spaic.Learning package

Submodules

spaic.Learning.Learner module

Created on 2020/8/12 @project: SPAIC @filename: Learner @author: Hong Chaofei @contact: hongchf@gmail.com

@description: 定义学习模块，包括各种学习算法对仿真计算过程中插入的各种计算模块，以及记录需要学习连接的接口

class spaic.Learning.Learner.Learner(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: BaseModule, ABC

Base learner model for all the learner model.

Parameters:

• parameters (dict) – The parameters for learner.

• super_parameters (dict) – Super parameters for future use.

• backend_functions (dict) – Contains all the learner model we can choose.

• name (str) – The typical name for the learner.

• preferred_backend (str) – Choose which kind of backend to use. Like “Pytorch”, “Tensorflow” or “Jax”.

• trainable_groups (dict) – Trainable container, includes nodes and layers to train.

• trainable_connections (dict) – Trainable container, includes connections to train.

• init_trainable – The initial state of this learner of whether it is trainable.

add_trainable(self, trainable)

Add target object (Network, Assembly, Connection, or list of them) to the trainable container

build(self, backend)

Build Learner, choose the backend as user wish, if we have already finished the api.

learning_algorithms = {'bp_rstdp': <class 'spaic.Learning.Backprop_RSTDP.Backprop_RSTDP'>, 'bp_rstdpet': <class 'spaic.Learning.Backprop_RSTDP.Backprop_RSTDP'>, 'conv2d_rstdp': <class 'spaic.Learning.Conv_RSTDP.Conv2d_RSTDP'>, 'conv2d_stdp': <class 'spaic.Learning.Conv_STDP.Conv2d_STDP'>, 'full_online_stdp': <class 'spaic.Learning.STDP_Learner.full_online_STDP'>, 'meta_nearest_online_stdp': <class 'spaic.Learning.STDP_Learner.Meta_nearest_online_STDP'>, 'nearest_online_stdp': <class 'spaic.Learning.STDP_Learner.nearest_online_STDP'>, 'postpreintstdp': <class 'spaic.Learning.STDP_Learner.PostPreIntSTDP'>, 'rate_modulate': <class 'spaic.Learning.Rate_Modulation.Rate_Modulate'>, 'rstdp': <class 'spaic.Learning.RSTDP.RSTDP'>, 'rstdpet': <class 'spaic.Learning.RSTDP.RSTDPET'>, 'sbp': <class 'spaic.Learning.SurrogateBP_Learner.SurrogateBP'>, 'stbp': <class 'spaic.Learning.STBP_Learner.STBP'>, 'stca': <class 'spaic.Learning.STCA_Learner.STCA'>, 'switch_rstdp': <class 'spaic.Learning.RSTDP.RewardSwitchSTDP'>, 'tempotron': <class 'spaic.Learning.Tempotron_Learner.Tempotron'>}

learning_optims = {}

optim_dict = {'ASGD': <class 'torch.optim.asgd.ASGD'>, 'Adadelta': <class 'torch.optim.adadelta.Adadelta'>, 'Adagrad': <class 'torch.optim.adagrad.Adagrad'>, 'Adam': <class 'torch.optim.adam.Adam'>, 'AdamW': <class 'torch.optim.adamw.AdamW'>, 'Adamax': <class 'torch.optim.adamax.Adamax'>, 'LBFGS': <class 'torch.optim.lbfgs.LBFGS'>, 'RMSprop': <class 'torch.optim.rmsprop.RMSprop'>, 'Rpop': <class 'torch.optim.rprop.Rprop'>, 'SGD': <class 'torch.optim.sgd.SGD'>, 'SparseAdam': <class 'torch.optim.sparse_adam.SparseAdam'>}

lr_schedule_dict = {'CosineAnnealingLR': <class 'torch.optim.lr_scheduler.CosineAnnealingLR'>, 'CosineAnnealingWarmRestarts': <class 'torch.optim.lr_scheduler.CosineAnnealingWarmRestarts'>, 'CyclicLR': <class 'torch.optim.lr_scheduler.CyclicLR'>, 'ExponentialLR': <class 'torch.optim.lr_scheduler.ExponentialLR'>, 'LambdaLR': <class 'torch.optim.lr_scheduler.LambdaLR'>, 'MultiStepLR': <class 'torch.optim.lr_scheduler.MultiStepLR'>, 'ReduceLROnPlateau': <class 'torch.optim.lr_scheduler.ReduceLROnPlateau'>, 'StepLR': <class 'torch.optim.lr_scheduler.StepLR'>}

add_trainable(trainable: list)

Add target object (Assembly, Connection, or list of them) to the trainable container :param trainable: The trainable target waiting for added. :type trainable: list

add_pathway(pathway: list)

Add target object (Assembly, Connection, or list of them) to the pathway container :param pathway: The pathway target waiting for added. :type pathway: list

build(backend: Backend)

