nepes NM500 User Guide

NM500 Software Development Guide

Python Wrapper API

Version 2.0

Contents

1. Introduction

2. Set up Development Environment.

2.1 Overview

2.2 Set up development environment for Linux Series

2.3 Set up development environment for macOS

2.3 Set up development environment for Windows

2.4 Using NM500 library APIs

3. NM500 APIs

3.1 set_lib_path

3.2 get_devices

3.3 connect

3.4 close

3.5 get_network_info

3.6 get_version

3.7 get_neuron_count

3.8 forget

3.9 reset

3.10 set_network_type

3.11 get_network_type

3.12 get_network_status

3.13 set_context

3.14 get_context

3.15 learn

3.16 learn_batch

3.17 clusterize

3.18 classify

3.19 read

3.20 write

3.21 read_neuron

3.22 read_neurons

3.23 write_neurons

3.24 get_model_info

3.25 get_model_stat

3.26 save_model

3.27 load_model

3.28 power_save

4. Result Codes

5. NeuroMem Registers

History

Date

Version

Description

2019-04-09

2.0

1. Introduction

This document is intended to use with nepes NM500 Software Development Kit. The nepes NM500

Software Development Kit allows you to manage nepes NM500 neuromorphic chip and implement

custom solutions by using libraries, utilities, and programming interfaces provided with the nepes

NM500 Software Development Kit.

The nepes NM500 performs pattern recognition upon a knowledge model contructed with

NeuroMem networks.

This document describes in detail how you can develop applications using the nepes NM500

Software Development Kit.

2. Set up Development Environment.

2.1 Overview

This nepes NM500 Software Development Kit contains NeuroMemEngine and USB(or SPI for

Raspberry PI) libraries for NM500 application boards. A developer can use this SDK to create

application that access NM500 neural network, build a knowledge model, and make decisions based

on pattern recognition.

2.2 Set up development environment for Linux Series

Pre-requisites:

1. libusb-1.0.x is required for functioning of the API library.

$ sudo apt-get install libusb-1.0-0-dev

Installation Steps:

1. Copy library into /usr/local/lib

[x64]

$ sudo cp lib/linux.x64/libnmengine.so.2.0.0 /usr/local/lib

[x86]

$ sudo cp lib/linux.x86/libnmengine.so.2.0.0 /usr/local/lib

[armv7I/usb]

$ sudo cp lib/linux.armv7I.usb/libnmengine.so.2.0.0 /usr/local/lib

[armv7I/spi] (e.g Raspberry PI with interface board support spi),

$ sudo cp lib/linux.armv7I.spi/libnmengine.so.2.0.0 /usr/local/lib

2. Make a link for library

$ sudo ln -s /usr/local/lib/libnmengine.so.2.0.0 /usr/local/lib/libnmengine.so

3. Add /usr/local/lib to LD_LIBRARY_PATH

e.g. export LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH:/usr/local/lib:

or set the path for library by using set_lib_path() func

2.3 Set up development environment for macOS

Pre-requisites:

1. libusb-1.0.x is required for functioning of the API library.

$ brew install libusb

Installation Steps:

1. Copy library into /usr/local/lib

$ sudo cp lib/macos/libnmengine.dylib.2.0.0 /usr/local/lib

2. Make a link for library

$ sudo ln -s /usr/local/lib/libnmengine.dylib.2.0.0 /usr/local/lib/libnmengine.dylib

3. Add /usr/local/lib to DYLD_LIBRARY_PATH

e.g. export DYLD_LIBRARY_PATH=.:$DYLD_LIBRARY_PATH:/usr/local/lib:

or set the path for library by using set_lib_path() func

2.3 Set up development environment for Windows

1. Unpack CypressDriverInstaller.zip in the Driver folder inside the package

and run the CypressDriverInstaller.exe file to install the driver.

2. Add the dll location to the system environment variable (like PATH)

lib/win.x32 or lib/win.x64

or place the dlls in the same location where the executable is located.

or set the path for library by using set_lib_path() func

2.4 Using NM500 library APIs

To use the NM500 library APIs, simply 'import nmengine'

3. NM500 APIs

3.1 set_lib_path

The set_lib_path function is used to set the local library path directly.

def set_lib_path(lib_path)

Args

 lib_path (str): A string of library path

Returns

 none (void)

The following is an example of setting the library path

import nmengine

# Set the library path directly if library can not load

nmengine.set_lib_path('<your library path>')

3.2 get_devices

The get_devices function is used to get the list of the NM500 devices detected.

def get_devices()

Args

 none

Returns

 result (int): An integer of result code. see Section 4. Result Codes.  devices ([Device]): Up to 5 detected device information.

