Add option to generate connections from NumPy arrays #1429

hakonsbm · 2020-02-13T09:41:09Z

Connections are created one-to-one based on NumPy arrays of source and target node IDs, and a synapse specification containing NumPy arrays of weights and delays. All arrays have to be of equal length. For efficiency, the PyNEST kernel bypasses SLI and passes a pointer for each array to the C++ level.

It can be used as in the following example:

nest.Create('iaf_psc_alpha', 10)
# sources and targets must be NumPy arrays of equal lengths, with integer dtype
sources = np.array([1, 5, 7], dtype=np.uint64)
targets = np.array([2, 4, 6], dtype=np.uint64)
# weights, delays, and receptor type must be NumPy arrays with lengths equal to 
# that of sources and targets, or unspecified
weights = np.ones(len(sources))
delays = np.ones(len(sources))

# conn_spec cannot be specified
nest.Connect(sources, targets, syn_spec={'weight': weights, 'delay': delays,
                                         'synapse_model': 'static_synapse'})

This replaces the implementation of Connect_nonunique.

Fixes #1423.

alberto-antonietti · 2020-02-14T15:51:47Z

Hi @hakonsbm!
Does this work also with non-unique sources and/or targets? Something like:

sources = np.array([1, 5, 7, 5], dtype=np.uint64)
targets = np.array([2, 4, 6, 6], dtype=np.uint64)

hakonsbm · 2020-02-17T10:26:26Z

@alberto-antonietti Yes, it is possible to use non-unique elements.

alberto-antonietti · 2020-02-17T17:24:18Z

@alberto-antonietti Yes, it is possible to use non-unique elements.

Great! That is what we would like to have with #1423 and HBP Ticket #482997.

Made weight and delay optional, added optional receptor_type specification.

jougs · 2020-02-25T09:54:02Z

@alberto-antonietti: the reference to this PR in your comment is most likely not what you wanted. Can you please check and update? Thanks!

jougs

Very, very nice! My comments are plenty but rather minor. Many thanks!

doc/guides/from_nest2_to_nest3.rst

pynest/nest/lib/hl_api_connections.py

pynest/nest/tests/test_connect_arrays.py

sarakonradi

Thanks a lot for your work! I agree with Jochen's comments. Also added a few minor suggestions!

doc/guides/from_nest2_to_nest3.rst

sarakonradi · 2020-02-25T11:01:03Z

nestkernel/node_collection.h

@@ -194,6 +194,14 @@ class NodeCollection
   */
  static NodeCollectionPTR create( const TokenArray& node_ids );

+  /**
+   * Create a NodeCollection from a single node ID. Results in a primitive.


Results in a primitive: Is this sentence complete?

A primitive is a NodeCollection that is contiguous and homogeneous. It is used as a noun throughout the file, so if the wording is confusing, it probably has to be changed everywhere.

Ah, I see. In that case it is best to leave primitive as it is.

alberto-antonietti · 2020-02-25T15:32:08Z

@alberto-antonietti: the reference to this PR in your comment is most likely not what you wanted. Can you please check and update? Thanks!

Hi @jougs, I updated the reference, sorry!

@hakonsbm I do not know if I can add something meaningful to the two reviews of @jougs and @sarakonradi, but I will try to test it as soon as I have a bit of time (given the incoming end of SGA2)

Co-Authored-By: Jochen Martin Eppler <jougs@gmx.net>

jougs

Many thanks for addressing all my comments in such a timely fashion. Good to go from my side.

alberto-antonietti

Thanks @hakonsbm for the great work!

I have added a few comments from the "user" perspective, to provide better information about the exact errors that could be triggered.

