Hi!
I am trying to implement weight normalization (as you show in your documentation). It works well (see below code) when executed "normally" (or when using" mpirun -np 1"), but it hangs (it just do not respond) when executed using mpirun. It hangs when I try to execute nest.Simulate() after weigth assignment. If I remove the "conn.weight = w_nparray" instruction in the following code, it works without hang.
I use NEST 3.6 and python3, compiled with python, mpi, sionlib and music.
Thanks a lot in advance!
import nest import numpy as np
pos = nest.spatial.grid(shape = [30,30] ) input = nest.Create('iaf_psc_alpha', positions=pos)
layer_0 = nest.Create('iaf_psc_alpha', positions=pos) layer_1 = nest.Create('iaf_psc_alpha', positions=pos)
conn_neur = {'rule':'pairwise_bernoulli', 'use_on_source': True, 'mask': {'grid':{'shape':[9,9]}} }
nest.CopyModel('static_synapse', 'syn_1_model') syn_0 = {'synapse_model': 'static_synapse'} syn_1 = {'synapse_model': 'syn_1_model'}
nest.Connect(input, layer_0, conn_neur, syn_0) nest.Connect(input, layer_1, conn_neur, syn_0)
nest.DumpLayerConnections(input, layer_0, 'static_synapse', 'conn.txt') nest.DumpLayerConnections(input, layer_1, 'syn_1_model', 'conn.txt')
nest.Simulate(100)
for neuron in layer_0: conn = nest.GetConnections(source=input, target=neuron, synapse_model='static_synapse') w = np.array(conn.weight) if (w.size>1): w_normed = w / sum(abs(w)) # L1-norm w_nparray = 2. * w_normed conn.weight = w_nparray
nest.Simulate(100)
Hi Xavier,
I think I can reproduce your problem and have an idea what it may cause it. How many MPI processes did you use?
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Wednesday, 7 February 2024 at 15:41 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Weights assignment error with mpirun [Some people who received this message don't often get email from xotazu@cvc.uab.cat. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
Hi!
I am trying to implement weight normalization (as you show in your documentation). It works well (see below code) when executed "normally" (or when using" mpirun -np 1"), but it hangs (it just do not respond) when executed using mpirun. It hangs when I try to execute nest.Simulate() after weigth assignment. If I remove the "conn.weight = w_nparray" instruction in the following code, it works without hang.
I use NEST 3.6 and python3, compiled with python, mpi, sionlib and music.
Thanks a lot in advance!
import nest import numpy as np
pos = nest.spatial.grid(shape = [30,30] ) input = nest.Create('iaf_psc_alpha', positions=pos)
layer_0 = nest.Create('iaf_psc_alpha', positions=pos) layer_1 = nest.Create('iaf_psc_alpha', positions=pos)
conn_neur = {'rule':'pairwise_bernoulli', 'use_on_source': True, 'mask': {'grid':{'shape':[9,9]}} }
nest.CopyModel('static_synapse', 'syn_1_model') syn_0 = {'synapse_model': 'static_synapse'} syn_1 = {'synapse_model': 'syn_1_model'}
nest.Connect(input, layer_0, conn_neur, syn_0) nest.Connect(input, layer_1, conn_neur, syn_0)
nest.DumpLayerConnections(input, layer_0, 'static_synapse', 'conn.txt') nest.DumpLayerConnections(input, layer_1, 'syn_1_model', 'conn.txt')
nest.Simulate(100)
for neuron in layer_0: conn = nest.GetConnections(source=input, target=neuron, synapse_model='static_synapse') w = np.array(conn.weight) if (w.size>1): w_normed = w / sum(abs(w)) # L1-norm w_nparray = 2. * w_normed conn.weight = w_nparray
nest.Simulate(100) _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi!
I used 2 and 64 number of processes.
An update ... the problem disappears when using "nest.SetStatus(conn,'weight',w_nparray)" instead of the direct assignment "conn.weight=w_nparray". In addition, it also hangs when using "conn.set()"
Xavier
Hi Xavier,
I have now created an issue for this on Github: https://github.com/nest/nest-simulator/issues/3099
But there is another problem with your approach: Even if the code worked with MPI, you would only collect the weights of the connections local to each MPI rank on that rank (i.e., connections with targets on that rank). If normalization per target neuron is what you want, this is fine, if you wanted to normalize all outgoing synapses of a neuron, this would be pretty much impossible when using MPI.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Wednesday, 7 February 2024 at 16:36 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu@cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi!