3.3 connect

The connect function is used to open selected device and perform the initialization of the NM500.

It will return ERROR_INIT_FAILED (104) if the NM500 is not properly initialized or selected device does

not support the NM500 functionality.

For more information on result codes, see Section 4. Result Codes.

def connect(device)

Args

 device (Device): A device selected for use.

Returns

 result (int): An integer of result code.

3.4 close

The close function is used to close a selected device handle. This should be called before

application exits.

def close(device)

Args

 device (Device): A device selected for disconnect.

Returns

 result (int): An integer of result code.

3.5 get_network_info

The get_network_info function is used to get information of the NM500 neural network(NeuroMem)

and device.

def get_network_info(device, info)

Args

 device (Device): A device selected for use.  info (NetworkInfo): A reference of NetworkInfo

Returns

 result (int): An integer of result code.

The NetworkInfo information is as follows;

.

 neuron_memory_size: the memory size of the neuron - fixed at 256  neuron_size: the total number of neurons on the NeuroMem network  version: the version of the device firmware.

The version value is represented as follows;

 [15:8] Board Identifier  [7:4] Board Version  [3:0] FPGA Version

[Warning] This function performs the initialization of the NM500 to count the total number of

neurons. Therefore, do not use this function during learning or recogition. This function is only useful

if you want to check the overall network information at the beginning.

3.6 get_version

The get_version function is used to get the version of the device firmware.

def get_version(device)

Args

 device (Device): A device selected for use.

Returns

 result (int): An integer of result code.  version (int): An integer of device firmware version.

The version value is represented as follows;

 [15:8] Board Identifier  [7:4] Board Version  [3:0] FPGA Version

3.7 get_neuron_count

The get_neuron_count function is used to get the number of neurons used on the NeuroMem

network.

def get_neuron_count(device)

Args

 device (Device): A device selected for use.

Returns

 result (int): An integer of result code.  count (int): An integer of the number of neuron used.

The count value is the number of neurons actually used for learning, not the total number of

neurons.

3.8 forget

The forget function is used to clear all knowledge on the NeuroMem network.

def forget(device)

Args

 device (Device): A device selected for use.

Returns

 result (int): An integer of result code.

It will only initialize the daisy-chain(index of neuron) on the NeuroMem network. In other words, It

will not erase the memory data(stored weight memory) of the neuron. Therefore, the neuron memory

data retrieved using readNeuron(s) function can contain the garbage data if you do not use

maximum memory size of neuron as input vector size. (the maximum memory size of neuron is 256).

However, it is faster than Reset function.

3.9 reset

The reset function is used to clear all knowledge on the NeuroMem network.

def reset(device)

Args

 device (Device): A device selected for use.

Returns

 result (int): An integer of result code.

It performs the initialization of all the NeuroMem network setting and stored weight momory data of

neurons.

3.10 set_network_type

The set_network_type function is used to set the NeuroMem network mode.

def set_network_type(device, nwtype)

Args

 device (Device): A device selected for use.  nwtype (int): An Integer of network type

The NeuroMem supports 2 types of network mode: RBF and KNN network. The RBF network

must be selected to build a knowledge model.

 0: RBF network mode  1: KNN network mode

Returns

 result (int): An integer of result code.

3.11 get_network_type

The get_network_type function is used to check the current network mode. To change network mode

is possible at any time during recognition(classification)

def get_network_type(device)

Args

 device (Device): A device selected for use.

Returns

 result (int): An integer value of result code.  type (int): An integer value of the current network type.

The values for the type are:

 0: RBF network mode,  1: KNN network mode

To change network type(mode) is possible at any time during recognition(classification)

3.12 get_network_status

The get_network_status function is used to get the current status of the NeuroMem network.

def get_network_status(device, status)

Args

 device (Device): A device selected for use.  status (NetworkStatus): A reference of NetworkStatus

Returns

 result (int): An integer value of result code.

The information of the NetworkStatus is as follows;

 network_type: the type of network mode: (0: RBF network mode, 1: KNN network mode)  network_used: the number of neurons currently used  context: the context ID that is currently used. It is same as global context ID  norm: the type of norm (L1 or Lsup) that is currently used.

