Hi,
I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows: - Initialise kernel - Get stimulation episodes - Create neurons (including spike generators) - Connect neurons - nest.Simulate() - nest.GetStatus(recording_neurons) for readout at relevant timepoints
I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so?
I have tried pickling
conn = nest.GetConnections()
but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is *no* connections, which means simply using
nest.SetConnections(conn)
prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need.
Cheers, Sock Ching
Dear Sock Ching,
When you say "I would like to save the network", could you elaborate on what exact part of the state you want to save and recall? For instance, in case of simulating plasticity, all the weights can be (re)set by calling SetStatus() on each connection object. The same goes for neuronal state variables. There are a few things that cannot be reset at the time of writing, such at the random number generators, but this is only an issue if you try to reproduce the same run *exactly*, multiple times in a row. The biggest issue might be that spike buffers (spikes in transit) cannot be reset. Potentially, you could work around this by allowing a small "startup transient", i.e. ignoring whatever your network produces in the first few hundred or so milliseconds after starting simulation. If this startup transient is causing you trouble, however, please feel invited to submit a feature request on our GitHub repository for spike buffer reset functionality. (Please provide as much technical detail as possible.)
There is already some prior discussion on this topic at https://github.com/nest/nest-simulator/issues/1618, where it was decided that just having a global "ResetNetwork()" function was not feasible, because it is not clear for the general case what this function would do. So in case you open a new GitHub issue, please try to specify as precisely as possible what parts of your network state need to be (re)set.
Hope this helps, please don't hesitate to share your further experiences.
Best regards, Charl
On Mon, Aug 3, 2020, at 19:19, LOW, SOCK CHING wrote:
Hi,
I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows:
- Initialise kernel
- Get stimulation episodes
- Create neurons (including spike generators)
- Connect neurons
- nest.Simulate()
- nest.GetStatus(recording_neurons) for readout at relevant timepoints
I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so?
I have tried pickling
conn = nest.GetConnections()
but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is *no* connections, which means simply using
nest.SetConnections(conn)
prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need.
Cheers, Sock Ching _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Charl,
Thanks for your prompt reply. What I mean to say is that I'd like to save the network parameters, specifically the connections (i.e. weights, delays etc.) and neuron parameters. There is no plasticity in the network, as I am simply initialising a network of neurons with probabilistic, random connections between each other (a liquid state machine, in other words).
My intention is to use the network in a real-time control architecture, so I don't need to replicate *exactly *the same output for a given input stimulation but I do need an output that is comparable. Currently, I am using the states of the network to train a classifier offline, but this classifier is fitted to the behaviour of the network that it trains on. To be able to use the classifier again, I will need to recreate that network, which leads me to the question of "saving the entire network". Once I can do so, it becomes possible to run simulations episode by episode rather than queueing all the episodes into a single batch for simulation, which is critical for my purpose. I hope that clarifies the question.
Cheers, Sock Ching
On Tue, 4 Aug 2020 at 09:01, Charl Linssen nest-users@turingbirds.com wrote:
Dear Sock Ching,
When you say "I would like to save the network", could you elaborate on what exact part of the state you want to save and recall? For instance, in case of simulating plasticity, all the weights can be (re)set by calling SetStatus() on each connection object. The same goes for neuronal state variables. There are a few things that cannot be reset at the time of writing, such at the random number generators, but this is only an issue if you try to reproduce the same run *exactly*, multiple times in a row. The biggest issue might be that spike buffers (spikes in transit) cannot be reset. Potentially, you could work around this by allowing a small "startup transient", i.e. ignoring whatever your network produces in the first few hundred or so milliseconds after starting simulation. If this startup transient is causing you trouble, however, please feel invited to submit a feature request on our GitHub repository for spike buffer reset functionality. (Please provide as much technical detail as possible.)
There is already some prior discussion on this topic at https://github.com/nest/nest-simulator/issues/1618, where it was decided that just having a global "ResetNetwork()" function was not feasible, because it is not clear for the general case what this function would do. So in case you open a new GitHub issue, please try to specify as precisely as possible what parts of your network state need to be (re)set.
Hope this helps, please don't hesitate to share your further experiences.