I used 2 and 64 number of processes.
An update ... the problem disappears when using "nest.SetStatus(conn,'weight',w_nparray)" instead of the direct assignment "conn.weight=w_nparray". In addition, it also hangs when using "conn.set()"
Xavier _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier
Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simula....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Thursday, 8 February 2024 at 19:46 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu@cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Thursday, 8 February 2024 at 21:42 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu@cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
_______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________
NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
Hi Xavier,
If you know the time course of you poisson process rates beforehand, you can use the inhomogeneous_poisson_generator, which changes to new rates at given time points.
The rate calculation of the poisson generator is precisely on of the cases where SetStatus inside a RM context causes problems. This will not change, because one of the key points about the RM context is to avoid time overhead for neuron configuration.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Friday, 9 February 2024 at 11:44 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu@cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________
NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
_______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hans,
Inhomogeneous spike generator is a very nice solution ... Thanks for the tip! I didn't thought about it :-) I will try it (although at this moment my main time consuming process is the normalization step, not the Simulate() preparation time :-/ )
Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
If you know the time course of you poisson process rates beforehand, you can use the inhomogeneous_poisson_generator, which changes to new rates at given time points.
The rate calculation of the poisson generator is precisely on of the cases where SetStatus inside a RM context causes problems. This will not change, because one of the key points about the RM context is to avoid time overhead for neuron configuration.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 11:44 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________
NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
Hans,
Hi Xavier,
I think I understand your MPI-hang problems now (see comment in the issue) and can suggest a work-around
if (w.size>1): w_normed = w / sum(abs(w)) # L1-norm w_nparray = 2. * w_normed conn.weight = w_nparray else: conn.weight = []
The important thing is that the SynapseCollection.set() method, which is called by the assignment, is executed by all MPI ranks, even if their conn is empty. Assigning to an empty conn just does nothing, except MPI communicate with the other ranks.
The other work around is to call SetStatus(), which avoids the MPI calls altogether and thus is most likely more efficient.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Friday, 9 February 2024 at 14:00 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu@cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hans,
Inhomogeneous spike generator is a very nice solution ... Thanks for the tip! I didn't thought about it :-) I will try it (although at this moment my main time consuming process is the normalization step, not the Simulate() preparation time :-/ )
Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
If you know the time course of you poisson process rates beforehand, you can use the inhomogeneous_poisson_generator, which changes to new rates at given time points.
The rate calculation of the poisson generator is precisely on of the cases where SetStatus inside a RM context causes problems. This will not change, because one of the key points about the RM context is to avoid time overhead for neuron configuration.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 11:44 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________
NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
Hans, _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Xavier,
I have now created a PR that should fix the problem with hanging simulations. You can find it at
https://github.com/nest/nest-simulator/pull/3101
I’d much appreciate it if you could check that it solves your problem.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Hans Ekkehard Plesser hans.ekkehard.plesser@nmbu.no Date: Sunday, 11 February 2024 at 16:46 To: NEST User Mailing List users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun
Hi Xavier,
I think I understand your MPI-hang problems now (see comment in the issue) and can suggest a work-around
if (w.size>1): w_normed = w / sum(abs(w)) # L1-norm w_nparray = 2. * w_normed conn.weight = w_nparray else: conn.weight = []
The important thing is that the SynapseCollection.set() method, which is called by the assignment, is executed by all MPI ranks, even if their conn is empty. Assigning to an empty conn just does nothing, except MPI communicate with the other ranks.
The other work around is to call SetStatus(), which avoids the MPI calls altogether and thus is most likely more efficient.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Friday, 9 February 2024 at 14:00 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu@cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hans,
Inhomogeneous spike generator is a very nice solution ... Thanks for the tip! I didn't thought about it :-) I will try it (although at this moment my main time consuming process is the normalization step, not the Simulate() preparation time :-/ )
Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
If you know the time course of you poisson process rates beforehand, you can use the inhomogeneous_poisson_generator, which changes to new rates at given time points.
The rate calculation of the poisson generator is precisely on of the cases where SetStatus inside a RM context causes problems. This will not change, because one of the key points about the RM context is to avoid time overhead for neuron configuration.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 11:44 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________
NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
Hans, _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Hans,
This "else" workaround doesn't work. It also hangs.