3.13 set_context

The set_context function is used to configure the current(global) context attributes on the NeuroMem

network.

def set_context(device, context)

Args

 device (Device): A device selected for use.  context (Context): A reference of Context

Returns

 result (int): An integer value of result code.

The Context information is as follows;

 context: the context ID that is used to split the NeuroMem network into subnetworks.

The neurons can be associated to different contexts and their use can be enabled or

disabled by selecting a context value

 norm: the type of norm (L1 or Lsup).

 minif(Minimum Influence Field): It represents minimum differences for dissimilarity

judgement and it is used to control uncertain domain.

 maxif(Maximum Influence Field): it represents maximum differences for similarity

judgement and it is used to adjust conservatism. The values range from 1 to 65535.

3.14 get_context

The get_context function is used to get the current context on the NeuroMem network

def get_context(device, context)

Args

 device (Device): A device selected for use.  context (Context): A reference of Context

For more information on Context, see Section 3.13 set_context

Returns

 result (int): An integer value of result code.

3.15 learn

The learn function is used to teach a neuron(s) with given vector on the NeuroMem network.

def learn(device, req)

Args

 device (Device): A device selected for use.  req (LearnReq): A reference of LearnReq

Returns

 result (int): An integer value of result code.

When a neuron learns input vector, its context information is set to the value of the current(global)

context. If neurons have been assigned with the same context, it says that they will be run with same

conditions, such as MinIF, MaxIF and Norm method.

When a new neuron is assigned to learn input vector, the NCOUNT register value of the NM500 is

incremented by one. If recognizable by existing neurons, new neurons are not allocated. Therefore,

the NCOUNT register value of the NM500 will not be changed.

The LearnReq information is as follow.

 Inputs

 query_affected: flag for whether to retrieve affected neuron information.  generally flagging is not required.  category: category of input data(vector).  vector_size: input data size.  vector: input data.

 Outputs

 status: network status of learning  affected_count: number of affected neurons  affected_neurons[affected_count]: list of affected neurons

The value of status represents the result status of the learning, and it is as follows.

 LEARN_ALREADY_KNOWN(0)  LEARN_SUCCESS(1)  LEARN_ADJUSTED (2)  LEARN_DEGENERATED (3)

The following is an example of learning and retrieves affected neurons after learning. If the

queryAffected is set to 0, the information of affected neurons will not be returned. If the

queryAffected is set to 1, it will return list of affected neurons. but the processing time will be

increased by more than 2 times.

#

# learn vector (10, 10, 10) with category 10

#

learn_req = nmengine.LearnReq()

vector = [10 for i in range(3)]

# Send max length to avoid gabages in neuron memory.

# If you set the size with actual vector length (< NEURON_MEMORY),

# the remaining value will be filled with garbege value

learn_req.size = nmengine.NEURON_MEMORY

learn_req.vector[:len(vector)] = vector

learn_req.category = 10

# This is optional field for retrieving affected neuron infromation

# If you set the value to 1, it takes more time to learn.

learn_req.query_affected = 1

print('learn req(cat, affected flag, len, vector)', learn_req.category, \

learn_req.query_affected, learn_req.size, tuple(learn_req.vector))

# Fro the learning result,

# refer to 'The result code for learning' in constants.py

print('train a vector to network, result:', nmengine.learn(target, learn_req), \

'status:', learn_req.status)

# Prints affected neurons

for i in range(learn_req.affected_count):

neuron = learn_req.affected_neurons[i]

print('A (nid, cat, aif) :', neuron.nid, neuron.cat, neuron.aif)

result, count = nmengine.get_neuron_count(target)

print('the number of neurons committed is', count)

3.16 learn_batch

The learn_batch function is used to teach a neurons with given vectors on the NeuroMem network.

def learn_batch(device, req)

Args

 device (Device): A device selected for use.  req (LearnBatchReq): A reference of LearnBatchReq

Returns

 result (int): An integer value of result code.

The LearnBatchReq information for batch learning is as follow.

 Inputs

 iterable: flag for whether to iterate batch learning.  iter_count: number of iteration (epoch).  vector_count: number of vectors.  vector_size: input data size.  vectors: list of input data.  categories: list of category of input data(vector).

 Outputs

 iter_result: the number of committed neurons from each iteration.

