NeuralNetworks
NeuralNetworks copied to clipboard
A resource-conscious neural network implementation for MCUs
Simple MLP - NeuralNetwork Library For Microcontrollers
Nothing "Import ant", just a simple library for implementing Neural-Networks(NNs) easily and effectively on any Arduino board and other microcontrollers.
Summary
NN Functions | Input Type (x) | Output Type (Y) | Action |
---|---|---|---|
getMeanSqrdError(x) |
Unsigned Int | float | Calculates & Returns the MSE (Mean Squared Error) = [ SSE (Sum Squared Error) divided by the Product of number-οf-οutputs and inputs-per-epoch aka batch-size]. |
*FeedForward(x) |
1D Float Array | 1D Float Array | "Feeds" the NN with X-input values and returns Y-Output Values, If needed. |
BackProp(x) |
1D Float Array | - | "Tells" to the NN if the outputs-Y were correct/the-expected/X-inputs and then, "teaches" it. |
Examples: Backpropagation_Single_Xor_Gate.ino
Understanding the Basics of a Neural Network:
0
1
2
3
4
5
6
7
8
9
10
11
12
.
Examples
* | Example Files (.ino) | About |
---|---|---|
1 | Backpropagation_Single_Xor_Gate |
Training a NN to behave as a xor-gate |
2 | Backpropagation_double_Xor |
Training a NN to behave as a 3-input-xor circuit |
3 | FeedForward_double_Xor |
Pre-trained NN behaving as a 3-input-xor circuit |
4 | Run_In_Loop |
Example of using a NN inside the loop |
5 | FeedForward_double_Xor_PROGMEM |
Weights & biases are stored in the programmable memmory |
6 | FeedForward_Individual_MNIST_PROGMEM |
Recognizing handwritten digits [proof of concept] |
7 | Any_Activation_Function_Per_Layer |
Example of using multiple Activation Functions per Layer. |
Important
- If you have an error with 'POINTER_REGS' Click Here
- I am NOT a professional in any of all those fields [...] I'm an idiot in some cases too.
- Make sure that you have used (4-byte)(32-bit)-precision variables when Training, Because Floats:"...are stored as 32 bits (4 bytes) of information...get more precision by using a double (e.g. up to 15 digits), on the Arduino, double is the same size as float."
Features
-
Current
-
-
+
Use of activation-functions per layer-to-layer.
-
-
-
+
Optimizations based on user's preference.
-
-
-
+
MSE/BCE/CCE loss-functions.
-
-
-
+
Many activation-functions.
-
-
-
+
Use of PROGMEM.
-
-
-
+
Simplicity!
-
-
In the future, i want
-
-
-
Better code.
-
-
-
-
Use of custom functions.
-
-
-
-
More Activation Functions.
-
-
-
-
Different weight initialization methods.
-
-
-
-
Even more properties, for many different needs.
-
Tested On
- Arduino Uno
- ATtiny85
-
- (since the first release I haven't tested it again on the ATtiny85 at least yet, so I am not sure)
Functions, Variables & [...]
(NN) Neural-Network's Constructors |
---|
NeuralNetwork() |
NeuralNetwork(const unsigned int *_layer , const unsigned int &NumberOflayers , byte* _ActFunctionPerLayer=nullptr ) |
NeuralNetwork(const unsigned int *layer_ , const unsigned int &NumberOflayers , const float &LRw , const float &LRb , byte *_ActFunctionPerLayer=nullptr ) |
NeuralNetwork(const unsigned int *layer_ , float *default_Weights , float *default_Bias , const unsigned int &NumberOflayers , byte *_ActFunctionPerLayer=nullptr ); |
Type
[Main Functions]
NN Functions | Input Type (x) | Output Type (Y) | Action |
---|---|---|---|
FeedForward_Individual(x ) |
Float | 1D Float Array | "Feeds" the NN with each one X-input Individually until it returns Y-Output Values, If needed. (Way Less RAM too) |
*FeedForword(x ) |
1D Float Array | 1D Float Array | "Feeds" the NN with X-input values and returns Y-Output Values, If needed. |
BackProp(x ) |
1D Float Array | - | Tells to the NN if the outputs-Y were right/the-expected-X-inputs and then, teaches it. |
print() | - | String | Serial.Prints the weights and biases of NN. (If B10000000 prints from PROGMEM) |
float
[Loss Functions]
No need for #define MEAN_SQUARED_ERROR
, MSE is the default loss and it is always enabled. The only case in which you will also need to define the MSE in your sketch, is only if you want to use it in relation with another loss-function. To use any other loss-function except from MSE just define it as seen below.
Loss Functions | Enabling MACRO |
---|---|
NN.getMeanSqrdError (unsigned int batch_size ) |
#define MEAN_SQUARED_ERROR |
NN.getBinaryCrossEntropy (unsigned int batch_size ) |
#define BINARY_CROSS_ENTROPY |
NN.getCategoricalCrossEntropy (unsigned int batch_size ) |
#define CATEGORICAL_CROSS_ENTROPY |
float
[Loss Variables]
To use any of the variables below, you first need to #define
a loss function as said above too.
