The Network Object

class leabra7.Net

The Net object is the primary point of interaction for scripts that use leabra7. It provides methods to construct the network, advance the network in time, and collect output data.

load(filename: str) → None:

Loads the network from a pickle file, overwriting the current network configuration. This function is insecure and could potentially execute arbitrary code; be sure not to load malicious or untrusted files.

Parameters:filename – The file from which to load the network.
save(filename: str) → None:

Saves the network as a pickle file.

Parameters:filename – Where to save the network.
new_layer(name: str, size: int, spec: lb.LayerSpec=None) → None:

Adds a new layer to the network.

Parameters:
  • name – The name of the layer.
  • size – How many units the layer should have.
  • spec – The layer spec. If None, the default layer spec will be used.
Raises:

ValidationError – If the spec contains an invalid parameter value.
new_projn(name: str, pre: str, post: str, size: int, spec: lb.ProjnSpec=None) → None:

Adds a new projection to the network.

Parameters:
  • name – The name of the projection.
  • pre – The name of the sending layer.
  • post – The name of the receiving layer.
  • spec – The projection spec. If None, the default spec will be used.
Raises:
  • ValueError – If pre or post do not match any existing layer names.
  • ValidationError – If the spec contains an invalid parameter value.
clamp_layer(name: str, acts: Sequence[float]) → None:

Clamps layer’s activations to the specified values, so that they do not change from cycle to cycle.

Parameters:
  • name – The name of the layer to clamp.
  • acts – A sequence containing the activations to which the layer’s units will be clamped. If its length is less than the number of units in the layer, it will be tiled. If its length is greater, the extra values will be ignored.
Raises:

ValueError – If name does not match any existing layer name.
unclamp_layer(name: str) → None:

Unclamps a previously-clamped layer. If the layer is not clamped, then nothing happens.

Parameters:name – The name of the layer to unclamp.
cycle() → None:

Cycles the network.

minus_phase_cycle(num_cycles: int = 50) → None:

Runs a series of cycles for the trial minus phase, signaling the network to compute the appropriate metrics at the beginning and end of the phase.

A minus phase is the trial phase where input patterns are clamped to the input layers, but output patterns are not clamped to the output layers. This clamping is the user responsibility.

Parameters:num_cycles – The number of cycles in the minus phase.
Raises:ValueError – If num_cycles is less than 1.
plus_phase_cycle(num_cycles: int = 25) → None:

Runs a series of cycles for the trial plus phase, which is like the minus phase except that target values are clamped on the output layers.

Parameters:num_cycles – The number of cycles in the plus phase.
Raises:ValueError – If num_cycles is less than 1.
learn() → None:

Updates the projection weights with the XCAL learning equation.

end_epoch() → None:

Signals the network that an epoch (one pass through the training data) has ended. This must be called by the user at the end of every epoch.

end_batch() → None:

Signals the network that a batch has ended (typically, a batch is a series of epochs). This must be called by the user at the end of every batch.

observe(name: str, attr: str) → pd.DataFrame:

This is like logging, but it only returns the current object state. There is no history. If you do not require historical observations, use this to avoid the performance penalty of logging.

Parameters:
  • name – The name of the object to observe.
  • attr – The name of the attribute to observe. This can be any loggable attribute. Check the object’s associated spec object for a list of valid loggable attributes.
Raises:

ValueError – If the object does not exist, does not support observations, or if the attribute is not a valid loggable attribute.
Returns:

A Pandas dataframe containing the observation result.
logs(freq: str, name: str) → Tuple[pd.DataFrame, pd.DataFrame]:

Retrieves logs (observations recorded over time) for an object in the network. Which logs should be recorded is specified in Spec objects at network creation, using the parameters "log_on_cycle", "log_on_trial", etc.

Parameters:
  • freq – The frequency at which the desired logs were recorded. One of ["cycle", "trial", "epoch", "batch"].
  • name – The name of the object for which the logs were recorded.
Raises:

ValueError – If the frequency name is invalid, or if no logs were recorded for the desired object.
Returns:

A tuple of Pandas dataframes. The first element of the tuple contains the logs for “whole” attributes, which are attributes of the object itself, like a layer’s "avg_act". The second element of the tuple contains the logs for “parts” attributes, which are attributes of the object’s constituents, like a layer’s "unit_act" attribute. For layers, parts attributes pertain to the units, and for projections, parts attributes pertain to the connections.
pause_logging(freq: str=None) → None:

Pauses logging in the network, if any logging is enabled. This is typically done for performance reasons; logging is quite slow.

Parameters:freq – The frequency for which to pause the logging, one of "cycle", "trial",, "epoch", or "batch". If None, pauses for all frequencies.
Raises:ValueError – If no frequency with name freq exists.
resume_logging(freq: str=None) → None:

Resumes logging in the network.

Parameters:freq – The frequency for which to resume the logging, one of "cycle", "trial",, "epoch", or "batch". If None, pauses for all frequencies.
Raises:ValueError – If no frequency with name freq exists.