I have not tried yet the PR (I have to download it and recompile and reinstall NEST). Will tell you later (today).
Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
I think I understand your MPI-hang problems now (see comment in the issue) and can suggest a work-around
if (w.size>1): w_normed = w / sum(abs(w)) # L1-norm w_nparray = 2. * w_normed conn.weight = w_nparray else: conn.weight = []The important thing is that the SynapseCollection.set() method, which is called by the assignment, is executed by all MPI ranks, even if their conn is empty. Assigning to an empty conn just does nothing, except MPI communicate with the other ranks.
The other work around is to call SetStatus(), which avoids the MPI calls altogether and thus is most likely more efficient.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 14:00 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hans,
Inhomogeneous spike generator is a very nice solution ... Thanks for the tip! I didn't thought about it :-) I will try it (although at this moment my main time consuming process is the normalization step, not the Simulate() preparation time :-/ )
Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
If you know the time course of you poisson process rates beforehand, you can use the inhomogeneous_poisson_generator, which changes to new rates at given time points.
The rate calculation of the poisson generator is precisely on of the cases where SetStatus inside a RM context causes problems. This will not change, because one of the key points about the RM context is to avoid time overhead for neuron configuration.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 11:44 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are added or synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context for higher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is so cumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig p? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me.
Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections() for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...)
for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...)
Many thanks in advance!
Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org Hans,
NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
Hans,
The PR works. There is no hang! I downloaded the master from the git and applied your patch. In comparison with my previous version, now I compiled NEST without MUSIC support.
In addition, the execution is MUCH faster. The time to perform the normalization (including the weight assignment) is around 30-40 times faster!!! May be it is because I use 64 mpi processes (mpirun -np 64)?
Best, Xavier
Xavier Otazu wrote:
Hi Hans,
This "else" workaround doesn't work. It also hangs.
I have not tried yet the PR (I have to download it and recompile and reinstall NEST). Will tell you later (today).
Xavier
Hans Ekkehard Plesser wrote:
Hi Xavier,
I think I understand your MPI-hang problems now (see comment in the issue) and can suggest a work-around
if (w.size>1): w_normed = w / sum(abs(w)) # L1-norm w_nparray = 2. * w_normed conn.weight = w_nparray else: conn.weight = []The important thing is that the SynapseCollection.set() method, which is called by the assignment, is executed by all MPI ranks, even if their conn is empty. Assigning to an empty conn just does nothing, except MPI communicate with the other ranks.
The other work around is to call SetStatus(), which avoids the MPI calls altogether and thus is most likely more efficient.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 14:00 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hans,
Inhomogeneous spike generator is a very nice solution ... Thanks for the tip! I didn't thought about it :-) I will try it (although at this moment my main time consuming process is the normalization step, not the Simulate() preparation time :-/ )
Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
If you know the time course of you poisson process rates beforehand, you can use the inhomogeneous_poisson_generator, which changes to new rates at given time points.
The rate calculation of the poisson generator is precisely on of the cases where SetStatus inside a RM context causes problems. This will not change, because one of the key points about the RM context is to avoid time overhead for neuron configuration.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Friday, 9 February 2024 at 11:44 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Inside an RM context, I am not only changing synaptic weights, but also Poisson spike generators rates. At every loop iteration I change the "input image", and I simulate it using a spike generator. I tried an RM context some time ago (previous NEST versions) and I realized that rate didn't change (I guess because the non-updated internal variables). Is this behaviour changed on 3.6 version?
Related to the neuroscientific question I understand you ask me about the temporal evolution of weights when externally modifying them. When I just used STDP, my weights grow indefinitely (a well known problem) or just saturates to W_max. Since in many models input weights are normalized, I applied normalization to avoid this growing problem (in addition, there are some biophyisical studies showing that this kind of normalization exists in biological systems). Another question is the "continuity" (first derivative) of temporal evolution of weights ... the answer is "I do not know" :-/
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
Actually, in your case the Prepare-Run-Cleanup approach should work, since you use SetStatus only on synaptic weights. The reason we warn about unpredictable behaviour in case of SetStatus inside a RunManager context is that Prepare() calls a pre_run_hook() on all network nodes to allow them to precompute internal variables. So if you changed neuron parameters inside an RM context, the nodes would update with an inconsistent set of externally set parameter values and internally not-updated variables. Synapses do not have such a pre-run-hook, so there you should be way safe. Note that you cannot Connect/Disconnect in a RM context, since the connection infrastructure needs to be updated by Prepare().