Build Learner, choose the backend as user wish, if we have already finished the api. :param backend: Backend we have. :type backend: backend

active()

deactive()

set_optimizer(optim_name, optim_lr, **kwargs)

set_schedule(lr_schedule_name, **kwargs)

get_param()

get_varname(key)

get_var_names()

build_optimizer()

build_lr_schedule()

optim_step()

optim_zero_grad()

optim_schedule()

static register(name, algorithm)

custom_rule(backend: Backend)

connection_rule(con: Connection, backend: Backend, obj_type='trainable')

neuron_rule(neuron: NeuronGroup, backend: Backend, obj_type='trainable')

is_overridden(func)

class spaic.Learning.Learner.ReSuMe(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

class spaic.Learning.Learner.FORCE(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

spaic.Learning.RSTDP module

@author: Yuan Mengwen @contact: mwyuan94@gmail.com @project: PyCharm @filename: RSTDP.py @time:2021/4/8 10:46 @description:

class spaic.Learning.RSTDP.RSTDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

Reward-modulated STDP. Adapted from (Florian 2007) <https://florian.io/papers/2007_Florian_Modulated_STDP.pdf>. :param lr: learning rate :type lr: int or float :param trainable: It can be network or neurongroups

Variables:

• tau_plus (int or float) – Time constant for pre -synaptic firing trace determines the range of interspike intervals over which synaptic occur.

• tau_minus (int or float) – Time constant for post-synaptic firing trace.

• a_plus (float) – Learning rate for post-synaptic.

• a_minus (flaot) – Learning rate for pre-synaptic.

weight_update(weight, eligibility, reward)

RSTDP learning rule for Connection subclass of AbstractConnection class. :param weight: weight between pre and post neurongroup :param eligibility: a decaying memory of the relationships between the recent pairs of pre and postsynaptic spike pairs :param reward: reward signal

build(backend)

class spaic.Learning.RSTDP.RSTDPET(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

Reward-modulated STDP with eligibility trace. Adapted from (Florian 2007) <https://florian.io/papers/2007_Florian_Modulated_STDP.pdf>. :param lr: learning rate :type lr: int or float :param trainable: It can be network or neurongroups

Variables:

• tau_plus (int or float) – Time constant for pre -synaptic firing trace determines the range of interspike intervals over which synaptic occur.

• tau_minus (int or float) – Time constant for post-synaptic firing trace.

• a_plus (float) – Learning rate for post-synaptic.

• a_minus (flaot) – Learning rate for pre-synaptic.

Notes

Batch_size for network using RSTDPET as learning algorithm must be 1.

weight_update(weight, eligibility_trace, reward)

RSTDPET learning rule for Connection subclass of AbstractConnection class. .. rubric:: Notes

The batch of pre and post should be 1

Parameters:

• weight – weight between pre and post neurongroup

• eligibility_trace – a decaying memory of the relationships between the recent pairs of pre and postsynaptic spike pairs

• reward – reward signal

build(backend)

class spaic.Learning.RSTDP.RewardSwitchSTDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

update(input, output, reward, input_trace, output_trace, pre_decay, post_decay, Apost_pos, Apost_neg, Apre_pos, Apre_neg, m_rate, homoestatic, weight)

build(backend)

spaic.Learning.Rate_Modulation module

Created on 2021/6/1 @project: SPAIC @filename: Rate_Modulation @author: Hong Chaofei @contact: hongchf@gmail.com

@description:

class spaic.Learning.Rate_Modulation.Rate_Modulate(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

neuron spiking rate modulation

build(backend: Backend)

property rate_loss

class spaic.Learning.Rate_Modulation.Modulation_method(weight, m_r, alpha, var_name)

Bases: object

record_rate(out)

modulate_step()

property rate_loss

spaic.Learning.STBP_Learner module

Created on 2021/3/30 @project: SPAIC @filename: STBP_Learner @author: Mengxiao Zhang @contact: mxzhangice@gmail.com

@description:

class spaic.Learning.STBP_Learner.STBP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

STBP learning rule.

Parameters:

• alpha (num) – The parameter alpha of STBP learning model.

• trainable – The parameter whether it can be trained.

build(self, backend)

Build the backend, realize the algorithm of STBP model.

threshold(self, x, v_th)

Get the threshold of the STBP model.

Example

Net._learner = STBP(0.5, Net)

Reference:

Yujie Wu et al. “Spatio-Temporal Backpropagation for Training High-performance Spiking Neural Networks” In: Frontiers in Neuroscience, 2018. Volume 12. pp. 331. doi:10.3389/fnins.2018.00331 url:ttps://www.frontiersin.org/article/10.3389/fnins.2018.00331

build(backend)

Build the backend, realize the algorithm of STBP model. Args：

backend: The backend we used to compute.

threshold(x, v_th)

Get the threshold of the STBP model.

Returns:

A method that use STBP model to compute the threshold.

spaic.Learning.STCA_Learner module

Created on 2020/11/9 @project: SPAIC @filename: STCA_Learner @author: Hong Chaofei @contact: hongchf@gmail.com

@description:

class spaic.Learning.STCA_Learner.ActFun(*args, **kwargs)

Bases: Function

Approximate firing func.

static forward(ctx, input, thresh, alpha)

static backward(ctx, grad_output)

spaic.Learning.STCA_Learner.firing_func(x, v_th, alpha)

class spaic.Learning.STCA_Learner.STCA(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

STCA learning rule.