If iterable is set as 1(true) and iter_count is greater than 0, the learning will be repeated on the basis

of the given vectors until learning is no longer possible or until the maximum iter_count value is

reached.

When learning is complete, iter_count returns the number of actually repeated iterations, and

iter_result returns the number of learned neurons for each iteration.

The following is an example of batch learning

#

# Learn vectors in batch mode

#

# Maximum iteration for learning

iter_count = 10

# The number of vectors for learning (for testing)

vector_count = 20

# The vector size

vector_size = 3

# Get the LearnBatchReq instance from factory method

req = nmengine.learn_batch_req_factory(iter_count, vector_count, vector_size)

# Flag for iteration,

# 1: automatic iterative learning

# 0: none

req.iterable = 1

#

# Generates dummy training dataset

#

# Gets the random value for cat

cat = randint(1, 50)

for i in range(vector_count):

# Keeps the same cat value for unknown case

if (i % 3) == 0:

cat = randint(1, 50)

# Gets the random value for vector

vec = randint(1, 50)

for j in range(vector_size):

req.vectors[i * vector_size + j] = vec

req.categories[i] = cat

print(i, 'batch >> cat:', cat, 'vec:', [vec for i in range(vector_size)])

print()

# Learn dataset in batch mode

result = nmengine.learn_batch(target, req)

print('batch learning done, iter:', req.iter_count, 'result:', result)

for i in range(req.iter_count):

print('epoch:', i, 'learned', req.iter_result[i])

# Print statistics of custerized dataset

# Get meta information of the model

info = nmengine.ModelInfo()

nmengine.get_model_info(target, info)

print("model count:", info.count, 'max ctx:', info.max_context \

, 'max cat:', info.max_category)

stat = nmengine.model_stat_factory(info.max_category)

# Get statistics of each context

stat.context = 1

nmengine.get_model_stat(target, stat)

dash = '-' * 30

for i in range(info.max_category + 1):

if i == 0:

print(dash)

print('{:<10}{:<10}{:<10}'.format('cat', 'neurons', 'degen'))

print(dash)

else:

print('{:<10}{:<10}{:<10}'.format(i, stat.histo_cat[i],stat.histo_deg[i]))

3.17 clusterize

The clusterize function is used to group a set of vectors given into a clusters on the NeuroMem

network.

def clusterize(device, req)

Args

 device (Device): A device selected for use.  req (ClusterizeReq): A reference of ClusterizeReq

Returns

 result (int): An integer value of result code.

The ClusterizeReq information for clusterizing is as follow.

 initial_category: initial category id (it must be greater than 0).  incrementof: unit of increasement of category id.  vector_count: number of vectors.  vector_size: input data size.  vectors: list of input data.

The minif and maxif will affect grouping of a set of vectors given. Therefore, both of them should be

set to the appropriate value before learning.

The following is an example of clusterizing

#

# Clusterize vectors on neuromem network

# It is unsupervised leanring method.

#

# The number of vectors for clusterization (for testing)

vector_count = 20

# The vector size

vector_size = 3

# Get the ClusterizeReq instance from factory method

req = nmengine.clusterize_req_factory(vector_count, vector_size)

# Initial category

req.initial_category = 1

# Flag for category value

# 0: none, use initial category for learning

# N: increase of N value of category when learning next vector

# if set 1, category will be 1, 2, 3 ...

req.incrementof = 1

#

# Generates dummy training dataset

#

for i in range(vector_count):

# Gets the random value for vector

vec = randint(1, 50)

for j in range(vector_size):

req.vectors[i * vector_size + j] = vec

print(i, 'data >>', [vec for i in range(vector_size)])

print()

# Need to set appropriate values for minimum/maximum influence field.

ctx = nmengine.Context()

ctx.context = 1

ctx.norm = 0

ctx.minif = 2

# 30 is just temporary value for testing

ctx.maxif = 30

nmengine.set_context(target, ctx)

#

# Clusterize the dataset

#

result = nmengine.clusterize(target, req)

print('clusterization done, result:', result, '\n')

# Print statistics of custerized dataset

# Get meta information of the model

info = nmengine.ModelInfo()

nmengine.get_model_info(target, info)

print("model count:", info.count, 'max ctx:', info.max_context \

, 'max cat:', info.max_category)

stat = nmengine.model_stat_factory(info.max_category)

