Scripts

run

class run.TASK_NAMES(value)

Bases: enum.Enum

An enumeration.

EXTRACT = 'extract_pt'

FINETUNE = 'finetune'

FIT = 'fit'

HDF5 = 'create_hdf5'

PREDICT = 'predict'

QUALIFY = 'qualify'

TEST = 'test'

run.launch_extract(config: omegaconf.dictconfig.DictConfig)

Training, evaluation, testing, or finetuning of a neural network.

run.launch_hdf5(config: omegaconf.dictconfig.DictConfig)

Build an HDF5 file from a directory with pairs of images, gt disparities and masks.

run.launch_qualify(config: omegaconf.dictconfig.DictConfig)

Training, evaluation, testing, or finetuning of a neural network.

run.launch_train(config: omegaconf.dictconfig.DictConfig)

Training, evaluation, testing, or finetuning of a neural network.

simlearner3d.train

simlearner3d.train.train(config: omegaconf.dictconfig.DictConfig) → pytorch_lightning.trainer.trainer.Trainer

Training pipeline (+ Test, + Finetuning)

Instantiates all PyTorch Lightning objects from config, then perform one of the following task based on parameter task.task_name:

fit:

Fits a neural network - train on a prepared training set and validate on a prepared validation set. Optionnaly, resume a checkpointed training by specifying config.model.ckpt_path.

test:

Tests a trained neural network on the test dataset of a prepared dataset (i.e. the test subdir which contains LAS files with a classification).

finetune:

Finetunes a checkpointed neural network on a prepared dataset, which must be specified in config.model.ckpt_path. In contrast to using fit, finetuning resumes training with altered conditions. This leads to a new, distinct training, and training state is reset (e.g. epoch starts from 0).

Typical use case are

• a different learning rate (config.model.lr) or a different scheduler (e.g. stronger config.model.lr_scheduler.patience)

• a different model to train on

• a different loss function …

Parameters

config (DictConfig) – Configuration composed by Hydra.

Returns

lightning trainer.

Return type

Trainer