Parameters:

• alpha (num) – The parameter alpha of STCA learning model.

• preferred_backend (list) – The backend prefer to use, should be a list.

• name (str) – The name of this learning model. Should be ‘STCA’.

• firing_func – The function of fire.

build(self, backend)

Build the backend, realize the algorithm of STCA model.

threshold(self, x, v_th)

Get the threshold of the STCA model.

Example

Net._learner = STCA(0.5, Net)

Reference:

Pengjie Gu et al. “STCA: Spatio-Temporal Credit Assignment with Delayed Feedback in Deep SpikingNeural Networks.” In:Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19. International Joint Conferences on Artificial Intelligence Organization, July 2019,pp. 1366–1372. doi:10.24963/ijcai.2019/189. url:https://doi.org/10.24963/ijcai.2019/189.

build(backend)

Build the backend, realize the algorithm of STCA model.

Args：

backend: The backend we used to compute.

torch_threshold(x, v_th)

Get the threshold of the STCA model.

Returns:

A method that use STCA model to compute the threshold.

spaic.Learning.STDP_Learner module

class spaic.Learning.STDP_Learner.Base_STDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Learner

build(backend)

class spaic.Learning.STDP_Learner.nearest_online_STDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Base_STDP

nearest_online STDP learning rule.

Parameters:

• Apost (num) – The parameter Apost of nearest_online STDP learning model.

• Apre (num) – The parameter Apre of nearest_online STDP learning model.

• trace_decay (num) – The parameter trace_decay of nearest_online STDP learning model.

• preferred_backend (list) – The backend prefer to use, should be a list.

• name (str) – The name of this learning model. Should be ‘nearest_online_STDP’.

initial_param(self, input, output)

initialize the output_trace and the input_trace for each batch.

nearest_online_stdp_weightupdate(self, input, output, weight)

calculate the update of weight

build(self, backend)

Build the backend, realize the algorithm of nearest_online STDP learning model.

Example

self._learner = BaseLearner(algorithm=’nearest_online_STDP’, lr=0.5, trainable=self, conn=self.connection1)

Reference:

Unsupervised learning of digit recognition using spike-timing-dependent plasticity. doi: 10.3389/fncom.2015.00099. url: http://journal.frontiersin.org/article/10.3389/fncom.2015.00099/abstract

nearest_online_stdp_weightupdate(dw, weight)

Parameters:

• dw – the change of weight

• weight – weight between pre and post neurongroup

Returns:

Updated weight.

build(backend)

class spaic.Learning.STDP_Learner.full_online_STDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Base_STDP

nearest_online STDP learning rule.

Parameters:

• Apost (num) – The parameter Apost of full_online STDP learning model.

• Apre (num) – The parameter Apre of full_online STDP learning model.

• trace_decay (num) – The parameter trace_decay of full_online STDP learning model.

• preferred_backend (list) – The backend prefer to use, should be a list.

• name (str) – The name of this learning model. Should be ‘full_online_STDP’.

initial_param(self, input, output)

initialize the output_trace and the input_trace for each batch.

nearest_online_stdp_weightupdate(self, input, output, weight)

calculate the update of weight

build(self, backend)

Build the backend, realize the algorithm of full_online STDP learning model.

Example

self._learner = BaseLearner(algorithm=’full_online_STDP’, lr=0.5, trainable=self, conn=self.connection1)

Reference:

Unsupervised learning of digit recognition using spike-timing-dependent plasticity. doi: 10.3389/fncom.2015.00099. url: http://journal.frontiersin.org/article/10.3389/fncom.2015.00099/abstract

full_online_stdp_weightupdate(dw, weight)

Parameters:

• dw – the change of weight

• weight – weight between pre and post neurongroup

Returns:

Updated weight.

build(backend)

class spaic.Learning.STDP_Learner.Meta_nearest_online_STDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Base_STDP

update(input, output, input_trace, output_trace, pre_decay, post_decay, Apost, Apre, weight)

build(backend)

class spaic.Learning.STDP_Learner.PostPreIntSTDP(trainable=None, pathway=None, algorithm=('STCA', 'STBP', 'RSTDP', '...'), **kwargs)

Bases: Base_STDP

Simple STDP rule involving both pre- and post-synaptic spiking activity, based on integer arithmetic. By default, pre-synaptic update is negative and the post-synaptic update is positive.

update(input, output, input_trace, output_trace, trace_decay, trace_scale, shift_spike_trace, max_threshold, post_thresh, weight)

build(backend)

spaic.Learning.SpikeProp_Learner module

spaic.Learning.TRUE_Learner module

Module contents

Created on 2020/8/11 @project: SPAIC @author: Hong Chaofei @contact: hongchf@gmail.com

@description: