seed: random seed for reproducibility.
checkpoint: path to a checkpoint to load.
lr: learning rate.
weight_decay: weight decay.
optimizer_type: optimizer type. Choose from adam, lamb, sgd, adam_w.
lr_schedulers_type: learning rate scheduler type. Choose from step, plateau, cosine, poly.
differential_lr: Whether to apply a learning rate multiplier to the backbone.
differential_lr_scaling: LR multiplier for backbone.
x_range, y_range, z_range: Range of the input point cloud.
voxel_size: Voxel size of the input point cloud.
num_queries: Number of queries for Mask2Former. Should be higher than the number of objects per scan.
predict_height: Whether to predict the height of the objects.
pc_point_dim: Number of dimensions of the input point cloud.
max_num_points: Maximum number of points per voxel to keep at the encoder stage.
encoder_feat_channels: Number of channels of the encoder features.
encoder_encoding_type: Encoding type of the encoder. Choose from vanilla, fourier, cosine.
encoder_fourier_enc_group: Number of groups for the Fourier encoding.
backbone_embed_dim: Embedding dimension of the backbone.
backbone_use_abs_emb: Whether to use absolute positional embeddings in the backbone.
backbone_path_size: Patch size of the Swin Transformer backbone.
backbone_window_size: Window size of the Swin Transformer backbone.
backbone_strides: Strides of the Swin Transformer backbone (int, int, int, int).
backbone_swap_dims: Whether to swap the dimensions of the Swin Transformer backbone.
head_feat_channels: Number of channels of the head features.
head_out_channels: Number of channels of the head output.
head_reverse_class_weight: Whether to reverse the class weight. This changes the default Mask2Former weighting of the
background class.
head_num_classes: Number of classes of the head output.
min_num_inst_pixels: Minimum number of pixels to keep an instance in the dataset.
batch_size: Batch size.
test_batch_size: Batch size for testing.
num_workers: Number of workers for the data loader.
test_num_workers: Number of workers for the data loader for testing.
pin_memory: Whether to pin memory for the data loader.
remove_unseen: Whether to remove instances that do not appear in a scan.
shuffle_train: Whether to shuffle the training data.
min_num_points: Minimum number of points to keep an instance in a scan.
log_every_n_step: Logging frequency during training.
limit_train_batches: Maximum number of batches to train on.
limit_val_batches: Maximum number of batches to validate on.
Available augmentations are available in maskbev.augmentations.
Select the ones you want using
augmentations:
- name: <augmentation name>
<augmentation parameters>
- name: <augmentation name>
<augmentation parameters>