Skip to content

Latest commit

 

History

History
263 lines (244 loc) · 13.4 KB

examples.rst

File metadata and controls

263 lines (244 loc) · 13.4 KB

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 :ref:`run-outputs`). The vision examples download MNIST or CIFAR-10 on first use, and the sequence examples generate their data (see :doc:`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. :ref:`config-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 :ref:`checkpoint-analysis`). For a checkpoint of the MNIST example of :doc:`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
configs/config_exp.yaml Annotated reference configuration with a comment for every population-network option: one CIFAR-10 layer of 32 inhibitory cells and 128 excitatory cells with [3, 3, 3] branch factors. See :ref:`reference-config`. dendritic-train
configs/config_exp_fsdp.yaml The same population architecture for fully sharded data-parallel training (distributed.mode: fsdp), with fixed indexed contacts and occupancy_quantile gate initialization. The single-process reference uses standard Top-K contacts and analytical gate initialization. See :ref:`distributed-training`. torchrun with train_experiments_fsdp
configs/sweep_config.yaml Annotated sweep template: synapse counts and excitatory branch factors of an MNIST population network, five seeds per combination (40 configurations). See :ref:`sweeps`. dendritic-sweep
configs/examples/population_network_feedforward.yaml One MNIST layer: an inhibitory population i (32 cells) and an excitatory population e (128 cells, [4, 4]), with input -> i, input -> e and same-step i -> e inhibition. Used in :doc:`quickstart`. dendritic-train
configs/examples/population_network_cifar10.yaml CIFAR-10 training setup: 128 excitatory cells with [3, 3, 3] branch factors and 32 inhibitory cells, 40 excitatory and 10 inhibitory synapses per branch, occupancy_quantile gates, synapse weight decay and weight boosting. dendritic-train
configs/examples/population_network_spatial_morphology_cifar10.yaml Every branch of 64 excitatory CIFAR-10 cells ([2, 2, 2, 2]) receives a fixed number of pixels from its own nested image region (structured_connectivity with method: spatial_morphology, topk_type: indexed). dendritic-train
configs/examples/population_network_dense_to_sparse_cifar10.yaml Learned pruning from dense allowed synapses down to the final synapse count during training (topk_type: dense_to_sparse). dendritic-train
configs/examples/population_network_indexed_rewire_mnist.yaml Indexed synapses that store only the active contacts of each branch and periodically replace the weakest ones with newly sampled inputs (topk_type: indexed_rewire). dendritic-train
configs/examples/population_network_normalized_additive_mnist.yaml Additive integration with voltage normalization (use_shunting: false, use_additive_normalization: true) on MNIST. dendritic-train
configs/examples/population_network_heterogeneous_feedforward.yaml Four populations in one MNIST layer (inhibitory pv and sst, excitatory pyr_flat and pyr_deep) with their own morphologies; population: blocks override options per population. dendritic-train
configs/examples/population_network_two_layer.yaml Two CIFAR-10 layers (128 excitatory and 32 inhibitory cells, then 64 and 16). The second layer reads layer0.e0 and layer0.i0 by qualified source names. dendritic-train
configs/examples/population_network_local_learning_mnist.yaml Local credit assignment (training.main.strategy: local_ca) on MNIST, with the information and synapse-turnover analyzers. See :ref:`local-credit-assignment`. dendritic-train
configs/examples/population_network_recurrent_pyr_pv_sst.yaml A recurrent microcircuit (pyr, pv, sst) on the context_dependent_decision sequence task: same-step input edges and delayed, probabilistic recurrent edges, trained with training.main.strategy: recurrent. dendritic-train
configs/examples/population_network_taus_and_delays.yaml Where the time constants live: the layer dt, the dendritic time constants (tau_base, tau_ratio, level_taus), receptor time constants and delayed edges. See :ref:`temporal-constants`. dendritic-train
configs/examples/population_network_synapse_types_ampa_nmda_gaba.yaml AMPA, NMDA, GABA-A and GABA-B receptor mixtures (synapse_types). See :ref:`synapse-types`. dendritic-train
configs/examples/population_network_spiking.yaml Leaky integrate-and-fire somas (dynamics_mode: spike) with a spike readout and the spike-train analyzer. See :ref:`spiking-soma`. dendritic-train
