|
CARLsim
3.1.3
CARLsim: a GPU-accelerated SNN simulator
|
|
CARLsim
3.1.3
CARLsim: a GPU-accelerated SNN simulator
|
Class ConnectionMonitor. More...
#include <connection_monitor.h>
Public Member Functions | |
| ConnectionMonitor (ConnectionMonitorCore *connMonCorePtr) | |
| ConnectionMonitor constructor. More... | |
| ~ConnectionMonitor () | |
| ConnectionMonitor destructor. More... | |
| std::vector< std::vector< float > > | calcWeightChanges () |
| Reports the weight changes since the last snapshot in a 2D weight matrix (pre x post) More... | |
| short int | getConnectId () |
| Returns the connection ID that this ConnectionMonitor is managing. More... | |
| int | getFanIn (int neurPostId) |
| Returns the number of incoming synapses for a specific post-synaptic neuron. More... | |
| int | getFanOut (int neurPreId) |
| Returns the number of outgoing synapses for a specific pre-synaptic neuron. More... | |
| double | getMaxWeight (bool getCurrent=false) |
| Returns the max weight in the connection. More... | |
| double | getMinWeight (bool getCurrent=false) |
| Returns the min weight in the connection. More... | |
| int | getNumNeuronsPost () |
| Returns the number of post-synaptic neurons. More... | |
| int | getNumNeuronsPre () |
| Returns the number of pre-synaptic neurons. More... | |
| int | getNumSynapses () |
| Returns the number of allocated synapses. More... | |
| int | getNumWeightsChanged (double minAbsChanged=1e-5) |
| Returns the number of weights that have changed since the last snapshot. More... | |
| int | getNumWeightsInRange (double minValue, double maxValue) |
| Returns the number of weights in the connection whose values are within some range (inclusive) More... | |
| int | getNumWeightsWithValue (double value) |
| Returns the number of weights in the connection with a particular value. More... | |
| double | getPercentWeightsChanged (double minAbsChanged=1e-5) |
| Returns the percentage of weights that have changed since the last snapshot. More... | |
| double | getPercentWeightsInRange (double minValue, double maxValue) |
| Returns the percentage of weights whose values are within some range (inclusive) More... | |
| double | getPercentWeightsWithValue (double value) |
| Returns the percentage of weights in the connection with a particular value. More... | |
| int64_t | getTimeMsCurrentSnapshot () |
| Returns the timestamp of the current snapshot (ms since beginning of simulation) More... | |
| int64_t | getTimeMsLastSnapshot () |
| Returns the timestamp of the last snapshot (ms since beginning of simulation) More... | |
| int64_t | getTimeMsSinceLastSnapshot () |
| Returns the timestamp difference of the current and last snapshot. More... | |
| double | getTotalAbsWeightChange () |
| Returns the absolute sum of all the weight changes since the last snapshot. More... | |
| void | print () |
| Prints the current weight state as a 2D matrix (pre x post) More... | |
| void | printSparse (int neurPostId=ALL, int maxConn=100, int connPerLine=4) |
| Prints the current weight state as a sparse list of weights. More... | |
| void | setUpdateTimeIntervalSec (int intervalSec) |
| Sets the time interval (seconds) for writing snapshots to file. More... | |
| std::vector< std::vector< float > > | takeSnapshot () |
| Takes a snapshot of the current weight state. More... | |
The ConnectionMonitor class allows a user record weights from a particular connection. First, the method CARLsim::setConnectionMonitor must be called on a specific pair of pre-synaptic and post-synaptic group. This method then returns a pointer to a ConnectionMonitor object, which can be queried for connection data.
By default, a snapshot of all the weights will be taken once per second and dumped to file. The easiest way to use a ConnectionMonitor is to call CARLsim::setConnectionMonitor with file string "default". This will create a file with path "results/conn_{name of pre-group}_{name of post-group}.dat". It is also possible to specify a custom file string instead. Alternatively, the user may suppress creation of the binary file by using file string "null" instead.
Periodic storing can be disabled by calling ConnectionMonitor::setUpdateTimeInterval with argument intervalSec=-1.
Additionally, during a CARLsim simulation, the ConnectionMonitor object returned by CARLsim::setConnectionMonitor can be queried for connection data. The user may take a snapshot of the weights at any point in time using the method ConnectionMonitor::takeSnapshot. Note that a snapshot taken with this method will also be stored in the binary file. However, the binary file will never contain the same snapshot (captured with a certain timestamp) twice.
If at least two snapshots have been taken, the method ConnectionMonitor::calcWeightChanges will calculate the weight changes since the last snapshot. To make sure you are comparing the right snapshots, compare the timestamps returend by ConnectionMonitor::getTimeMsCurrentSnapshot and ConnectionMonitor::getTimeMsLastSnapshot.