# Get statistics of each context

stat.context = 1

nmengine.get_model_stat(target, stat)

dash = '-' * 30

for i in range(info.max_category + 1):

if i == 0:

print(dash)

print('{:<10}{:<10}{:<10}'.format('cat', 'neurons', 'degen'))

print(dash)

else:

print('{:<10}{:<10}{:<10}'.format(i, stat.histo_cat[i],stat.histo_deg[i]))

#

# Read neuron(s) information of knowledge model

#

result, count = nmengine.get_neuron_count(target)

print('the number of neurons committed', count, '\n')

neurons = (nmengine.Neuron * count)()

result, count = nmengine.read_neurons(target, neurons, count)

print(count, "neuron(s) read\n")

for i in range(count):

print(i, 'neuron id', neurons[i].nid)

print(i, 'neuron size', neurons[i].size)

print(i, 'neuron ncr', neurons[i].ncr)

print(i, 'neuron aif', neurons[i].aif)

print(i, 'neuron minif', neurons[i].minif)

print(i, 'neuron cat', neurons[i].cat)

print(i, 'neuron model', tuple(neurons[i].model))

print()

3.18 classify

The classify function is used to identify to which of a set of categories input vector belongs.

def classify(device, req)

Args

 device (Device): A device selected for use.  req (ClassifyReq): A reference of ClassifyReq

Returns

 result (int): An integer value of result code.

The ClassifyReq information is as follows.

 Inputs

 vector_size: input data size.  vector: input data.  k: number of returns matched.

 Outputs

 status: network status of classifying.  matched_count: number(n) of matched.  nid[n]: id of neuron matched.  category[n]: category of neuron matched.  degenerated[n]: degenerated flag of neuron matched.

1: degenerated

 distance[n]: distance between input data and prototype of neuron matched.

The k parameter means the number of neurons returned. and the default value is 9. If network mode

is RBF, it can be less than 9 (k).

The value of status represents the result status of the classification(recognition), and it is as follows.

 CLASSIFY_UNKNOWN (0): It cannot be classified into the learned model.  CLASSIFY_UNCERTAIN (4): It can be classified into the learned model, but it is judged to be

one or more other categories and is unclear, and additional judgment is necessary.

 CLASSIFY_IDENTIFIED (8): It can be clearly classified into the learned model.

The following is an example of classifying

#

# Classify a vector using trained model

#

vector = [20 for i in range(3)]

classify_req = nmengine.ClassifyReq()

classify_req.size = nmengine.NEURON_MEMORY

classify_req.vector[:len(vector)] = vector

# Set the value for maximum

classify_req.k = 3

print('classify_req(k, len, vector)', classify_req.k, \

classify_req.size, tuple(classify_req.vector))

print('classify a vector using trained model. result:', \

nmengine.classify(target, classify_req))

# Fro the learning result,

# refer to 'The result code for classifying' in constants.py

print('status:', classify_req.status, 'matched:', classify_req.matched_count)

# Prints matched neurons

for i in range(classify_req.matched_count):

print('M (nid, dis, cat, degen) :', classify_req.nid[i], classify_req.distance[i], \

classify_req.category[i], classify_req.degenerated[i] )

print()

3.19 read

The read function is used to read the value of the NeuroMem register specified.

def read(device, reg)

Args

 device (Device): A device selected for use.  reg (int): An Integer value of NeuroMem Register ID

Returns

 result (int): An integer value of result code.  value (int): An integer value of the value of specified NeuroMem Register

For the NeuroMem network registers accessible, refer to 5. NeuroMem Register.

The power_save function must be called after using this function directly.

3.20 write

The write function is used to set the value of the NeuroMem network register specified directly.

def write(device, reg, data)

Args

 device (Device): A device selected for use.  reg (int): An Integer value of NeuroMem register ID  data (int) An Integer value to set in the specified register (reg)

Returns

 result (int): An integer value of result code.

For the NeuroMem network registers accessible, refer to 5. NeuroMem Register.

The power_save function must be called after using this function directly.

3.21 read_neuron

The read_neuron function is used to get neuron information specified.

def read_neurons(device, neuron)

Args

 device (Device): A device selected for use.  neuron (Neuron): A reference of Neuron

Returns

 result (int): An integer value of result code.  count (int): An integer value of the number of neurons actually read. (1 or 0)

The Neuron information of each neuron is as the following.

 nid: neuron id  size: vector length used in neuron memory. (vector size)  ncr: neuron context. (context id, norm)  aif: active influence field. (threshold of activation function)  minif: minimum influence field.  model: prototype (stored weight memory/rbf center).

