Repository navigation
Releases
DendriNet 0.5.0 and DendriNet Studio 0.1.0
DendriNet is a PyTorch library for training networks of dendritic neurons. Version 0.5.0 includes runnable tutorials. DendriNet Studio 0.1.0 adds a local browser interface for designing populations, inspecting the excitatory and inhibitory inputs to a neuron, and training a model.
Install the core library in a fresh Python 3.10 or later environment and run the synthetic-data tutorial. Core on PyPI
python -m pip install dendritic-modeling==0.5.0
dendritic-example synthetic_population --output tutorial.yaml
dendritic-train tutorial.yaml --epochs 1To use Studio, install the optional interface on your computer and open the local address printed by the launcher. Studio also installs a compatible core library. Studio on PyPI · Illustrated Studio guide
python -m pip install dendrinet-studio==0.1.0
dendrinet-studio --workspace studio-demoEmpirical paper. Download empirical-examples.zip for eight representative training configurations, the per-seed data and the analysis scripts for Figs. 3D and 7F of the revised manuscript. With the numerical dependencies listed in the paper, core v0.5.0 reproduced both reference tables and the PDF of Supplementary Fig. S29 byte for byte. The original native neuronal experiments used v0.3.0. The figure index records the separate AlexNet checkpoint sources. Empirical paper · Examples at v0.5.0
DendriNet 0.4.0
DendriNet 0.3.0
Upgrading from 0.2.0
- Configuration files now carry
config_version: 4. Standard dendritic E/I cores are validated as biological neurons unless they declarebiological_neuron: false, which stays available for exploratory models. - Run
dendritic-upgrade-configon configuration files written for earlier versions: it writes the historicalbiological_neuron: false, so they keep describing the same models and runs. A file withoutconfig_versionis read as format 4. - Synaptic weight normalization (
weight_norm_order) now normalizes rows whose norm lies below machine epsilon instead of dividing them by machine epsilon. Rows at or above machine epsilon are unchanged bit for bit, but checkpoints with such small rows can give different outputs, so compare before moving existing results to 0.3.0.
New
- Checked-identity
nonnegativeinputs that reject negative or non-finite rates, declaredpaired_afferents(distinct, identically tuned external E/I sources) and opt-in positive input projections (input_projection_weight_transform: softplus) for population and unified E/I networks. - Opt-in
shared_projection_shuntingcore: normalized sparse projections shared across native DendriNet banks, with signed block pooling and an affine output. - Biological configuration preparation and a source-pinned recipe validator (
dendritic_modeling.scripts.tools.biological_recipes) for migrating experiment recipes. dendritic-collect-reactivation --require-clean-analysis-sourcerejects runs whose analysis source is not one clean commit.
Fixes and performance
- Recurrent population and unified E/I networks check nonnegative inputs once per sequence instead of at every step; outputs, recurrent states and gradients are unchanged.
- Weight normalization below machine epsilon is exact, also when weights are subnormal or the L2 norm underflows;
expandsoftplusrows keep finite gradients, and the normalization traces as a single graph undertorch.compile. - Saved 0.2.0 core records stay recognized; flat population and unified E/I payloads accept
biological_neuron: false; section-wrappedshared_projection_shuntingmappings build; dictionary AlexNet recipes are no longer modified; the reactivation collection configuration rejects unknown keys and mistyped values. - Biological operator cards cover the neuronal core only (operator-card schema v3).
DendriNet 0.2.0
The first release since 0.1.0 (tag pre-refactor).
- Configuration: every option is a typed field, and unknown keys raise. Files carry
config_version: 3; dendritic-upgrade-config rewrites older files so that they
describe the same runs. - Reproducibility: runs are seeded and record their resolved configuration, source
commit, plugins and environment. - Corrections to models, training rules and analyses. Results produced earlier
reproduce at the library revision that produced them. - Plugins: installed packages can add architectures, synapse layers, activations,
datasets, trainer strategies, losses, optimizers and analyzers through the
dendritic_modeling.plugins entry point. - Compact checkpoint export (dendritic-compact-checkpoint).
- Transformer and vision replacement training is no longer part of the library.
v0.1.0
What's Changed
- Iss311 hs by @houman1359 in #313
- added and integrated default (old) initialization strategy by @houman1359 in #316
- add weight analysis / weight tranfer / analytical initializations by @houman1359 in #319
- fix non-shunting scaling by @houman1359 in #321
- merge weight transfer MR/weight boost to devTest by @houman1359 in #325
- merge weight transfer and local learning to pre-merge by @houman1359 in #328
- refactor config structure by @houman1359 in #330
- Add local learning, refactor config structure, add weight transfers and initializations by @houman1359 in #331
- added CompartmentStatisticsAnalyzer by @houman1359 in #334
- Iss333 add statistics compartment and more analysis, test new sweep setup by @houman1359 in #339
- Iss333 - add more analysis and fix issues/ add video visualizations / fix stochastic noise parametrization by @houman1359 in #344
- add more test coverage by @houman1359 in #348
- add RNN, add new strucutre for networks, improve information/ablation/plotting by @houman1359 in #351
- add LDS, update docs by @houman1359 in #352
- update docs by @houman1359 in #353
- restore docs deployment workflow by @houman1359 in #354
- dendrinetCNN pretrained implementation, ablation/merge by @houman1359 in #357
- update readme by @houman1359 in #358
Full Changelog: refactored...pre-neurips
What's Changed
- Iss311 hs by @houman1359 in #313
- added and integrated default (old) initialization strategy by @houman1359 in #316
- add weight analysis / weight tranfer / analytical initializations by @houman1359 in #319
- fix non-shunting scaling by @houman1359 in #321
- merge weight transfer MR/weight boost to devTest by @houman1359 in #325
- merge weight transfer and local learning to pre-merge by @houman1359 in #328
- refactor config structure by @houman1359 in #330
- Add local learning, refactor config structure, add weight transfers and initializations by @houman1359 in #331
- added CompartmentStatisticsAnalyzer by @houman1359 in #334
- Iss333 add statistics compartment and more analysis, test new sweep setup by @houman1359 in #339
- Iss333 - add more analysis and fix issues/ add video visualizations / fix stochastic noise parametrization by @houman1359 in #344
- add more test coverage by @houman1359 in #348
- add RNN, add new strucutre for networks, improve information/ablation/plotting by @houman1359 in #351
- add LDS, update docs by @houman1359 in #352
- update docs by @houman1359 in #353
- restore docs deployment workflow by @houman1359 in #354
- dendrinetCNN pretrained implementation, ablation/merge by @houman1359 in #357
- update readme by @houman1359 in #358
- update rnn and docs, add two more information metrics by @houman1359 in #359
- Integrate recurrent EI core config and init compatibility fixes by @houman1359 in #360
- Add recurrent E/I analysis framework follow-up by @houman1359 in #361
- update docs by @houman1359 in #362
Full Changelog: refactored...pre-drafts
v0.0.2
New refactored structure with new analysis and visualization. New capabilities like fstp, mixed precision, etc are added.
v0.0.1
This is the pre-refactor version of the codebase.