Weights can be visualized in the Matlab Offline Analysis Toolbox (OAT) using the ConnectionMonitor utility. The OAT offers ways to plot 2D weight matrices, as well as receptive fields and response fields.
Weights can also be visualized in C++ using ConnectionMonitor::print and ConnectionMonitor::printSparse.
Example to store weights in binary every second:
Example that stores weight at the beginning and end of a simulation:
Example that periodically stores to binary and does some analysis on returned weight vector
writeToFile to true. Definition at line 142 of file connection_monitor.h.
| ConnectionMonitor | ( | ConnectionMonitorCore * | connMonCorePtr | ) |
Creates a new instance of the ConnectionMonitor class.
| ~ConnectionMonitor | ( | ) |
Cleans up all the memory upon object deletion.
| std::vector< std::vector<float> > calcWeightChanges | ( | ) |
This function calculates the difference between the current state of the weight matrix and what it was when taking the last snapshot. Weight change is reported for every synapse, in a 2D vector where the first dimension corresponds to pre-synaptic neuron ID and the second dimension corresponds to post-synaptic neuron ID. For example, element wtChange[3][8] of the return argumentr will indicate the signed weight change since the last snapshot for the synapse that connects preNeurId=3 to postNeurId=8. Synapses that are not allocated (i.e., that do not exist) are marked as float value NAN in the weight matrix. Synapses that do exist, but have zero weight, are marked as 0.0f in the weight matrix.
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
| short int getConnectId | ( | ) |
This function returns the connection ID that this ConnectionMonitor is managing. It is equivalent to the return argument of CARLsim::connect.
| int getFanIn | ( | int | neurPostId | ) |
This function returns the number of incoming synapses for a specific post-synaptic neuron ID.
| int getFanOut | ( | int | neurPreId | ) |
This function returns the number of outgoing synapses for a specific pre-synaptic neuron ID.
| double getMaxWeight | ( | bool | getCurrent = false | ) |
This function returns the maximum weight value of all synapses in the connection.
If getCurrent is set to true, then the function will return the currently largest weight value. In a plastic connection, this value might be different from the upper bound of the weight range specified when setting up the network (i.e., the max field of the RangeWeight struct). In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
If getCurrent is set to false, then the upper bound of the configured weight range will be returned.
| [in] | getCurrent | whether to return the currently largest weight value (true) or the upper bound of the weight range specified during setup (false). Default: false. |
| double getMinWeight | ( | bool | getCurrent = false | ) |
This function returns the minimum weight value of all synapses in the connection.
If getCurrent is set to true, then the function will return the currently smallest weight value. In a plastic connection, this value might be different from the lower bound of the weight range specified when setting up the network (i.e., the min field of the RangeWeight struct). In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
If getCurrent is set to false, then the lower bound of the configured weight range will be returned.
| [in] | getCurrent | whether to return the currently smallest weight value (true) or the lower bound of the weight range specified during setup (false). Default: false. |
| int getNumNeuronsPost | ( | ) |
This function returns the number of neurons in the post-synaptic group.
| int getNumNeuronsPre | ( | ) |
This function returns the number of neurons in the pre-synaptic group.
| int getNumSynapses | ( | ) |
This function returns the number of allocated synapses in the connection.
| int getNumWeightsChanged | ( | double | minAbsChanged = 1e-5 | ) |
This function returns the number of weights whose absolute value has changed at least minAbsChanged (inclusive) since the last snapshot was taken.
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
| [in] | minAbsChanged | the minimal value (inclusive) a weight has to have changed in order for it to be counted towards the number of changed synapses |
| int getNumWeightsInRange | ( | double | minValue, |
| double | maxValue | ||
| ) |
This function returns the number of synaptic weights whose values are within some specific range e[minVal,maxVal] (inclusive).
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
| [in] | minValue | the lower bound of the weight range (inclusive) |
| [in] | maxValue | the upper bound of the weight range (inclusive) |
| int getNumWeightsWithValue | ( | double | value | ) |
This function returns the number of synaptic weights that have exactly some specific value. It could be used to determine the sparsity of the connection matrix (wtValue==0.0f).
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
Machine epsilon (FLT_EPSILON) is used for floating point equality. That is, the weight value is considered equal to the input value if fabs(wt-value)<=FLT_EPSILON (inclusive).
This is a convenience function whose result is equivalent to getNumWeightsInRange(value-FLT_EPSILON,value+FLT_EPSILON).
| [in] | value | the exact weight value to look for |
| double getPercentWeightsChanged | ( | double | minAbsChanged = 1e-5 | ) |
This function returns the percentage of weights whose absolute has changed at least minAbsChanged (inclusive) since the last snapshot was taken.