Do you think that the limitation of setting only synapse_model, weights, delays, and receptors could be relaxed?
I do not know how hard that could be, but I can think of some of my simulations where I connect non-unique neurons (therefore I cannot use the standard nest.Connect() between NodeCollections) and I set other syn_spec, especially if I have synaptic plasticity.

alberto-antonietti · 2020-03-05T15:28:25Z

doc/guides/from_nest2_to_nest3.rst

+   nest.Create('iaf_psc_alpha', 10)
+   # Node IDs in the arrays must address existing nodes, but may occur multiple times.
+   sources = np.array([1, 5, 7, 5], dtype=np.uint64)
+   targets = np.array([2, 2, 4, 4], dtype=np.uint64)
+   nest.Connect(sources, targets)


If I run this code snippet, I get an error:

----> 5 nest.Connect(sources, targets) ~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/ll_api.py in stack_checker_func(*args, **kwargs) 245 def stack_checker_func(*args, **kwargs): 246 if not get_debug(): --> 247 return f(*args, **kwargs) 248 else: 249 sr('count') ~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/lib/hl_api_connections.py in Connect(pre, post, conn_spec, syn_spec, return_synapsecollection) 224 if return_synapsecollection: 225 raise ValueError("SynapseCollection cannot be returned when connecting two arrays of node IDs") --> 226 weights = numpy.array(processed_syn_spec['weight']) if 'weight' in processed_syn_spec else None 227 delays = numpy.array(processed_syn_spec['delay']) if 'delay' in processed_syn_spec else None 228 receptor_type = (numpy.array(processed_syn_spec['receptor_type']) TypeError: argument of type 'NoneType' is not iterable

I do not get it if I specify the synapse model ( nest.Connect(sources, targets, syn_spec={"synapse_model" : "static_synapse"}) )

Would it be useful to force the specification of synapse_model?

Yes, we are forcing the specification of synapse_model. I just forgot to update the documentation, but it's fixed now.

alberto-antonietti · 2020-03-05T16:02:06Z

pynest/nest/lib/hl_api_connections.py

+        if len(set(processed_syn_spec.keys()) - set(['weight', 'delay', 'synapse_model', 'receptor_type'])) != 0:
+            raise ValueError("When connecting two arrays of node IDs, the synapse specification dictionary can "
+                             "only contain weights, delays, synapse model, and r_port.")
+        connect_arrays(pre, post, weights, delays, receptor_type, synapse_model)


I see that its documented, however if one calls by mistake something like nest.Connect(pre, post, syn_spec={"synapse_model" : "static_synapse", "weight": 2.0}), the error message he gets is not very informative.

~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/ll_api.py in stack_checker_func(*args, **kwargs) 245 def stack_checker_func(*args, **kwargs): 246 if not get_debug(): --> 247 return f(*args, **kwargs) 248 else: 249 sr('count') ~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/lib/hl_api_connections.py in Connect(pre, post, conn_spec, syn_spec, return_synapsecollection) 237 raise ValueError("When connecting two arrays of node IDs, the synapse specification dictionary can " 238 "only contain weights, delays, synapse model, and r_port.") --> 239 connect_arrays(pre, post, weights, delays, receptor_type, synapse_model) 240 return 241 pynestkernel.pyx in pynestkernel.NESTEngine.connect_arrays() TypeError: len() of unsized object

Apparently one can create a 0-dimension NumPy array with for example np.array(2.0), which we didn't check for when connecting. I have added some checks for it now.

alberto-antonietti · 2020-03-05T16:09:55Z

pynest/nest/lib/hl_api_connection_helpers.py

@@ -233,9 +232,3 @@ def fixdict(d):
    projections = fixdict(projections)


if "weight" is a list/np.array of len != len(sources) and len(targets), the following error is Raised.
e.g.

n=10 nest.Create("iaf_cond_exp") pre = np.random.randint(1, n+1, 100) post = np.copy(pre) np.random.shuffle(post) nest.Connect(pre, post, syn_spec={"synapse_model" : "static_synapse", "weight": [2.0, 3.0]})

Error:

~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/lib/hl_api_connections.py in Connect(pre, post, conn_spec, syn_spec, return_synapsecollection) 217 # to the right formats. 218 processed_syn_spec = _process_syn_spec( --> 219 syn_spec, processed_conn_spec, len(pre), len(post)) 220 221 # If pre and post are arrays of node IDs, and conn_spec is unspecified, ~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/lib/hl_api_connection_helpers.py in _process_syn_spec(syn_spec, conn_spec, prelength, postlength) 74 raise kernel.NESTError( 75 "'" + key + "' has to be an array of " ---> 76 "dimension " + str(prelength) + ", a " 77 "scalar or a dictionary.") 78 else: NESTError: 'weight' has to be an array of dimension 100, a scalar or a dictionary.