Finally, one could wonder about how the external manipulation of the synaptic weights interacts with the internal evolution of the weights through STDP, i.e., what the actual time-evolution equations for your weights become when taking normalization into account. But that is a neuroscientific question, not a NEST-technical one.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 21:42 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktig på https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans,
Great news to be able to call GetConnections() just once ... it speedsA LOT my runnings (from 230-ish seconds to 90-ish) :-)
I tried the Prepare/Run/Cleanup some time ago (nice solution), but since I have to use SetStatus() to update normalized weights between Prepare() and Cleanup(), I cannot use it :-(
Best, Xavier
Hans Ekkehard Plesser wrote: Hi Xavier,
The result of `GetConnections()` remains valid as long as no new synapses are addedor synapses deleted, so in your case, you only need to call it once. When you then do a `conn.weight`, you will get the weights at that point in time. But note that NEST only updates synaptic weights when a spike actually passes through the synapse, so you will get for each synapse the weight as it was at the end of the plasticity update when the most recent spike passed through that synapse.
Instead of calling Simulate() repeatedly, you can also use a RunManager context forhigher efficiency, see https://nest-simulator.readthedocs.io/en/stable/nest_behavior/running_simul%....
In case you should be wondering why global instantaneous normalization is socumbersome in NEST: We have focused on support for biologically plausible networks so far and instantaneous global normalization is just not biologically plausible (nor physically feasible with a view to causality).
Best, Hans Ekkehard -- Prof. Dr. Hans Ekkehard Plesser Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.no<mailto:hans.ekkehard.plesser@nmbu.no> Home http://arken.nmbu.no/~plesser From: Xavier Otazu <xotazu(a)cvc.uab.cat> Date: Thursday, 8 February 2024 at 19:46 To: users(a)nest-simulator.org <users(a)nest-simulator.org> Subject: [NEST Users] Re: Weights assignment error with mpirun [Du mottar ikke ofte e-post fra xotazu(a)cvc.uab.cat. Finn ut hvorfor dette er viktigp? https://aka.ms/LearnAboutSenderIdentification ]
Hi Hans, Thanks for your information. Fortunately, I do not normalize output connections (only input), hence, this code works for me. Another question. Take that I call nest.Simulate() inside a for loop and I want to normalize weights after every iteration. In order to avoid calling GetConnections()for every loop iteration, may I get the connections calling GetConnections() just once? May I safely suppose that after every Simulate() new weights of my connections (they are changed because I use STDP) will be available (for reading) using the same variable? An example code would be something like:
my_conn = GetConnections(...) for iter in range(...) nest.Simulate(...) normalize_connections(my_conn) # function that normalizes weights and calls nest.SetStatus(my_conn, ...) Many thanks in advance! Xavier _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org_______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org Hans, _______________________________________________ NEST Users mailing list -- users(a)nest-simulator.org To unsubscribe send an email to users-leave(a)nest-simulator.org
A small clarification ... I always used 64 mpi processes to run my code (hence, the 30-40 speedup performace is because of your new code).
Xavier
Dear Xavier,
30–40 times faster is indeed a very pleasant result 😊. The reason is that the PyNEST interface induced MPI communication for each of the weight updates, and if 64 ranks need to sync all the time, the going gets slow.
Best, Hans Ekkehard
--
Prof. Dr. Hans Ekkehard Plesser
Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway
Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.nomailto:hans.ekkehard.plesser@nmbu.no Home http://arken.nmbu.no/~plesser
From: Xavier Otazu xotazu@cvc.uab.cat Date: Monday, 12 February 2024 at 15:12 To: users@nest-simulator.org users@nest-simulator.org Subject: [NEST Users] Re: Weights assignment error with mpirun [Some people who received this message don't often get email from xotazu@cvc.uab.cat. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
A small clarification ... I always used 64 mpi processes to run my code (hence, the 30-40 speedup performace is because of your new code).
Xavier _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
-
Hans Ekkehard Plesser -
Xavier Otazu