Config reference

Generated from visin_fusion/config/config_schema.py by tools/make_config_reference.py; do not edit. Schema version: 2.0. Export JSON Schema with visin-fusion schema.

Unknown keys in these sections are errors. Model sections (CLFT, CLFTv2, MaskFormer, Mask2Former, DeepLabV3Plus) take the settings of their model; start from its preset in visin_fusion/config/presets/.

CLI.backbone	Model section	Modes
clft	CLFT	rgb, lidar, cross_fusion
clftv2	CLFTv2	rgb, lidar, cross_fusion
maskformer	MaskFormer	rgb, lidar, cross_fusion
mask2former	Mask2Former	rgb, lidar, cross_fusion
deeplabv3plus	DeepLabV3Plus	rgb, lidar, fusion

Top level

The sections of a config and its name.

Key	Type	Default	Description
Summary	str	""	Run name, e.g. in Visin
description	str	""	What this config or preset is for
tags	list of str	[]
plugins	list of str		Modules to import before the config is checked, so they can register models (register_model)
CLI	section	required
General	section	required
Log	section	required
Dataset	section	required
CLFT	dict (optional)
CLFTv2	dict (optional)
MaskFormer	dict (optional)
Mask2Former	dict (optional)
DeepLabV3Plus	dict (optional)

CLI

The model and its inputs.

Key	Type	Default	Description
backbone	str	required	The model: clft, clftv2, maskformer, mask2former, deeplabv3plus, or one added with register_model
mode	rgb, lidar, fusion, cross_fusion	required	Inputs: camera (rgb), LiDAR (lidar) or both (fusion; cross_fusion is the same)

General

Training settings.

Key	Type	Default	Description
device	str	"cuda:0"	Torch device; the CPU is used when CUDA is not available
epochs	int	required
batch_size	int	required
seed	int	0
accumulate_batches	int	1	Batches whose gradients are averaged per optimizer step: effective batch size = batch_size x this
precision	fp16, bf16, fp32 (optional)		Training arithmetic on CUDA: fp16 mixed precision with loss scaling, bf16 mixed precision, or fp32. Empty keeps the model's default (fp16 for most models). Try bf16 or fp32 when a run goes NaN
max_nonfinite_batches	int	3	Training batches with a non-finite loss or gradients that are skipped per epoch; one more stops the run, keeping the last good checkpoint. 0 stops at the first
num_workers	int (optional)		DataLoader worker processes; 0 loads in the main process. Default: CPU count, at most 8 and at most one per batch
resume_training	bool	False	Continue from the latest checkpoint in Log.logdir
reset_lr	bool	False	When resuming, start the learning-rate schedule again
early_stop_patience	int	required	Epochs without a better validation mIoU before stopping
max_checkpoints	int	required	Best checkpoints kept
model_path	str	""	Checkpoint to start from; empty for the latest in Log.logdir
transfer_learning	bool	False	Start from model_path trained on other classes
source_classes	list (optional)		The classes model_path was trained on
callbacks	list of str		Callback classes to load in every stage, as "package.module:ClassName", built with the config
create_new_training	bool	False	Start a new Visin run even when resuming
hub_repo	str (optional)		Hugging Face model repo (org/name) the best checkpoint is published to when training ends; empty publishes nothing. Needs the hf extra and your own HF_TOKEN
hub_private	bool	True	Create hub_repo as a private repo; false makes it public
publish_model	visin (optional)		Store the best inference checkpoint in Visin when training ends; needs the visin extra and an enabled run
model_license	str (optional)		Licence id of a stored model, e.g. cc-by-4.0; unset models cannot be downloaded
model_credit	str (optional)		Attribution shown with the stored model
suite	str (optional)		Visin suite version, slug@version, the test stage records its results on as an evaluation of the tested checkpoint; empty records none. Needs a Visin that has suites, and a visin package that has evaluate
suite_file	str (optional)		The Visin suite file (JSON, or YAML with visin[yaml]) the test stage scores against. Its digest is sent as the protocol that ran, together with the digest of the test sets' frame lists, which marks the evaluation observed rather than reported; its slug@version is used when suite is empty

Log

Where a run writes.

Key	Type	Default	Description
logdir	str	required	Epoch logs, checkpoints, test results, visualizations and benchmarks

Dataset

The data, and the classes to learn. With a dataset.json at dataset_root, most of this comes from the manifest.

Key	Type	Default	Description
name	str	required
dataset_root	str	required	Frames in the split files are relative to this; may use environment variables, e.g. "$DATA_ROOT/zod"
train_split	str	required
val_split	str	required
split_dir	str (optional)		Where test and visualization splits are (visin_fusion/config/splits.py)
test_splits	dict of str, str (optional)		Test sets: name -> split file
visualization_split	str (optional)
annotation_path	str (optional)		Annotation folder that replaces a frame's camera folder
layout	dict of str, str (optional)		Frame folders, e.g.
dataset_classes	list of section	required
train_classes	list of section	required
transforms	section	required

Dataset.transforms

Input size, augmentation and normalization.

Key	Type	Default	Description
resize	int	required	Square input size
random_rotate_range	float	0	Training rotation range, degrees
p_flip	float	0
p_crop	float	0
p_rot	float	0
image_mean	list of float	required
image_std	list of float	required
lidar_mean	list of float (optional)
lidar_std	list of float (optional)

Dataset.dataset_classes[]

The dataset's label values.

Key	Type	Default	Description
name	str	required
index	int	required	Label value in the annotation images

Dataset.train_classes[]

The classes the model learns.

Key	Type	Default	Description
name	str	required
index	int	required	Model output channel; 0..n-1 without gaps, 0 is background
weight	float	1.0	Loss weight
dataset_mapping	list of int	required	Dataset class indices merged into this class
color	list of int	required	RGB color in visualizations