If we are connective np.arrays, the error should say that 'weight' has to be an array of dimension 100, right?

Nice catch! I've updated the error messages.

Silmathoron · 2020-04-03T16:14:07Z

Is there a reason against just converting lists or numpy arrays to NodeCollections if such inputs are provided?

heplesser · 2020-04-03T18:10:46Z

Is there a reason against just converting lists or numpy arrays to NodeCollections if such inputs are provided?

Any NodeID can occur only once in a NodeCollection, while repetitions are permitted in arrays. Support for NumPy arrays is mainly aimed at cases where connections are to be generated from explicitly given connection data where repeated NodeIDs will be rather common.

Silmathoron · 2020-04-04T06:14:15Z

Any NodeID can occur only once in a NodeCollection, while repetitions are permitted in arrays.

Sure, on the other hand, connection to and from devices have a priori no reason to feature duplicate ids, and an error could simply be raised if the length of the converted NodeCollection differs from that of the initial list...
What I'm trying to say is that the current situation seems unnecessarily restrictive to me.
If there is no strong reason against it, I can provide a PR so that this check and conversion are done automatically (since a check is performed anyway, I can write a single function that checks and converts without degrading code readability or maintainability).

hakonsbm · 2020-04-07T14:51:35Z

Would it be possible to also include the possibility of connecting from and to numpy with devices?
This would require adapting things below L257 in hl_api_connections.py (Connect).

@Silmathoron I'm not sure this is what you're asking about, but when connecting you can use node IDs of devices as well, e.g.:

nest.Create('iaf_psc_alpha', 10)  # Node IDs 1-10
nest.Create('spike_detector')  # Node ID 11
sources = np.array([1, 5, 7], dtype=np.uint64)
targets = np.array([11, 11, 11], dtype=np.uint64)
weights = np.ones(len(sources))
delays = np.ones(len(sources))

nest.Connect(sources, targets, syn_spec={'weight': weights, 'delay': delays,
                                         'synapse_model': 'static_synapse'})

Silmathoron · 2020-04-08T05:35:47Z

@hakonsbm @heplesser
Here is a short code example; currently only the first Connect call works.
What I propose is to make all of them work, especially the simpler ones, since having to add the whole syn_spec part makes it very heavy.
Reasons for it:

convenience (if you work with graphs, your ids will not be NodeCollection but lists or arrays because this is what graph libraries use, always having to convert back and forth is tedious)
consistency (if you can use arrays in one case, it's more intuitive if you can always use them)

import nest
import numpy as np

# create neurons and devices
neuron = nest.Create("iaf_psc_alpha", 10)

vm = nest.Create("voltmeter", 10)

# set weights
weights = np.ones(len(neuron))
delays = np.ones(len(neuron))

# numpy version
arr_vm  = np.array(vm.tolist(), dtype=int)
arr_nrn = np.array(neuron.tolist(), dtype=int)

# works
nest.Connect(arr_vm, arr_nrn, syn_spec={'weight': weights, 'delay': delays, 'synapse_model': 'static_synapse'})

# does not work
nest.Connect(arr_vm, neuron.tolist(), syn_spec={'weight': weights, 'delay': delays, 'synapse_model': 'static_synapse'})
nest.Connect(vm.tolist(), neuron.tolist(), syn_spec={'weight': weights, 'delay': delays, 'synapse_model': 'static_synapse'})

nest.Connect(arr_vm, arr_nrn)
nest.Connect(arr_vm, neuron, syn_spec={'weight': weights, 'delay': delays, 'synapse_model': 'static_synapse'})
nest.Connect(arr_vm, neuron)
nest.Connect(vm.tolist(), neuron)

the way I see it, requiring syn_spec is not necessary and we could either default to the numpy connect if one of the entries is not a NodeCollection, or first check whether conversion to NodeCollection is possible (pretty harmless in my view but maybe I am missing something...) since we're already checking the type anyway (cf. previous message).

alberto-antonietti · 2020-04-15T18:15:21Z

I stressed a bit the new function :-)

If I try to connect two NumPy arrays where pre and/or post neuronshave a single element, e.g.:

nest.Create('iaf_psc_alpha', 10)
sources = np.array(1, dtype=np.uint64)
targets = np.array(2, dtype=np.uint64)
nest.Connect(sources, targets, syn_spec={'synapse_model': 'static_synapse'})

I get

----> 8 nest.Connect(sources, targets, syn_spec={'synapse_model': 'static_synapse'})

~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/ll_api.py in stack_checker_func(*args, **kwargs)
245 def stack_checker_func(*args, **kwargs):
246 if not get_debug():
--> 247 return f(*args, **kwargs)
248 else:
249 sr('count')

~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/lib/hl_api_connections.py in Connect(pre, post, conn_spec, syn_spec, return_synapsecollection)
218 # to the right formats.
219 processed_syn_spec = _process_syn_spec(
--> 220 syn_spec, processed_conn_spec, len(pre), len(post), connect_np_arrays)
221
222 # If pre and post are arrays of node IDs, and conn_spec is unspecified,

TypeError: len() of unsized object

Another similar error in the parameters provided can be the following one:

nest.Create('iaf_psc_alpha', 10)
sources = np.array([1], dtype=np.uint64)
targets = np.array([2], dtype=np.uint64)
nest.Connect(sources, targets, syn_spec={'synapse_model': 'static_synapse', 'weight': np.array(1.0)})
# or
nest.Connect(sources, targets, syn_spec={'synapse_model': 'static_synapse', 'delay': np.array(1.0)})

----> 4 nest.Connect(sources, targets, syn_spec={'synapse_model': 'static_synapse', 'weight': np.array(1.0)})

~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/ll_api.py in stack_checker_func(*args, **kwargs)
245 def stack_checker_func(*args, **kwargs):
246 if not get_debug():
--> 247 return f(*args, **kwargs)
248 else:
249 sr('count')

~/workspace/nest-simulator/b/lib/python3.6/site-packages/nest/lib/hl_api_connections.py in Connect(pre, post, conn_spec, syn_spec, return_synapsecollection)
249 syn_param_values = None
250
--> 251 connect_arrays(pre, post, weights, delays, synapse_model, syn_param_keys, syn_param_values)
252 return
253

pynestkernel.pyx in pynestkernel.NESTEngine.connect_arrays()

TypeError: weights must be a NumPy array

I added check for the first case and slightly modified the error messages for the second case.
(see hakonsbm#28)

…from_data small changes for managing 0-dimensional inputs

alberto-antonietti

Thank you for the effort!
I am now reviewing the related PR #1485

heplesser · 2020-04-20T09:37:26Z

This fixes #1423 and HBP Ticket #482997.

hakonsbm added 3 commits February 13, 2020 09:42

Replaced 'connect_nonunique' with connect_arrays

66935ea

Tests of connect_arrays

72cbcd3

Ability to choose synapse model when connecting with NumPy arrays

3652292

terhorstd added ZC: PyNEST DO NOT USE THIS LABEL I: External API Developers of extensions or other language bindings may need to adapt their code ZP: PR Created DO NOT USE THIS LABEL S: High Should be handled next T: Enhancement New functionality, model or documentation labels Feb 19, 2020

hakonsbm added 3 commits February 19, 2020 14:00

Changes to connecting with Numpy arrays

963b2ee

Made weight and delay optional, added optional receptor_type specification.

Refactored connect_arrays to directly create connections

3d8899b

Added threading support in connect_arrays

21f9c93

heplesser requested review from alberto-antonietti, jougs and sarakonradi February 20, 2020 09:04

Updated documentation on connection with arrays

9d5841e

jougs requested changes Feb 25, 2020

View reviewed changes

sarakonradi suggested changes Feb 25, 2020

View reviewed changes

hakonsbm and others added 5 commits February 27, 2020 09:13

Apply suggestions by @jougs from code review

7946025

Co-Authored-By: Jochen Martin Eppler <jougs@gmx.net>

Raise the correct error in connect_arrays

2c0e122

Corrected spelling of NumPy

50bdc12

Adjustments to documentation based on code review

1921b72

Putting comment on a single line

a871a7d

jougs approved these changes Feb 27, 2020

View reviewed changes

Silmathoron mentioned this pull request Feb 28, 2020

Remove pylab and delay pyplot import #1447

Merged

alberto-antonietti suggested changes Mar 5, 2020

View reviewed changes

heplesser assigned hakonsbm Apr 7, 2020

heplesser removed ZC: PyNEST DO NOT USE THIS LABEL ZP: PR Created DO NOT USE THIS LABEL labels Apr 7, 2020