Loss variables | Sum variables |
---|---|
NN.MeanSqrdError | NN.sumSquaredError |
NN.BinaryCrossEntropy | NN.sumOfBinaryCrossEntropy |
NN.CategoricalCrossEntropy | NN.sumOfCategoricalCrossEntropy |
float
[Activation Functions]
Because of (my uncertainty and) the strict RAM optimization that allows the library to use one array that stores only the values after the activation instead of two arrays storing values before and after the activation, the use of some derivative functions in backpropagation are not supported by this library at this moment, as also seen by the MACRO NO_BACKPROP
below. This means that if you want to use and #define
any function from 8-13 under the section "NO_BACKPROP
support" , you won't be able to use backpropagation.
Enabling MACRO | Activation Functions | Action | |
---|---|---|---|
0 | #define Sigmoid |
NN.layers->Sigmoid (&x ) |
Returns 1/(1+e^(-x)) |
1 | #define Tanh |
NN.layers->Tanh (&x ) |
Returns (e^(2*x)-1)/(e^(2*x)+1) |
2 | #define ReLU |
NN.layers->ReLU (&x ) |
Returns (x>0)?x:0 |
3 | #define LeakyELU |
NN.layers->LeakyELU (&x ) |
Returns (x>0)?x:AlphaLeaky*x |
4 | #define ELU |
NN.layers->ELU (&x ) |
Returns (x>0)?x:AlphaELU*(e^(x)-1) |
5 | #define SELU |
NN.layers->SELU (&x ) |
Returns (x>0)?x:AlphaSELU*(e^(x)-1) |
6 | #define Softmax |
NN.layers->Softmax (&x ) |
void "complicated implementation" |
7 | #define Identity |
NN.layers->Identity (&x ) |
Returns x |
NO_BACKPROP support |
|||
8 | #define BinaryStep |
NN.layers->BinaryStep (&x ) |
Returns (x < 0) ? 0 : 1 |
9 | #define Softplus |
NN.layers->Softplus (&x ) |
Returns log(1 + exp(x)) |
10 | #define SiLU |
NN.layers->SiLU (&x ) |
Returns x / (1 + exp(-x)) |
11 | #define GELU |
NN.layers->GELU (&x ) |
Returns (1/2) * x * (1 + erf(x / sqrt(x))) |
12 | #define Mish |
NN.layers->Mish (&x ) |
Returns x * Tanh(log(1 + exp(x))) |
13 | #define Gaussian |
NN.layers->Gaussian (&x ) |
Returns exp(-(x*x)) |
0 | #define Sigmoid |
NN.layers->SigmoidDer (&fx ) |
Returns fx-fx*fx |
1 | #define Tanh |
NN.layers->TanhDer (&fx ) |
Returns 1-fx*fx |
2 | #define ReLU |
NN.layers->ReLUDer (&fx ) |
Returns (fx>0)?1:0 |
3 | #define LeakyELU |
NN.layers->LeakyELUDer (&fx ) |
Returns (fx>0)?1:AlphaLeaky |
4 | #define ELU |
NN.layers->ELUDer (&fx ) |
Returns (fx>0)?1:fx+AlphaELU |
5 | #define SELU |
NN.layers->SELUDer (&fx ) |
Returns (fx>0)?LamdaSELU:fx+AlphaSELU*LamdaSELU |
6 | #define Softmax |
NN.layers->SoftmaxDer (&fx ) |
Returns fx * (1 - fx) |
7 | #define Identity |
NN.layers->IdentityDer (&x ) |
Returns x |
if you want to use other activation function from the default one, just define one other:
#define Sigmoid //[default] No need definition, for single activation across network
#define Tanh
#define ReLU
#define LeakyELU
#define ELU
#define SELU
...
Use any activation function per layer-to-layer, like :
#define ACTIVATION__PER_LAYER
#include <NeuralNetwork.h>
unsigned int layers[] = {3, 4, ..., 2, 1};
byte Actv_Functions[] = { 1, ..., 2, 0};
// Tanh > ... > ReLU > Sigmoid
If you want to drastically reduce ROM & slightly RAM size you can Define which Functions to use/compile, like:
#define ACTIVATION__PER_LAYER
#define Sigmoid // 0
//#define Tanh
//#define ReLU
//#define LeakyELU
#define ELU // 1
#define SELU // 2
...