Best regards, Charl
On Mon, Aug 3, 2020, at 19:19, LOW, SOCK CHING wrote:
Hi,
I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows:
- Initialise kernel
- Get stimulation episodes
- Create neurons (including spike generators)
- Connect neurons
- nest.Simulate()
- nest.GetStatus(recording_neurons) for readout at relevant timepoints
I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so?
I have tried pickling
conn = nest.GetConnections()
but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is *no* connections, which means simply using
nest.SetConnections(conn)
prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need.
Cheers, Sock Ching _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hey all,
this sound similarly to my later suggestion in the reset issue: https://github.com/nest/nest-simulator/issues/1618#issuecomment-633972232 https://github.com/nest/nest-simulator/issues/1618#issuecomment-633972232
I wanted to do some more research on how loading and saving is already currently supported in the library (PyNest and c++ back-end side) before working on the implementation. I would be interested in working on this implementation or gladly join anyone, who wants to work on this.
Kind regards, Benedikt
Am 04.08.2020 um 10:15 schrieb LOW, SOCK CHING sockching.low@upf.edu:
Hi Charl,
Thanks for your prompt reply. What I mean to say is that I'd like to save the network parameters, specifically the connections (i.e. weights, delays etc.) and neuron parameters. There is no plasticity in the network, as I am simply initialising a network of neurons with probabilistic, random connections between each other (a liquid state machine, in other words).
My intention is to use the network in a real-time control architecture, so I don't need to replicate exactly the same output for a given input stimulation but I do need an output that is comparable. Currently, I am using the states of the network to train a classifier offline, but this classifier is fitted to the behaviour of the network that it trains on. To be able to use the classifier again, I will need to recreate that network, which leads me to the question of "saving the entire network". Once I can do so, it becomes possible to run simulations episode by episode rather than queueing all the episodes into a single batch for simulation, which is critical for my purpose. I hope that clarifies the question.
Cheers, Sock Ching
On Tue, 4 Aug 2020 at 09:01, Charl Linssen <nest-users@turingbirds.com mailto:nest-users@turingbirds.com> wrote: Dear Sock Ching,
When you say "I would like to save the network", could you elaborate on what exact part of the state you want to save and recall? For instance, in case of simulating plasticity, all the weights can be (re)set by calling SetStatus() on each connection object. The same goes for neuronal state variables. There are a few things that cannot be reset at the time of writing, such at the random number generators, but this is only an issue if you try to reproduce the same run *exactly*, multiple times in a row. The biggest issue might be that spike buffers (spikes in transit) cannot be reset. Potentially, you could work around this by allowing a small "startup transient", i.e. ignoring whatever your network produces in the first few hundred or so milliseconds after starting simulation. If this startup transient is causing you trouble, however, please feel invited to submit a feature request on our GitHub repository for spike buffer reset functionality. (Please provide as much technical detail as possible.)
There is already some prior discussion on this topic at https://github.com/nest/nest-simulator/issues/1618 https://github.com/nest/nest-simulator/issues/1618, where it was decided that just having a global "ResetNetwork()" function was not feasible, because it is not clear for the general case what this function would do. So in case you open a new GitHub issue, please try to specify as precisely as possible what parts of your network state need to be (re)set.
Hope this helps, please don't hesitate to share your further experiences.
Best regards, Charl
On Mon, Aug 3, 2020, at 19:19, LOW, SOCK CHING wrote:
Hi,
I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows:
- Initialise kernel
- Get stimulation episodes
- Create neurons (including spike generators)
- Connect neurons
- nest.Simulate()
- nest.GetStatus(recording_neurons) for readout at relevant timepoints
I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so?
I have tried pickling
conn = nest.GetConnections()
but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is no connections, which means simply using
nest.SetConnections(conn)
prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need.
Cheers, Sock Ching _______________________________________________ NEST Users mailing list -- users@nest-simulator.org mailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org mailto:users-leave@nest-simulator.org
NEST Users mailing list -- users@nest-simulator.org mailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org mailto:users-leave@nest-simulator.org _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hello all,
I remember seeing a presentation some years back where a large connectome was read directly into NEST from disk. I think it was stored in HDF5. One of the main complications was the round robin neuron placement complicating manners when a MPI rank size different then the stored version was used: all synapses needed to be transported to a different node. It was expensive.