If the length of the learned data is less than 256, you can specify the length of data to read through

the size variable.

The following is an example of getting a neuron information.

neuron = nmengine.Neuron()

# Specifies the ID of the neuron to read.

neuron.nid = nid

result, count = nmengine.read_neuron(target, neuron)

print("the number of neurons read",count)

if count > 0:

print('neuron id', neuron.nid)

print('neuron size', neuron.size)

print('neuron ncr', neuron.ncr)

print('neuron aif', neuron.aif)

print('neuron minif', neuron.minif)

print('neuron cat', neuron.cat)

print('neuron model', tuple(neuron.model))

print()

3.22 read_neurons

The read_neurons function returns information about the knowledge model that is a list of all the

neurons learned.

def read_neurons(device, neurons, count)

Args

 device (Device): A device selected for use.  neurons ([Neuron]): A reference of list of Neuron  count (int) An Integer value of the number of neurons to read

Returns

 result (int): An integer value of result code.  count (int): An integer value of the number of neurons actually read.

The following is an example of querying up to the number of committed neurons.

result, count = nmengine.get_neuron_count(target)

print('the number of neurons used', count)

neurons = (nmengine.Neuron * count)()

result, count = nmengine.read_neurons(target, neurons, count)

print("read",count)

for i in range(count):

print(i, 'neuron id', neurons[i].nid)

print(i, 'neuron size', neurons[i].size)

print(i, 'neuron ncr', neurons[i].ncr)

print(i, 'neuron aif', neurons[i].aif)

print(i, 'neuron minif', neurons[i].minif)

print(i, 'neuron cat', neurons[i].cat)

print(i, 'neuron model', tuple(neurons[i].model))

print()

3.23 write_neurons

The write_neurons function is used to upload the knowledge model to the NeuroMem network.

def write_neurons(device, neurons, count)

Args

 device (Device): A device selected for use.  neurons ([Neuron]): A reference of list of Neuron  count (int) An Integer value of the number of neurons to write

Returns

 result (int): An integer value of result code.  count (int): An integer value of the number of neurons actually wrote.

The following is an example of uploading dummy neurons.

neurons = []

for i in range(iter_cnt):

vector = [i+1 for v in range(v_size)]

neuron = nmengine.Neuron()

neuron.nid = i

neuron.size = v_size

neuron.ncr = 1

neuron.aif = (i+1)*100

neuron.minif = 2

neuron.cat = i+1

neuron.model[:v_size] = vector

neurons.append(neuron)

result, count = nmengine.write_neurons(target, neurons, len(neurons))

print('write neurons', count, 'result:',result)

3.24 get_model_info

The get_model_info function is used to get the meta information of the knowledge model

def get_model_info(device, info)

Args

 device (Device): A device selected for use.  info (ModelInfo): A reference of ModelInfo

Returns

 result (int): An integer value of result code.

The ModelInfo information of model is as the following.

 count: number of neurons used/committed  max_context: the largest context id (1~127)

 max_category: the largest cateogry id (1~32766)

The following is an example of getting meta information of the knowledge model

# Print statistics of custerized dataset

# Get meta information of the model

info = nmengine.ModelInfo()

nmengine.get_model_info(target, info)

print("model count:", info.count, 'max ctx:', info.max_context \

, 'max cat:', info.max_category)

3.25 get_model_stat

The get_model_stat function is used to get the statistics information of the knowledge model. It

shows distribution of neuron per category

def get_model_stat(device, stat)

Args

 device (Device): A device selected for use.  stat (ModelStat): A reference of ModelStat

Returns

 result (int): An integer value of result code.

The ModelStat information of model statistics is as the following.

