Example configurations
The repository ships several kinds of YAML files under configs/: reference
configurations, training examples for the main model families and features,
sweep templates, analysis profiles and smaller building blocks. This page
lists every one of them and the command that uses it. Run the commands from
the root of the repository checkout.
Using the examples
A training example trains with dendritic-train; check an edited copy
with --validate-only before a long run, and shorten a first run with
--epochs:
dendritic-train configs/examples/population_network_two_layer.yaml --validate-only
dendritic-train configs/examples/population_network_two_layer.yaml --epochs 1
Each example names its own outputs.results_dir and outputs.run_name
(see Run directory layout). The vision examples download MNIST or CIFAR-10 on
first use, and the sequence examples generate their data (see
Datasets).
Complete experiment files use config_version: 4. Their settings define
specific examples, rather than the library defaults. Analysis profiles,
deployment policies and operator presets are reusable sections with their own
readers; they cannot be passed directly to dendritic-train. A sweep embeds
an experiment under base_config. Configuration format version explains how to
upgrade earlier experiment files.
A sweep template expands into training configurations, which you then train one by one or submit as Slurm jobs:
dendritic-sweep --config configs/examples/indexed_backends/rnn_large.yaml \
--generate-only --output-dir outputs/indexed_backends
An analysis profile holds only an analysis: section. Run it on a trained
checkpoint with dendritic-analyze-checkpoint, giving the run’s
config.json (see Analyzing saved checkpoints). For a checkpoint of the
MNIST example of Quickstart:
dendritic-analyze-checkpoint \
--config "<run directory>/config.json" \
--checkpoint "<run directory>/final_model.pt" \
--analysis-config configs/examples/synaptic_source_support_analysis.yaml \
--output-dir outputs/synaptic_source_support
Some analysis profiles name a module path, a number of somas or an image size of the model they were written for; adapt those fields to your model first.
Example map
Configuration |
What it shows |
Used with |
|---|---|---|
|
Annotated reference configuration with a comment for every
population-network option: one CIFAR-10 layer of 32 inhibitory cells
and 128 excitatory cells with |
|
|
The same population architecture for fully sharded data-parallel
training ( |
|
|
Annotated sweep template: synapse counts and excitatory branch factors of an MNIST population network, five seeds per combination (40 configurations). See Parameter sweeps. |
|
|
One MNIST layer: an inhibitory population |
|
|
CIFAR-10 training setup: 128 excitatory cells with |
|
|
Every branch of 64 excitatory CIFAR-10 cells ( |
|
|
Learned pruning from dense allowed synapses down to the final synapse
count during training ( |
|
|
Indexed synapses that store only the active contacts of each branch
and periodically replace the weakest ones with newly sampled inputs
( |
|
|
Additive integration with voltage normalization
( |
|
|
Four populations in one MNIST layer (inhibitory |
|
|
Two CIFAR-10 layers (128 excitatory and 32 inhibitory cells, then 64
and 16). The second layer reads |
|
|
Local credit assignment ( |
|
|
A recurrent microcircuit ( |
|
|
Where the time constants live: the layer |
|
|
AMPA, NMDA, GABA-A and GABA-B receptor mixtures ( |
|
|
Leaky integrate-and-fire somas ( |
|
|
Feedback from the previous time step’s soma output into the dendritic
branches ( |
|
|
A two-layer network for the context-cued EMNIST task
( |
|
|
Two inhibitory classes that inhibit different dendritic levels of the
excitatory cells (connection |
|
|
A dense |
|
|
Sweep template for a three-layer feedforward |
|
|
The same backend variants for a recurrent |
|
|
Analysis profile: local sensitivity of one dendritic tree, including
the integration rule and the gate derivatives at held-out operating
points ( |
|
|
Analysis profile: paired interventions on image regions of a
feedforward classifier, with the layer-by-layer E/I responses
( |
|
|
Analysis profile: interventions on whole local inhibitory populations
of a feedforward population network
( |
|
|
Analysis profile: information carried by dendritic paths on fixed,
label-independent supports ( |
|
|
Analysis profile: interventions on a fixed number of paths of one
dendritic tree ( |
|
|
Analysis profile: linear and nonlinear probes and class information of
the soma populations of a population network
( |
|
|
Analysis profile: class tuning of the input coordinates each synapse
contacts, for MNIST images ( |
|
|
The same analyzer for inputs that are not images (pixels mixed with context cues). |
|
|
Analysis profile: how often each input coordinate is contacted, and
with which effective conductance, per dendritic depth ( |
|
|
Analysis profile: overlap of the contacted image region with the
pixels of each soma’s class on MNIST ( |
|
|
Five |
|
|
Deployment policy for exporting a trained model with frozen sparse topology. See Compact checkpoints. |
|
|
Selection of the best epoch of each run from validation metrics
collected by |
|
|
Every analyzer section with comments, to copy into an |
Copy into a configuration |
|
Recommended indexed-synapse runtime settings (backend, index type, workspace, memory-efficient gates) for structured cores and population blocks. |
Copy into a configuration |
Most population-network examples use indexed synapses
(topk_type: indexed_rewire, or indexed in the spatial-morphology
example) with indexed_index_dtype: auto, the recompute projection
backend and memory-efficient gates (reactivation_memory_efficient: true).
The context-cued EMNIST example with stochastic release uses the eager
backend with 32-bit indices. The example with level-specific inhibitory
targets uses variance Top-K over a dense candidate bank, so indexed-backend
settings do not apply. Projection backends explains the indexed options.