Looking ahead, how some connectomes are generated in EBRAINS. Large pre-processing pipelines with complicated touch detectection and 3d placement of neurons (Markram Cell 2015) having this functionality might come in handy in the future.
2cents.
Wouter
On 04-Aug-20 10:33, Benedikt S. Vogler wrote:
Hey all,
this sound similarly to my later suggestion in the reset issue: https://github.com/nest/nest-simulator/issues/1618#issuecomment-633972232
I wanted to do some more research on how loading and saving is already currently supported in the library (PyNest and c++ back-end side) before working on the implementation. I would be interested in working on this implementation or gladly join anyone, who wants to work on this.
Kind regards, Benedikt
Am 04.08.2020 um 10:15 schrieb LOW, SOCK CHING <sockching.low@upf.edu mailto:sockching.low@upf.edu>:
Hi Charl,
Thanks for your prompt reply. What I mean to say is that I'd like to save the network parameters, specifically the connections (i.e. weights, delays etc.) and neuron parameters. There is no plasticity in the network, as I am simply initialising a network of neurons with probabilistic, random connections between each other (a liquid state machine, in other words).
My intention is to use the network in a real-time control architecture, so I don't need to replicate /exactly /the same output for a given input stimulation but I do need an output that is comparable. Currently, I am using the states of the network to train a classifier offline, but this classifier is fitted to the behaviour of the network that it trains on. To be able to use the classifier again, I will need to recreate that network, which leads me to the question of "saving the entire network". Once I can do so, it becomes possible to run simulations episode by episode rather than queueing all the episodes into a single batch for simulation, which is critical for my purpose. I hope that clarifies the question.
Cheers, Sock Ching
On Tue, 4 Aug 2020 at 09:01, Charl Linssen <nest-users@turingbirds.com mailto:nest-users@turingbirds.com> wrote:
__ Dear Sock Ching, When you say "I would like to save the network", could you elaborate on what exact part of the state you want to save and recall? For instance, in case of simulating plasticity, all the weights can be (re)set by calling SetStatus() on each connection object. The same goes for neuronal state variables. There are a few things that cannot be reset at the time of writing, such at the random number generators, but this is only an issue if you try to reproduce the same run *exactly*, multiple times in a row. The biggest issue might be that spike buffers (spikes in transit) cannot be reset. Potentially, you could work around this by allowing a small "startup transient", i.e. ignoring whatever your network produces in the first few hundred or so milliseconds after starting simulation. If this startup transient is causing you trouble, however, please feel invited to submit a feature request on our GitHub repository for spike buffer reset functionality. (Please provide as much technical detail as possible.) There is already some prior discussion on this topic at https://github.com/nest/nest-simulator/issues/1618, where it was decided that just having a global "ResetNetwork()" function was not feasible, because it is not clear for the general case what this function would do. So in case you open a new GitHub issue, please try to specify as precisely as possible what parts of your network state need to be (re)set. Hope this helps, please don't hesitate to share your further experiences. Best regards, Charl On Mon, Aug 3, 2020, at 19:19, LOW, SOCK CHING wrote:Hi, I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows: - Initialise kernel - Get stimulation episodes - Create neurons (including spike generators) - Connect neurons - nest.Simulate() - nest.GetStatus(recording_neurons) for readout at relevant timepoints I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so? I have tried pickling conn = nest.GetConnections() but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is /no/ connections, which means simply using nest.SetConnections(conn) prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need. Cheers, Sock Ching _______________________________________________ NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org>_______________________________________________ NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org>
NEST Users mailing list -- users@nest-simulator.org mailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org mailto:users-leave@nest-simulator.org
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
-- Wouter Klijn w.klijn@fz-juelich.de
Team Leader Multiscale simulation and design SimLab Neuroscience Jülich Supercomputing Centre Institute for Advanced Simulation Forschungszentrum Jülich http://www.fz-juelich.de/ias/jsc/slns
Office: +49 2461 61-3523 Fax # : +49 2461 61-6656
------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------ Forschungszentrum Juelich GmbH 52425 Juelich Sitz der Gesellschaft: Juelich Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Volker Rieke Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender), Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt ------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------
Hi everyone,
It seems like there is no straight-forward method of saving the network as a single object, so I think the simplest solution would be to save the relevant connection information and reconnect it accordingly when I reinitialise the network again. Since my network is pretty simple, this is not excessively demanding...the only real issue I've had is saving SLILiteral typed objects, but that's usually solved by turning it into a string.
Just in case someone else unfamiliar with nest finds themselves in this situation too, here are some details on the method to extract the information:
- Ensure that the handles of all distinct populations of created neurons are stored in variables (e.g. pop1 = nest.Create("iaf_psc_alpha", 50) retains a list of node handles in pop1). - Using nest.GetConnections(args): - Without args; returns *all *connections - args = pop1; returns all connections coming from the specified nodes - args = pop1, pop2; returns connections between the two populations - Using nest.GetStatus(conn, keys="weight") to get the parameters of the connection. - Save the parameters, there are many numbers of ways e.g. pickle.dump(), numpy.save()
If you know the populations in the network, you can get the connections between them (in both directions) and obtain their parameters with GetStatus (available keys: ['delay', 'receptor', 'sizeof', 'source', 'synapse_model', 'target', 'weight']).
To recreate the network, I choose to first create all the same populations then reinstate the connections using nest.Connect(pop1, pop2, 'one_to_one', syn_spec), with syn_spec being {'delay': [d1, d2,...,dn], 'weight': [w1, w2,...,wn], etc.} reconstructed from the saved data.
Cheers, Sock
On Wed, 5 Aug 2020 at 09:28, Wouter Klijn w.klijn@fz-juelich.de wrote:
Hello all,
I remember seeing a presentation some years back where a large connectome was read directly into NEST from disk. I think it was stored in HDF5. One of the main complications was the round robin neuron placement complicating manners when a MPI rank size different then the stored version was used: all synapses needed to be transported to a different node. It was expensive.
Looking ahead, how some connectomes are generated in EBRAINS. Large pre-processing pipelines with complicated touch detectection and 3d placement of neurons (Markram Cell 2015) having this functionality might come in handy in the future.
2cents.
Wouter
On 04-Aug-20 10:33, Benedikt S. Vogler wrote:
Hey all,
this sound similarly to my later suggestion in the reset issue:
https://github.com/nest/nest-simulator/issues/1618#issuecomment-633972232
I wanted to do some more research on how loading and saving is already currently supported in the library (PyNest and c++ back-end side) before working on the implementation. I would be interested in working on this implementation or gladly join anyone, who wants to work on this.
Kind regards, Benedikt
Am 04.08.2020 um 10:15 schrieb LOW, SOCK CHING <sockching.low@upf.edu mailto:sockching.low@upf.edu>:
Hi Charl,
Thanks for your prompt reply. What I mean to say is that I'd like to save the network parameters, specifically the connections (i.e. weights, delays etc.) and neuron parameters. There is no plasticity in the network, as I am simply initialising a network of neurons with probabilistic, random connections between each other (a liquid state machine, in other words).
My intention is to use the network in a real-time control architecture, so I don't need to replicate /exactly /the same output for a given input stimulation but I do need an output that is comparable. Currently, I am using the states of the network to train a classifier offline, but this classifier is fitted to the behaviour of the network that it trains on. To be able to use the classifier again, I will need to recreate that network, which leads me to the question of "saving the entire network". Once I can do so, it becomes possible to run simulations episode by episode rather than queueing all the episodes into a single batch for simulation, which is critical for my purpose. I hope that clarifies the question.
