GRASS logo

NAME

m.neural_network.train - Trains a neural network for semantic segmentation including the steps for initial training and finetuning.

KEYWORDS

raster, training, finetuning, neural network, classification

SYNOPSIS

m.neural_network.train
m.neural_network.train --help
m.neural_network.train data_dir=name img_size=integer in_channels=integer [out_classes=integer] [model_arch=string] [encoder_name=string] [encoder_weights=string] [epochs=integer] [batch_size=integer] [input_model_path=string] output_model_path=string output_train_metrics_path=string [--help] [--verbose] [--quiet] [--ui]

Flags:

--help
Print usage summary
--verbose
Verbose module output
--quiet
Quiet module output
--ui
Force launching GUI dialog

Parameters:

data_dir=name [required]
Name of the input data directory containing subfolders with training data
Name of input directory
img_size=integer [required]
Image size in pixels
Default: 512
in_channels=integer [required]
Number of channels used as input for classification: R-G-B-I-ndsm (default 5)
Default: 5
out_classes=integer
Number of classes for classification: tree/ no-tree (default 2)
Default: 2
model_arch=string
Model architecture for classification (default Unet)
Default: Unet
encoder_name=string
Encoder for the classification (default resnet34)
Default: resnet34
encoder_weights=string
Weights for encoder (default imagenet)
Default: imagenet
epochs=integer
Number of training epochs (default 50)
Default: 50
batch_size=integer
Number of tiles used per training unit (default 8)
Default: 8
input_model_path=string
Name of the input model directory for finetuning.
output_model_path=string [required]
Name of the output model directory.
output_train_metrics_path=string [required]
Name of the output directory containing the metrics of the training.

Table of contents

DESCRIPTION

m.neural_network.train trains or fine-tunes a neural network using the segmentation_models.pytorch framework for semantic segmentation.

A new model can be trained by specifying the options img_size, out_classes, model_arch, encoder_name, encoder_weights, in_channels. A locally saved model can be further trained (fine-tuned) by giving the path to the previously saved model. For more information about available encoder-decoder combinations, see the smp documentation.

A larger batchsize is generally better for training, but requires more GPU RAM. A bit of experimentation is needed to find a batch size that still fits into the GPU RAM.

NOTES

It is expected that all data lie in the directory structure and naming format as created by m.neural_network.preparedata_part1.

EXAMPLES

  m.neural_network.train data_dir=path/to/data/train/ output_model_path=path/to/model

SEE ALSO

v.import, g.region r.mapcalc, v.to.rast,

REQUIREMENTS

AUTHORS

Victoria-Leandra Brunn, mundialis GmbH & Co. KG

SOURCE CODE

Available at: m.neural_network.train source code (history)

Accessed: Monday Jul 27 12:26:22 2026

Main index | Miscellaneous index | Topics index | Keywords index | Graphical index | Full index

© 2003-2026 GRASS Development Team, GRASS 8.5.1dev Reference Manual