heplesser added this to the NEST 3.0 milestone Apr 7, 2020

Merge branch 'master' into connection_generation_from_data

d08e99f

alberto-antonietti added 2 commits April 15, 2020 19:55

modified error messages when providing NumPy arrays with dimensions != 1

b917a19

added check for zero-dimensional sources and target arrays

0e7175b

alberto-antonietti mentioned this pull request Apr 15, 2020

small changes for managing 0-dimensional inputs hakonsbm/nest-simulator#28

Merged

Merge pull request #28 from alberto-antonietti/connection_generation_…

04211cd

…from_data small changes for managing 0-dimensional inputs

alberto-antonietti approved these changes Apr 16, 2020

View reviewed changes

heplesser requested a review from sarakonradi April 16, 2020 12:54

sarakonradi approved these changes Apr 20, 2020

View reviewed changes

heplesser merged commit f529b42 into nest:master Apr 20, 2020

This was referenced Apr 20, 2020

Full support for numpy arrays and lists in Connect #1523

Closed

Enhance Connect to work seamlessly with arrays and lists #1539

Merged

hakonsbm deleted the connection_generation_from_data branch April 29, 2020 11:17

hakonsbm mentioned this pull request Aug 5, 2020

Connect with NumPy arrays of node IDs does not work with MPI #1710

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add option to generate connections from NumPy arrays #1429

Add option to generate connections from NumPy arrays #1429

hakonsbm commented Feb 13, 2020 •

edited

Loading

alberto-antonietti commented Feb 14, 2020

hakonsbm commented Feb 17, 2020

alberto-antonietti commented Feb 17, 2020 •

edited

Loading

jougs commented Feb 25, 2020

jougs left a comment

sarakonradi left a comment

sarakonradi Feb 25, 2020

hakonsbm Feb 27, 2020

sarakonradi Apr 20, 2020

alberto-antonietti commented Feb 25, 2020

jougs left a comment

alberto-antonietti left a comment

alberto-antonietti Mar 5, 2020

hakonsbm Mar 6, 2020

alberto-antonietti Mar 5, 2020

hakonsbm Mar 6, 2020

alberto-antonietti Mar 5, 2020

hakonsbm Mar 6, 2020

Silmathoron commented Apr 3, 2020

heplesser commented Apr 3, 2020

Silmathoron commented Apr 4, 2020 •

edited

Loading

hakonsbm commented Apr 7, 2020

Silmathoron commented Apr 8, 2020 •

edited

Loading

alberto-antonietti commented Apr 15, 2020

alberto-antonietti left a comment

heplesser commented Apr 20, 2020

		@@ -233,9 +232,3 @@ def fixdict(d):
		projections = fixdict(projections)

Add option to generate connections from NumPy arrays #1429

Add option to generate connections from NumPy arrays #1429

Conversation

hakonsbm commented Feb 13, 2020 • edited Loading

alberto-antonietti commented Feb 14, 2020

hakonsbm commented Feb 17, 2020

alberto-antonietti commented Feb 17, 2020 • edited Loading

jougs commented Feb 25, 2020

jougs left a comment

Choose a reason for hiding this comment

sarakonradi left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alberto-antonietti commented Feb 25, 2020

jougs left a comment

Choose a reason for hiding this comment

alberto-antonietti left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Silmathoron commented Apr 3, 2020

heplesser commented Apr 3, 2020

Silmathoron commented Apr 4, 2020 • edited Loading

hakonsbm commented Apr 7, 2020

Silmathoron commented Apr 8, 2020 • edited Loading

alberto-antonietti commented Apr 15, 2020

alberto-antonietti left a comment

Choose a reason for hiding this comment

heplesser commented Apr 20, 2020

hakonsbm commented Feb 13, 2020 •

edited

Loading

alberto-antonietti commented Feb 17, 2020 •

edited

Loading

Silmathoron commented Apr 4, 2020 •

edited

Loading

Silmathoron commented Apr 8, 2020 •

edited

Loading