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
This is a convenience function whose result is equivalent to getNumWeightsChanged()*100.0/getNumSynapses().
| [in] | minAbsChanged | the minimal value (inclusive) a weight has to have changed in order for it to be counted towards the percentage of changed synapses |
| double getPercentWeightsInRange | ( | double | minValue, |
| double | maxValue | ||
| ) |
This function returns the percentage of synaptic weights whose values are within some specific range e[minVal,maxVal] (inclusive).
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
This is a convenience function whose result is equivalent to getNumWeightsInRange(minValue,maxValue)*100.0/getNumSynapses().
| [in] | minValue | the lower bound of the weight range (inclusive) |
| [in] | maxValue | the upper bound of the weight range (inclusive) |
| double getPercentWeightsWithValue | ( | double | value | ) |
This function returns the percentage of synaptic weights that have exactly some specific value. It could be used to determine the sparsity of the connection matrix (wtValue==0.0f).
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
Machine epsilon (FLT_EPSILON) is used for floating point equality. That is, the weight value is considered equal to the input value if fabs(wt-value)<=FLT_EPSILON (inclusive).
This is a convenience function whose result is equivalent to getNumWeightsWithValue(value)*100.0/getNumSynapses().
| [in] | value | the exact weight value to look for |
| int64_t getTimeMsCurrentSnapshot | ( | ) |
This function returns the timestamp of the current weight snapshot, reported as the amount of time that has passed since the beginning of the simulation (in milliseconds). It will not take a snapshot by itself, so the time reported here is not necessarily equal to the time reported by CARLsim::getSimTime.
| int64_t getTimeMsLastSnapshot | ( | ) |
This function returns the timestamp of the last weight snapshot, reported as the amount of time that has passed since the beginning of the simulation (in milliseconds).
| int64_t getTimeMsSinceLastSnapshot | ( | ) |
This function returns the amount of time that has been passed between the current and last weight snapshot (reported in ms).
This is a convenience function whose result is equivalent to getTimeMsCurrentSnapshot()-getTimeMsLastSnapshot().
| double getTotalAbsWeightChange | ( | ) |
This function calculates the absolute sum of weight changes since the last snapshot was taken.
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
| void print | ( | ) |
This function prints the current state of the weights in a readable 2D weight matrix, where the first dimension corresponds to pre-synaptic neuron ID and the second dimension corresponds to post-synaptic neuron ID. Synapses that are not allocated (i.e., that do not exist) are marked as float value NAN in the weight matrix. Synapses that do exist, but have zero weight, are marked as 0.0f in the weight matrix.
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
| void printSparse | ( | int | neurPostId = ALL, |
| int | maxConn = 100, |
||
| int | connPerLine = 4 |
||
| ) |
This function prints the current state of the weights as a readable sparse list of weights. This is also the standard format used for ConnectionMonitor to report a run summary (see CARLsim::runNetwork). A weight will be reported as [preId,postId] wt (+/- weight change in the last x ms). User can control for which post-synaptic neurons the list should be generated, and set limits on how many connections to print in total and per line.
In order to get the current state of the weight matrix, this function will take a snapshot itself, but will not write it to file.
| [in] | neurPostId | The neuron ID of the post-synaptic group for which to generate the sparse weight list. Set to ALL to generate the list for all post-synaptic neurons. |
| [in] | maxConn | The maximum number of weights to print. |
| [in] | connPerLine | The number of weights to print per line. |
| void setUpdateTimeIntervalSec | ( | int | intervalSec | ) |
This function sets the time interval (seconds) for writing weight snapshots to file. The first snapshot will be written at time t=0, and every intervalSec seconds later.
In order to disable the periodic storing of weights to file alltogether, set intervalSec to -1. In this case, only weight snapshots acquired via ConnectionMonitor::takeSnapshots will end up in the binary.
| [in] | intervalSec | The update time interval (number of seconds) for writing snapshots to file. Set to -1 to disable periodic weight storing. Default: 1 (every second). |
| std::vector< std::vector<float> > takeSnapshot | ( | ) |
This function takes a snapshot of the current weight matrix, stores it in the binary as well as returns it as a 2D vector where the first dimension corresponds to pre-synaptic neuron ID and the second dimension corresponds to post-synaptic neuron ID. For example, element wtChange[3][8] of the return argumentr will indicate the signed weight change since the last snapshot for the synapse that connects preNeurId=3 to postNeurId=8. Synapses that are not allocated (i.e., that do not exist) are marked as float value NAN in the weight matrix. Synapses that do exist, but have zero weight, are marked as 0.0f in the weight matrix.
1.8.11