The parameters present in a config.json file allow one to configure a training session. Each of these parameters is described below:
Where to find and how to handle the data
- classes: list[string] A list of class names representing the subset of classes to train with all other classes will be masked
- original_file_suffix: string The suffix of all PLY files present in the dataset featuring non-uniform point sampling
- path: string An absolute path to the root of the dataset
- refresh_cache: bool Whether to invalidate an existing cache and create it again from scratch
- special_weights:
- empty: float The weight given to empty unoccupied voxels in the loss function
- masked: float The weight given to masked voxels in the loss function
- training_samples:
- num_points: int Number of points present in a training sample
- spatial_size: list[float] The spatial size in x, y and z dimensions of a training sample. Must be divisible by 0.6
- uniform_file_suffix: string The suffix of all PLY files present in the dataset featuring uniform point sampling use ClassyVoxelizer to uniformly sample meshes/point clouds
- validity_thresholds:
- empty: float The maximum percentage of empty volume relative to a sample's entire volume permitted for that sample to be considered valid
- empty: float The maximum percentage of masked points relative to a sample's entire points permitted for that sample to be considered valid
Hyperparameters of the model
- filters:
- abstraction:
- points_to_15cm: list[int] Depths of the filters in the simplified PointNets
- 15cm_to_30cm: list[int] Depths of the 3D convolution and following 1x1x1 convolution filters abstracting from the 15cm to the 30cm spatial scale
- 30cm_to_60cm: list[int] Depths of the 3D convolution and following 1x1x1 convolution filters abstracting from the 30cm to the 60cm spatial scale
- skip:
- 15cm: int Filter depth of the skip connection propagating features at the 15cm spatial scale
- 30cm: int Filter depth of the skip connection propagating features at the 30cm spatial scale
- 45cm: int Filter depth of the skip connection propagating features at the 45cm spatial scale
- 60cm: int Filter depth of the skip connection propagating features at the 60cm spatial scale
- 90cm: int Filter depth of the skip connection propagating features at the 90cm spatial scale
- 180cm: int Filter depth of the skip connection propagating features at the 180cm spatial scale
- spatial_pool: int Filter depth of the skip connection propagating spatially-pooled features at the 60cm spatial scale
- upsampling:
- 60cm_to_30cm: list[int] Depths of the 3D deconvolution and following 1x1x1 convolution filters upsampling from the 60cm to the 30cm spatial scale
- 30cm_to_15cm: list[int] Depths of the 3D convolution and following 1x1x1 convolution filters abstracting from the 30cm to the 15cm spatial scale
- 15cm_to_5cm: list[int] Depths of the 3D convolution and following 1x1x1 convolution filters abstracting from the 15cm to the 5cm spatial scale
- abstraction:
- pointnet:
- neighbors: int Number of neighboring points to group around a PointNet's center
- spacing: float Distance between the centers of two PointNets
- spatial_pool_radius: float The radial distance from the center of a voxel where the responses of its neighboring voxels will be weighted the highest
How training will be performed
- batch_size: int Number of samples in a batch
- checkpoint_path: string Path to a Tensorflow checkpoint. If non-empty, this new training session will start with weights from the specified checkpoint
- checkpoints_to_keep: int Number of checkpoints to keep in session directory before deleting
- data_augmentation:
- jitter:
- max: float Maximum distance permitted to move a point
- sigma: float Sigma value of the normal distribution used to jitter points
- random_dropout:
- max: float Maximum permitted percentage of a training sample that can be removed
- random_rotation: bool Whether to randomly rotate a training sample about the +Z axis
- shuffle: bool Whether to shuffle the samples in the training set at the beginning of each epoch
- jitter:
- description: string Textual description of the training session
- finetune_on_new_classes: bool Set to true when finetuning a pretrained model on a new dataset with a different set of trainable classes. If true, the weights of the model's final layer will not be loaded
- gpu:
- allow_growth: bool Whether to incrementally reserve more memory on the GPU as opposed to allocating all available memory from the beginning
- allow_soft_placement: bool Whether to permit certain ops being executed by another device cpu/gpu if its not possible to execute it on the preferred device
- id: int Numerical identifier of the GPU to use for training
- log_every_n_batches: int Number of batches after which to log the current state of the training
- max_epochs: int Maximum number of epochs to train for
- optimizer: Model is trained using the Adam optimizer
- batch_normalization:
- decay_rate: float Rate at which batch normalizatio momentum exponentially decays
- decay_step: int Number of training samples after which the batch normalization momentum is decreased one step
- initial_decay: float Initial decay of batch normalization momentum
- learning_rate:
- decay_rate: float Rate at which the learning rate exponentially decays
- decay_step: int Number of training samples after which the learning rate is decreased one step
- initial: float Initial learning rate
- random_rotation: bool Whether to randomly rotate a training sample about the +Z axis
- shuffle: bool Whether to shuffle the samples in the training set at the beginning of each epoch
- batch_normalization:
- save_every_n_batches: int Number of batches after which to save a snapshot of the model
- session_id: int Numerical identifier for this session
- upsampling_only: bool Whether to only train the final upsampling layers of the model. Useful for finetuning a pretrained model on a smaller dataset