Version 10 (modified by apdavison, 5 years ago)

--

API documentation: | spikes? | analysis | benchmark? | parameters | stgen | utilities | facets? |

The parameters module

We consider it to be best practice to cleanly separate the parameters of a model from the model itself. At the least, parameters should be defined in a separate section at the start of a file. Ideally, they should be defined in a separate file entirely. This makes version control easier, since the model code typically changes less often than the parameters, and makes it easier to track a simulation project, since the parameter sets can be stored in a database, displayed in a GUI, etc.

Parameters

At their simplest, individual parameters consist of a name and a value. The value is either a simple type such as a numerical value or a string, or an aggregate of such simple types, such as a set, list or array.

However, we may also wish to specify the physical dimensions of the parameter, i.e., its units, and the range of permissible values.

It is also often useful to specify an object that generates numerical values or strings, such as a random number generator, and treat that object as the parameter.

To support all these uses, we define the Parameter and ParameterRange classes, and various subclasses of the ParameterDist abstract class, such as GammaDist, NormalDist and UniformDist.

The Parameter class

Here are some examples of creating Parameter objects:

>>> i1 = Parameter(3)
>>> f1 = Parameter(6.2)
>>> f2 = Parameter(-65.3, "mV")
>>> s1 = Parameter("hello", name="message_to_the_world")

The parameter name, units, value and type can be accessed as attributes:

>>> i1.value
3
>>> f1.type
<type 'float'>
>>> f2.units
'mV'
>>> s1.name
'message_to_the_world'

Parameter objects are not hugely useful at the moment. The units are not used for checking dimensional consistency, for example, and Parameter objects are not drop-in replacements for numerical values - you must always use the value attribute to access the value, whereas it might be nice to define, for example, a class IntegerParameter which was a subclass of the built-in int type.

The ParameterRange class

When investigating the behaviour of a model or in doing sensitivity analysis, it is often useful to run a model several times using a different value for a certain parameter each time (also see the iter_range_keys() and similar methods of the ParameterSet class, below). The ParameterRange class supports this. Some usage examples:

>>> tau_m_range = ParameterRange([10.0, 15.0, 20.0], "ms", "tau_m")
>>> tau_m_range.name
'tau_m'
>>> tau_m_range.next()
10.0
>>> tau_m_range.next()
15.0
>>> [2*tau_m for tau_m in tau_m_range]
[20.0, 30.0, 40.0]

The ParameterDist classes

As with taking parameter values from a series or range, it is often useful to pick values from a particular random distribution. Three classes are available: UniformDist, GammaDist and NormalDist. Examples:

>>> ud = UniformDist(min=-1.0, max=1.0)
>>> gd = GammaDist(mean=0.5, std=1.0)
>>> nd = NormalDist(mean=-70, std=5.0)
>>> ud.next()
array([-0.56342352])
>>> gd.next(3)
array([ 0.04061142,  0.05550265,  0.23469344])
>>> nd.next(2)
array([-76.18506715, -68.71229944])

[Note that very similar functionality is available with the RandomDistribution class in the pyNN.random module. We should look at the best ways to avoid duplication].

Parameter sets

A problem with parameter sets for large-scale, detailed models is that the list of parameters gets very long and unwieldy, and due to the typically hierarchical nature of such models, the individual parameter names can also get very long, e.g., v1_layer5_pyramidal_apical_dend_gbar_na.

A solution to this is to give the parameter set a hierarchical structure as well, which allows the top-level list of parameters to be very short (e.g. v1, retina and lgn for a visual system simulation) since the top-level parameters are themselves parameter sets.

The simplest way to implement this in Python is using nested dicts. One disadvantage of this is that accessing deeply-nested parameters can be very verbose, e.g. v1['layer5']['pyramidal']['apical_dend']['na']['gbar']. A second disadvantage is that it is tedious to flatten the hierarchy when this becomes necessary, e.g. for serialisation - writing to file, etc.

For these reasons we have created a ParameterSet class, which:

  1. allows a more convenient notation;

2. enables subsets of the parameters, lower in the hierarchy, to be passed around by themselves;

3. provides convenient methods for reading from/writing to file and for determining the differences between two different parameter sets.

An example of the notation is v1.layer5.pyramidal.apical_dend.na.gbar, which requires only a single . for each level in the hierarchy rather than two "'"s, a "[" and a "]". This is not much shorter than v1_layer5_pyramidal_apical_dend_gbar_na - the difference is that v1.layer5.pyramidal is itself a ParameterSet object that can be passed as an argument to the pyramidal cell object, which doesn't care about v1.layer4.spinystellate, let alone retina.ganglioncell.magno.tau_m (while v1_layer5_pyramidal is just a NameError).

The ParameterSet class

Creation

ParameterSet objects may be created from a dict:

>>> sim_params = ParameterSet({'dt': 0.11, 'tstop': 1000.0})

or loaded from a URL:

>>> exc_cell_params = ParameterSet("https://neuralensemble.org/svn/NeuroTools/trunk/doc/example.param")

They may be nested:

>>> inh_cell_params = ParameterSet({'tau_m': 15.0, 'cm': 0.5})
>>> network_params = ParameterSet({'excitatory_cells': exc_cell_params, 'inhibitory_cells': inh_cell_params})
>>> P = ParameterSet({'sim': sim_params, 'network': network_params}, label="my_params")

Note that although we show here only numerical parameter values, Parameter, ParameterRange and ParameterDist objects, as well as strings, may also be parameter values.