Cheers, Sock Ching
On Tue, 4 Aug 2020 at 09:01, Charl Linssen <nest-users@turingbirds.com mailto:nest-users@turingbirds.com> wrote:
__ Dear Sock Ching, When you say "I would like to save the network", could you elaborate on what exact part of the state you want to save and recall? For instance, in case of simulating plasticity, all the weights can be (re)set by calling SetStatus() on each connection object. The same goes for neuronal state variables. There are a few things that cannot be reset at the time of writing, such at the random number generators, but this is only an issue if you try to reproduce the same run *exactly*, multiple times in a row. The biggest issue might be that spike buffers (spikes in transit) cannot be reset. Potentially, you could work around this by allowing a small "startup transient", i.e. ignoring whatever your network produces in the first few hundred or so milliseconds after starting simulation. If this startup transient is causing you trouble, however, please feel invited to submit a feature request on our GitHub repository for spike buffer reset functionality. (Please provide as much technical detail as possible.) There is already some prior discussion on this topic at https://github.com/nest/nest-simulator/issues/1618, where it was decided that just having a global "ResetNetwork()" function was not feasible, because it is not clear for the general case what this function would do. So in case you open a new GitHub issue, please try to specify as precisely as possible what parts of your network state need to be (re)set. Hope this helps, please don't hesitate to share your further experiences. Best regards, Charl On Mon, Aug 3, 2020, at 19:19, LOW, SOCK CHING wrote:Hi, I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows: - Initialise kernel - Get stimulation episodes - Create neurons (including spike generators) - Connect neurons - nest.Simulate() - nest.GetStatus(recording_neurons) for readout at relevant timepoints I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so? I have tried pickling conn = nest.GetConnections() but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is /no/ connections, which means simply using nest.SetConnections(conn) prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need. Cheers, Sock Ching _______________________________________________ NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org>_______________________________________________ NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org>
NEST Users mailing list -- users@nest-simulator.org mailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org mailto:users-leave@nest-simulator.org
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
-- Wouter Klijn w.klijn@fz-juelich.de
Team Leader Multiscale simulation and design SimLab Neuroscience Jülich Supercomputing Centre Institute for Advanced Simulation Forschungszentrum Jülich http://www.fz-juelich.de/ias/jsc/slns
Office: +49 2461 61-3523 Fax # : +49 2461 61-6656
Forschungszentrum Juelich GmbH 52425 Juelich Sitz der Gesellschaft: Juelich Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Volker Rieke Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender), Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi,
I'm a bit late, but since this might help users who have similar workflows, I just wanted to signal that you can have a look at NNGT: nngt.readthedocs.io/ http://nngt.readthedocs.io/ and https://github.com/Silmathoron/NNGT
And specifically how to interact with NEST: https://nngt.readthedocs.io/en/stable/user/tutorial.html#real-neuronal-cultu...
This would enable you to create a network beforehand and save it to a file, then just load it and send it to NEST using network = nngt.load_from_file("the-file") and network.to_nest().
Note that this works only with static weights and will not capture any information about the evolution of weights or connections via STDP or structural plasticity.
(Disclaimer: I am the library's main developer)
Best, Tanguy
On 05/08/2020 10:33, LOW, SOCK CHING wrote:
Hi everyone,
It seems like there is no straight-forward method of saving the network as a single object, so I think the simplest solution would be to save the relevant connection information and reconnect it accordingly when I reinitialise the network again. Since my network is pretty simple, this is not excessively demanding...the only real issue I've had is saving SLILiteral typed objects, but that's usually solved by turning it into a string.
Just in case someone else unfamiliar with nest finds themselves in this situation too, here are some details on the method to extract the information:
- Ensure that the handles of all distinct populations of created neurons are stored in variables (e.g. pop1 = nest.Create("iaf_psc_alpha", 50) retains a list of node handles in pop1).
- Using nest.GetConnections(args): o Without args; returns /all /connections o args = pop1; returns all connections coming from the specified nodes o args = pop1, pop2; returns connections between the two populations
- Using nest.GetStatus(conn, keys="weight") to get the parameters of the connection.
- Save the parameters, there are many numbers of ways e.g. pickle.dump(), numpy.save()
If you know the populations in the network, you can get the connections between them (in both directions) and obtain their parameters with GetStatus (available keys: ['delay', 'receptor', 'sizeof', 'source', 'synapse_model', 'target', 'weight']).
To recreate the network, I choose to first create all the same populations then reinstate the connections using nest.Connect(pop1, pop2, 'one_to_one', syn_spec), with syn_spec being {'delay': [d1, d2,...,dn], 'weight': [w1, w2,...,wn], etc.} reconstructed from the saved data.
