Command line interface

The console script hetero is bound to hetero.cli.main(). The parser is built by hetero.cli.get_parser(), whose flags are grouped into the sections listed below; hetero.cli.create_param_dicts() turns the parsed namespace into the nested parameter dictionary that hetero.sim.Simulation consumes. Every default comes from hetero.config.CFG, so the parser stays in sync with the config automatically.

Input and output

--root

Result directory. Default: hetero.config.res_root.

--stype

Storage dtype of the simulated states (float64 for accurate forecasting, float16 otherwise). Default: float16.

--ncore

Number of worker processes; -1 uses all cores. Default: -1.

--seed

Random number seed. Default: 1234.

Stage skips

Each of these skips one stage of the run, so a finished run can be extended or inspected piecemeal.

--noint

Skip integration and reuse the states on disk.

--nobm

Skip benchmarking and reuse the scores on disk.

--nosvd

Skip the SVD computation, i.e. skip the svals/erank/angle csv files. This is a computation flag, not a plot flag: --noviz still writes the numbers to disk.

--noviz

Skip every figure of the run. Implies --noconn, --notau, --nouvx, --noscore, --noangle and --nosumplot (see hetero.cli.PLOT_SKIP_FLAGS). The csv files are still computed. Use it while sweeping parameters.

--replot

Re-render every figure of a finished run, loading whatever is already on disk and computing only what is missing. Overrides --noviz and the individual --no* plot flags, and also recomputes missing integration/benchmark results even under --noint/--nobm.

The individual plot flags are --noconn (connectivity), --notau (time-constant distribution), --nouvx (input/voltage/activation traces), --noscore (readout score map), --noangle (angle-vs-score) and --nosumplot (angle CDF, singular values, effective rank). The summary figures aggregate every state on disk, so in a sweep they are re-rendered from all previously computed cases after every single run.

Heterogeneity

--distro

Distribution of the neuronal time constants, e.g. lognormal, gamma, normal, uniform. Any numpy.random-compatible distribution. Default: lognormal.

--n_means

Number of mean-timescale levels. Default: 3.

--n_vars

Number of heterogeneity levels per mean level. Default: 3.

--log_distance

Log-10 distance between two consecutive mean levels. Default: 0.5.

Network

--N

Number of neurons. Default: 250.

--p

Connection probability (sparsity). Default: 0.1.

--f

Fraction of excitatory neurons. Default: 0.8.

--mue

Mean excitatory synaptic weight. Default: 1.0.

--sig0

Standard deviation of the excitatory and inhibitory weights. Default: 1.0.

--autapse

Whether self-connections are allowed. Default: False.

--delay

Synaptic delay in units of dt. Default: 0.

Gains

--Jn

Gain of the white noise. Default: 0.1.

--J

Gain of the recurrent weights. Default: 1.0.

--Ju

Gain of the feedforward weights. Default: 1.0.

Stimulation

--stim

One or more stimulus labels; the timeseries are concatenated. See the stim block of hetero.config.CFG for the available labels (lorenz, mackey_glass-10, mackey_glass-50, mackey_glass-80, sin, sin-5, sinabs, sign, sign-5, narma, constant) and their parameters.

Task

--task

taylor, nostradamus or tayloramus. Default: tayloramus.

Dataset

--n_trn

Length of the training set, in units of the input base timescale. Default: 20.

--n_tst

Length of the test set, in the same units. Default: 10.

--disable_scale_by_size

Turn off the automatic scaling of the training-set size with the network size. By default enlarging the network by a factor of x enlarges the training set by the same factor; the test set is unaffected.

Dynamics

--dyn

LI (leaky integrator, rate) or LIF (leaky integrate-and-fire, brian2). Default: LI.

--solver

Solver used by the rate dynamics. Default: euler.

Readout

--readout_activation

How the neuronal state is turned into a readout signal. Rate networks take an activation function (possig1, tanh1), spiking networks take a temporal kernel (array, g50, gnc50, d50, d10, a50, ma50). An incompatible choice is corrected with a warning; see hetero.config.VALID_ACTIVATIONS. Default: possig1.

Optimization

--alg

Any scikit-learn-compatible regressor, e.g. ridge, ols, lasso, elasticnet. Default: ridge.

--reg

Regularization strength, where the algorithm uses one. Default: 1e-6.

--fit_bias

Whether to fit an intercept. Default: True.

Uncertainty quantification

--n_trials

Number of non-overlapping contiguous intervals over which a readout is trained. The spread of the corresponding test scores estimates the performance uncertainty. Default: 3.

--n_realizations

Number of independent realizations of the topology and the weights. The spread of the scores across realizations estimates the network uncertainty. Default: 1.

API

hetero.cli.get_options(arg)

hetero.cli.get_parser()

Creates a parser that is able to capture all variables necessary to construct the paramter dictionaries.

hetero.cli.create_param_dicts(args, verbose=True)

Creates proper dictionaries from args

hetero.cli.main()