Viewing and saving

To see the entire parameter set at once, nicely formatted use the pretty() method:

>>> print P.pretty()
{
  "network": {
    "excitatory_cells": url("https://neuralensemble.org/svn/NeuroTools/trunk/doc/example.param")
    "inhibitory_cells": {
      "tau_m": 15.0,
      "cm": 0.75,
    },
  },
  "sim": {
    "tstop": 1000.0,
    "dt": 0.11,
  },
}

By default, if the ParameterSet contains other ParameterSets that were loaded from URLs, these will be represented with a url() function in the output, but there is also the option to expand all URLs and show the full contents:

>>> print P.pretty(expand_urls=True)
{
  "network": {
    "excitatory_cells": {
      "tau_refrac": 0.11,
      "tau_m": 10.0,
      "cm": 0.25,
      "synI": {
        "tau": 10.0,
        "E": -75.0,
      },
      "synE": {
        "tau": 1.5,
        "E": 0.0,
      },
      "v_thresh": -57.0,
      "v_reset": -70.0,
      "v_rest": -70.0,
    },
    "inhibitory_cells": {
      "tau_m": 15.0,
      "cm": 0.75,
    },
  },
  "sim": {
    "tstop": 1000.0,
    "dt": 0.11,
  },
}

If a ParameterSet was loaded from a URL, it may be modified then saved back to the same URL, provided the protocol supports writing:

>>> exc_cell_params.save()
Traceback (most recent call last):
  File "<stdin>", line 1, in ?
  File "parameters.py", line 266, in save
    raise Exception("Saving using the %s protocol is not implemented" % scheme)
Exception: Saving using the https protocol is not implemented

or saved to a different URL:

>>> exc_cell_params.save(url="file:///tmp/exc_params")

The file format is the same as that produced by the pretty() method.

Copying and converting

A ParameterSet can be used simply as a dictionary, but can also be converted explicitly to a dict if required:

>>> print sim_params.as_dict()
{'tstop': 1000.0, 'dt': 0.11}

[need to say something about tree_copy()]

Iteration

There are several different ways to iterate over all or part of the ParameterSet object. keys(), values() and items() work as for dict``s. For the sake of more readable code, ``names() is provided as an alias for keys() and parameters() as an alias for items():

>>> P.names()
['network', 'sim']
>>> exc_cell_params.parameters()
[('tau_refrac', 0.11), ('tau_m', 10.0), ('cm', 0.25),
 ('synI', {'tau': 10.0, 'E': -75.0}), ('synE', {'tau': 1.5, 'E': 0.0}),
 ('v_thresh', -57.0), ('v_reset', -70.0), ('v_rest', -70.0)]

To flatten nested parameter sets, i.e., the iterate recursively over all branches of the tree, the the flatten() method returns a dict with keys created by joining the names at each hierarchical level with a separator character ('.' by default):

>>> network_params.flatten()
{'excitatory_cells.synI.E': -75.0, 'excitatory_cells.v_rest': -70.0,
 'excitatory_cells.tau_refrac': 0.11, 'excitatory_cells.v_reset': -70.0,
 'excitatory_cells.v_thresh': -57.0, 'excitatory_cells.tau_m': 10.0,
 'excitatory_cells.synI.tau': 10.0, 'excitatory_cells.cm': 0.25,
 'inhibitory_cells.cm': 0.75, 'excitatory_cells.synE.tau': 1.5,
 'excitatory_cells.synE.E': 0.0, 'inhibitory_cells.tau_m': 15.0}

while the flat() method returns a generator which yields (name, value) tuples.:

>>> for x in network_params.flat():
...   print x
...

The ParameterSpace class

The ParameterSpace class is a subclass of ParameterSet that is allowed to contain ParameterRange and ParameterDist objects as parameters. This turns the single point in parameter space represented by a ParameterSet into a set of points. For example, the following definition creates a set of six points in parameter space, which can be obtained in turn using the iter_inner() method:

>>> PS = ParameterSpace({
...        'x': 999,
...        'y': ParameterRange([10, 20]),
...        'z': ParameterRange([-1, 0, 1])
... })
>>> for P in PS.iter_inner():
...     print P
{'y': 10, 'x': 999, 'z': -1}
{'y': 20, 'x': 999, 'z': -1}
{'y': 10, 'x': 999, 'z': 0}
{'y': 20, 'x': 999, 'z': 0}
{'y': 10, 'x': 999, 'z': 1}
{'y': 20, 'x': 999, 'z': 1}

Putting parameter distribution objects inside a ParameterSpace allows an essentially infinite number of points to be generated:

>>> PS2 = ParameterSpace({
...    'x': UniformDist(min=-1.0, max=1.0),
...    'y': GammaDist(mean=0.5, std=1.0),
...    'z': NormalDist(mean=-70, std=5.0)
... })
>>> for P in PS2.realize_dists(n=3):
...     print P
{'y': 1.81311773668, 'x': 0.883293989399, 'z': -73.5871002759}
{'y': 0.299391158731, 'x': 0.371474054049, 'z': -68.6936045978}
{'y': 2.90108202422, 'x': -0.388218831787, 'z': -68.6681724449}

This document was last updated for r274.