+visualize_single_img_preprocessing: False
+
+## Performance options
+use_jit: False # Only thoroughly tested for inference

config/hyperparameters.yaml

@@ -0,0 +1,33 @@
+# Training Parameters -------------------------------
+## Loss Computation
+batch_size: 1 # batch_size > 1 currently only supported if single image dims are used (vertical and horizontal cells)
+# In general: larger batches currently implemented rather primitively
+learning_rate: 0.00001 #1e-5
+lambda_po2pl: 1.0 # 10.0
+## Network training options
+use_dropout: False
+## Data Augmentation
+random_point_cloud_rotations: False # Data augmentation by rotating point cloud during training
+random_rotations_only_yaw: False # If set to true, and random_point_cloud_rotations set to true --> rotations only about yaw axis
+magnitude_random_rot: 4.0 # in degrees
+## Loss Functions
+normal_loss: "squared" # "squared" or "linear"
+rotation_losses_only_plane_to_plane: False
+point_to_point_loss: False
+point_to_plane_loss: True
+plane_to_plane_loss: True
+po2po_alone: False
+
+# Network Parameters ---------------------------------
+## Data Format
+normalization_scaling: False # If set to true --> lambda_po2pl should be increased
+## Model parameters
+###############num_input_channels: 8 # No impact yet
+activation_fct: "tanh" # "tanh" or "relu"
+resnet_outputs: 1000 # Output of the resnet, then fed into the rot/trans fully connected networks
+###############conv_size: [3, 5] # No impact yet
+pre_feature_extraction: False # Additional feature extraction layers before concatenation of input
+layers: [ 2, 2, 2, 2 ] # 4 down-sampling steps, num resnet blocks (2 layers each) per down-sampling (2 here)
+                       # [ 1, 1, 1, 1 ] e.g. results in half as many layers
+factor_fewer_resnet_channels: 1 # 4 leads to approximately 15 times fewer weights
+use_single_mlp_at_output: False # Single big MLP at output