Cheers, Sock
Hey all,
this sound similarly to my later suggestion in the reset issue: https://github.com/nest/nest-simulator/issues/1618#issuecomment-633972232 https://github.com/nest/nest-simulator/issues/1618#issuecomment-633972232
I wanted to do some more research on how loading and saving is already currently supported in the library (PyNest and c++ back-end side) before working on the implementation. I would be interested in working on this implementation or gladly join anyone, who wants to work on this.
Kind regards, Benedikt
Am 04.08.2020 um 10:15 schrieb LOW, SOCK CHING sockching.low@upf.edu:
Hi Charl,
Thanks for your prompt reply. What I mean to say is that I'd like to save the network parameters, specifically the connections (i.e. weights, delays etc.) and neuron parameters. There is no plasticity in the network, as I am simply initialising a network of neurons with probabilistic, random connections between each other (a liquid state machine, in other words).
My intention is to use the network in a real-time control architecture, so I don't need to replicate exactly the same output for a given input stimulation but I do need an output that is comparable. Currently, I am using the states of the network to train a classifier offline, but this classifier is fitted to the behaviour of the network that it trains on. To be able to use the classifier again, I will need to recreate that network, which leads me to the question of "saving the entire network". Once I can do so, it becomes possible to run simulations episode by episode rather than queueing all the episodes into a single batch for simulation, which is critical for my purpose. I hope that clarifies the question.
Cheers, Sock Ching
On Tue, 4 Aug 2020 at 09:01, Charl Linssen <nest-users@turingbirds.com mailto:nest-users@turingbirds.com> wrote: Dear Sock Ching,
When you say "I would like to save the network", could you elaborate on what exact part of the state you want to save and recall? For instance, in case of simulating plasticity, all the weights can be (re)set by calling SetStatus() on each connection object. The same goes for neuronal state variables. There are a few things that cannot be reset at the time of writing, such at the random number generators, but this is only an issue if you try to reproduce the same run *exactly*, multiple times in a row. The biggest issue might be that spike buffers (spikes in transit) cannot be reset. Potentially, you could work around this by allowing a small "startup transient", i.e. ignoring whatever your network produces in the first few hundred or so milliseconds after starting simulation. If this startup transient is causing you trouble, however, please feel invited to submit a feature request on our GitHub repository for spike buffer reset functionality. (Please provide as much technical detail as possible.)
There is already some prior discussion on this topic at https://github.com/nest/nest-simulator/issues/1618 https://github.com/nest/nest-simulator/issues/1618, where it was decided that just having a global "ResetNetwork()" function was not feasible, because it is not clear for the general case what this function would do. So in case you open a new GitHub issue, please try to specify as precisely as possible what parts of your network state need to be (re)set.
Hope this helps, please don't hesitate to share your further experiences.
Best regards, Charl
On Mon, Aug 3, 2020, at 19:19, LOW, SOCK CHING wrote:
Hi,
I am new to pynest and am using it to implement a liquid state machine. It works marvelously for generating the network. My workflow is as follows:
- Initialise kernel
- Get stimulation episodes
- Create neurons (including spike generators)
- Connect neurons
- nest.Simulate()
- nest.GetStatus(recording_neurons) for readout at relevant timepoints
I would like to save the network so I can run nest.Simulate() on other episodes in a real-time application, is there a straight-forward way to do so?
I have tried pickling
conn = nest.GetConnections()
but to my knowledge that does not include a lot of details about the connection, like the weight. It also does not return where there is no connections, which means simply using
nest.SetConnections(conn)
prior to nest.Simulate() will not work to reproduce the previously generated network even if I initialise with all the same variables. I have found the functions GetNetwork() and GetNodes() but I'm not sure how I can use them, or if they are even useful for what I need.
Cheers, Sock Ching _______________________________________________ NEST Users mailing list -- users@nest-simulator.org mailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org mailto:users-leave@nest-simulator.org
NEST Users mailing list -- users@nest-simulator.org mailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org mailto:users-leave@nest-simulator.org _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
-
Benedikt S. Vogler -
Benedikt S. Vogler -
Charl Linssen -
LOW, SOCK CHING -
Tanguy Fardet -
Wouter Klijn