DendriNet ========= DendriNet is a PyTorch library for networks of dendritic neurons. Its Python package is ``dendritic_modeling``. A DendriNet neuron is a tree of dendritic branches: each branch reads a sparse set of excitatory and inhibitory inputs, integrates them by shunting (divisive) or additive inhibition, and passes its output toward the soma. Neurons are grouped into named excitatory and inhibitory populations that can be feedforward or recurrent, rate-based or spiking. Experiments are described in YAML files that the library loads, validates, trains, analyzes and sweeps. .. figure:: _static/architecture/system-overview.svg :alt: Overview: YAML configuration is loaded into typed settings, the factory builds an encoder, a dendritic core and a decoder, and trainers, sweeps and the analysis manager produce run artifacts. :width: 100% From a YAML file to trained models and analysis results. Install the released library and train a small population network on generated data for one CPU epoch: .. code-block:: bash python -m pip install dendritic-modeling==0.5.0 dendritic-example synthetic_population --output demo.yaml dendritic-train demo.yaml --epochs 1 For a graphical first run, install the companion package with ``python -m pip install dendrinet-studio==0.1.0`` and follow :ref:`studio-first-run`. Every model has three parts: an **encoder** that prepares the input, a **core** that holds the dendritic network (or a baseline), and a **decoder** that maps the core output to the task. :doc:`network_architectures` describes this structure and the families of cores. Where to start -------------- .. list-table:: :header-rows: 1 :widths: 40 60 * - Goal - Page * - Install DendriNet and train a first model - :doc:`installation`, :doc:`quickstart` * - Draw a network, trace E and I inputs and compare training runs - :doc:`studio`, :ref:`studio-first-run` * - Find a configuration to start from - :doc:`examples` * - Build populations of dendritic neurons - :doc:`population_networks` * - Choose integration rules, gates and synapse types - :doc:`dendritic_neurons`, :doc:`synapses`, :doc:`temporal_dynamics` * - Write and validate a configuration file - :doc:`configuration_guide` * - Choose a learning rule or train on several GPUs - :doc:`training_strategies`, :doc:`distributed_training` * - Find the files a run writes - :doc:`runs_and_outputs` * - Analyze trained models and saved checkpoints - :doc:`analysis`, :doc:`analyzers`, :doc:`checkpoint_analysis` * - Run parameter sweeps - :doc:`sweeps` * - Pin a project's code to an exact library commit - :doc:`project_provenance` * - Look up a component, an option or a command - :doc:`available_components`, :doc:`options_reference`, :doc:`cli_reference` * - Add components from your own package - :doc:`extending` .. toctree:: :maxdepth: 1 :caption: Getting started installation releases studio quickstart examples .. toctree:: :maxdepth: 1 :caption: Building models network_architectures dendritic_neurons synapses population_networks temporal_dynamics structured_ei_cores conductance_options comparison_models .. toctree:: :maxdepth: 1 :caption: Configuration configuration_guide config_reference .. toctree:: :maxdepth: 1 :caption: Training training_strategies distributed_training performance_kernels .. toctree:: :maxdepth: 1 :caption: Data and analysis datasets analysis analyzers checkpoint_analysis .. toctree:: :maxdepth: 1 :caption: Running experiments runs_and_outputs sweeps model_compression project_provenance .. toctree:: :maxdepth: 1 :caption: Reference available_components options_reference cli_reference api_reference .. toctree:: :maxdepth: 1 :caption: Development extending engineering_architecture contributing