configs/examples/population_network_soma_feedback.yaml Feedback from the previous time step's soma output into the dendritic branches (soma_feedback_enabled). See :ref:`soma-feedback`. dendritic-train
configs/examples/population_network_contextual_remapping.yaml A two-layer network for the context-cued EMNIST task (data.dataset_name: contextual_remapping) trained on a stream of freshly sampled presentations (presentation_stream), with stochastic-release synapses (topk_type: indexed_release) and endpoint_tanh gates. It needs the EMNIST download and a CUDA GPU for data generation. dendritic-train
configs/examples/population_network_target_levels.yaml Two inhibitory classes that inhibit different dendritic levels of the excitatory cells (connection target_levels), with variance Top-K synapses (topk_variance_score: relative), on the same task. See :ref:`population-connections`. dendritic-train
configs/examples/contextual_remapping_affine_mlp.yaml A dense affine_mlp baseline (816-246-127-47 units) for the same task. See :ref:`affine-mlp`. dendritic-train
configs/examples/indexed_backends/feedforward_large.yaml Sweep template for a three-layer feedforward einet on CIFAR-10 (8,192 excitatory and 1,024 inhibitory cells per layer) with one variant per indexed projection backend (eager, recompute, triton_transposed, auto). See :ref:`indexed-backends`. dendritic-sweep
configs/examples/indexed_backends/rnn_large.yaml The same backend variants for a recurrent unified_ei network on a sequence task. dendritic-sweep
configs/examples/branch_local_sensitivity_analysis.yaml Analysis profile: local sensitivity of one dendritic tree, including the integration rule and the gate derivatives at held-out operating points (branch_local_sensitivity). dendritic-analyze-checkpoint
configs/examples/input_region_intervention_analysis.yaml Analysis profile: paired interventions on image regions of a feedforward classifier, with the layer-by-layer E/I responses (input_region_intervention). dendritic-analyze-checkpoint
configs/examples/local_inhibitory_population_intervention_analysis.yaml Analysis profile: interventions on whole local inhibitory populations of a feedforward population network (local_inhibitory_population_intervention). dendritic-analyze-checkpoint
configs/examples/path_matched_information_analysis.yaml Analysis profile: information carried by dendritic paths on fixed, label-independent supports (path_matched_information). dendritic-analyze-checkpoint
configs/examples/path_matched_intervention_analysis.yaml Analysis profile: interventions on a fixed number of paths of one dendritic tree (path_matched_intervention). dendritic-analyze-checkpoint
configs/examples/representation_accessibility_analysis.yaml Analysis profile: linear and nonlinear probes and class information of the soma populations of a population network (representation_accessibility). dendritic-analyze-checkpoint
configs/examples/source_tuning_support_analysis.yaml Analysis profile: class tuning of the input coordinates each synapse contacts, for MNIST images (source_tuning_support). dendritic-analyze-checkpoint
configs/examples/conductance_source_tuning_analysis.yaml The same analyzer for inputs that are not images (pixels mixed with context cues). dendritic-analyze-checkpoint
configs/examples/synaptic_source_support_analysis.yaml Analysis profile: how often each input coordinate is contacted, and with which effective conductance, per dendritic depth (weight analyzer). dendritic-analyze-checkpoint
configs/examples/synaptic_source_region_alignment_analysis.yaml Analysis profile: overlap of the contacted image region with the pixels of each soma's class on MNIST (weight analyzer). dendritic-analyze-checkpoint
configs/examples/operator_presets/00_mlp_flat.yaml to 04_shunting_tree.yaml Five model.core bodies for the einet core, in steps from a signed flat layer to a shunting E/I tree; see the README.md in that directory. model.core with type: einet
configs/examples/deployment/compact_inference.yaml Deployment policy for exporting a trained model with frozen sparse topology. See :ref:`compact-checkpoints`. dendritic-compact-checkpoint --compression-config
configs/examples/validation_epoch_selection.yaml Selection of the best epoch of each run from validation metrics collected by dendritic-collect-sweep-metrics. dendritic-select-validation-epochs --config
configs/templates/analysis_full_template.yaml Every analyzer section with comments, to copy into an analysis: section. See :ref:`analyzer-catalogue`. Copy into a configuration
configs/templates/indexed_runtime_template.yaml 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. :ref:`indexed-backends` explains the indexed options.