 Inputs

 context: target context id for analysis

 Outputs

 count: number of neurons used/committed in given context  histo_cat[the largest cateogry id + 1]: number of neurons per cateogory  histo_deg[the largest cateogry id + 1]: nember of degenerated neuron per category

The following is an example of getting the statistics information of the knowledge model.

stat = nmengine.model_stat_factory(info.max_category)

# Get statistics of each context

stat.context = 1

nmengine.get_model_stat(target, stat)

dash = '-' * 30

for i in range(info.max_category + 1):

if i == 0:

print(dash)

print('{:<10}{:<10}{:<10}'.format('cat', 'neurons', 'degen'))

print(dash)

else:

print('{:<10}{:<10}{:<10}'.format(i, stat.histo_cat[i],stat.histo_deg[i]))

3.26 save_model

The save_model function is used to save the current knowledge model to file.

def save_model(device, neurons, count, path)

Args

 device (Device): A device selected for use.  neurons ([Neuron]): A reference of list of Neurons  count (int): A integer value of the number of neurons to save  path (str): A string of file path to save

Returns

 result (int): An integer value of result code.

The following is an example of saving the model to file:

result, count = nmengine.get_neuron_count(target)

print('the number of neurons used', count)

neurons = (nmengine.Neuron * count)()

result, count = nmengine.read_neurons(target, neurons, count)

print('save model', count, path, 'result:', nmengine.save_model(target, neurons, count, path))

3.27 load_model

The load_model function is used to load a knowledge model from file.

def load_model(device, neurons, count, path)

Args

 device (Device): A device selected for use.  neurons ([Neuron]): A reference of list of Neurons  count (int): A integer value of the number of neurons to load (maximum)  path (str): A string of file path to load

Returns

 result (int): An integer value of result code.  count (int): A integer value of the number of neurons actually loaded.

The following is an example of loading the model from file:

# Gets network information

# Notice, this function performs the initialization of the NM500

# Do not use this function during learning/classfying

network_info = nmengine.NetworkInfo()

nmengine.get_network_info(target, network_info)

print('get network info', network_info.neuron_count, network_info.neuron_memory_size,

network_info.version)

count = network_info.neuron_count

neurons = (nmengine.Neuron * count)()

result, count = nmengine.load_model(target, neurons, count, path)

for i in range(count):

print(i, 'neuron id', neurons[i].nid)

print(i, 'neuron size', neurons[i].size)

print(i, 'neuron ncr', neurons[i].ncr)

print(i, 'neuron aif', neurons[i].aif)

print(i, 'neuron minif', neurons[i].minif)

print(i, 'neuron cat', neurons[i].cat)

print(i, 'neuron model', tuple(neurons[i].model))

print()

3.28 power_save

The power_save function is used to set the power mode of the NM500. It immediately changes to the

power saving mode.

def power_save(device)

[Notice] It is recommended that you call power_save function during idle periods after using low-

level functions such as the read() and write() functions.

4. Result Codes

All result codes are integers. Symbolic names for all result codes are defined in nmengine.h header

file.

Result Code

Name

Description

0

SUCCESS 100

ERROR_UNKNOWN

101

ERROR_DEVICE_NOT_FOUND,

,

102

ERROR_DEVICE_INFO_FETCH_FAILED,

,

103

ERROR_DEVICE_OPEN_FAILED,

104

ERROR_INIT_FAILED,

105

ERROR_INVALID_PARAMETER,

106

ERROR_NOT_SUPPORT,

107

ERROR_IO_TIMEOUT,

108

ERROR_IO_FAILED,

109

ERROR_RESOURCE

5. NeuroMem Registers

Acronym

Description

Address

Normal

mode

SR

mode

16-bit

default

NSR

Network Status Register

0x0D

RW W 0x0000

GCR

Global Control Register

0x0B

RW 0x0001

MINIF

Minimum Influence Field

0x06

RW

0x0002

MAXIF

Maximum Influence Field

0x07

RW 0x4000

NCR

Neuron Context Register

0x00

RW

0x0001

COMP

Component

0x01

W

RW

0x0000

LCOMP

Last Component

0x02

W 0x0000

INDEXCOMP

Component index

0x03

W W 0x0000

DIST

Distance register

0x03

R R 0xFFFF

CAT

Category register

0x04

RW

0xFFFF

AIF

Active Influence Field

0x05

RW

0x4000

NID

Neuron Identifier

0x0A

R R 0x0000

POWERSAVE

PowerSave

0x0E

W n/a

FORGET

Forget

0x0F

W n/a

NCOUNT

Count of committed

neurons

0x0F

R R 0x0000

RESETCHAIN

Points to the first neuron

0x0C

W n/a

TESTCOMP

Test Component

0x08

W 0x0000

TESTCAT

Test Category

0x09

W 0x0000

nepes NM500 User Guide

Specifications and Main Features

Frequently Asked Questions

User Manual