#include <NeuralNetwork.h>
unsigned int layers[] = {3, 4, ..., 2, 1};
byte Actv_Functions[] = { 1, ..., 2, 0};
// ELU > ... > SELU > Sigmoid
⚠️ have in mind that because I used f(x) for derivatives there might be chances of mistakes (?)
float
[Variables Of Activation Functions]
Enabling MACRO | Activation Variables | Default | Explenation |
---|---|---|---|
#define LeakyELU |
NN.AlphaLeaky | 0.01 | the α of Leaky |
#define ELU |
NN.AlphaELU | 1 | the α of ELU |
#define SELU |
NN.AlphaSELU | 1.6733 | the α of SELU |
#define SELU |
NN.LamdaSELU | 1.0507 | the λ of SELU |
Type
[Other Variables]
Type | NN's Variables | Explenation |
---|---|---|
byte* |
NN.ActFunctionPerLayer |
if ACTIVATION__PER_LAYER defined |
float |
NN.LearningRateOfWeights |
The Learning-Rate-Of-Weights |
float |
NN.LearningRateOfBiases |
The Learning-Rate-Of-Biases |
float* |
NN.weights |
If B00010000 defined |
Layer* |
NN.layers |
Layers of NN |
float* |
NN.layers[i].bias |
The bias of an individual layer[i] |
float* |
NN.layers[i].outputs [] |
The Output array of an individual layer[i] |
float** |
NN.layers[i].weights [][] |
if not B00010000 defined |
float* |
NN.layers[i].preLgamma [] |
The γ-error of previous layer[i-1] |
unsigned int |
NN.layers[i]._numberOfInputs |
The number-Of-Inputs for an individual layer[i] |
unsigned int |
NN.layers[i]._numberOfOutputs |
The number-Of-Outputs for an individual layer[i] |
#define
[MACRO Properties]
#define _1_OPTIMIZE B00000000
_1_OPTIMIZE | Action |
---|---|
B00000000 |
Nothing |
B10000000 |
⚠️ Enables the use of PROGMEM" instead of RAM |
B01000000 |
⚠️ Deletes previous layer's Outputs: Reduces RAM by a factor of ((the_sum_of_each_layer'_s _numberOfOutputs) - (_numberOfOutputs of_biggest_layer) *4 ) approximately i think ? |
B00100000 |
🔴 Partialy Reduces RAM for Weights, level 1 |
B00010000 |
Reduces RAM for Weights, level 2 by a factor of number_of_layers*2 - 2 bytes |
B00001000 |
🟢 Deletes previous layer's Gamma |
B00000100 |
Reduce RAM by a factor of 2*(number_of_layers - 1 or 2)bytes slightly different usage with multiple NNs |
B00000010 |
Recomended | Disables default MSE loss function | Reduces ROM, RAM & CPU consumption |
⚠️ = Backpropagation is not allowed
🔴 = Not yet implimented
🟢 = Always enabled (not switchable yet.)
Searches Across Internet
Here most of the resources I came across the internet, I recomend you to have a look if you want to (but please stay aware of the fact that for some of those sites, I had only opened them checked something and then closed them in a matter of seconds [so, please don't get them all siriously])
THANK YOU!
I also want especially to thanks Underpower Jet for his amazing tutorial, by bringing it more to the surface. Because after all the videos and links I came across, he was the one that made the most significant difference to my understanding of neural networks. Plus, I would like to thanks: giant_neural_network for this and this, 3Blue1Brown for this, the Arduino community and everyone else who gave me the oportunity to learn and make this library possible to exist [...]
New Searches:
12\08\2021
-
Neural Network Related
-
-
Videos:
-
-
-
Softmax:
-
-
-
StackOverflow\Exchange:
-
-
-
General:
-
-
Code Related:
-
-
Tools:
-
-
-
Macros:
-
-
-
Arduino:
-
-
-
-
StackOverflow:
-
-
-
-
-
General:
-
-
-
General:
-
-
Math:
-
-
-
Grammar:
-
-
-
Just "Random":
-
xx\xx\202x
-
Neural Network Related
-
-
kind of Intresting To me
-
-
-
General
-
-
-
Activation Function Related
-
-
-
Gradient Explosion and clipping Related
-
-
-
MNIST Related
-
-
Related to Programming
-
-
to C-type Languages
-
-
-
to Python
-
-
-
Other
-
-
Arduino Related
-
#MACROS / pre-processor directives
-
- [(Very Useful) Generate Functions](https://stackoverflow.com/a/1254012/11465149 )
Old Searches:
|
| Intresting |NN.
| Neural Network(s) |A.
| Arduino etc. |-
| Mostly .NET & Other |*
| Maybe Intresting?
Outro
Forgive me if I have made any mistakes and please don't take me seriously with every claim i made, I am mainly "self taught" in this field of NeuralNetworks, I am not a professional programmer nor do I have a good knowledge in many of the fields used to create this library [...]
I need help, I need to learn,
I want to learn, I want to help [...]
I wish love and happiness to Everyone! 🤍
Σημασία έχει, ότι ως αποτέλεσμα έχει την αγάπη.
if you want to help me and others to educate ourselves better and if you have a love and passion for sharing and helping others, then I highly suggest you to join our discord server 🤍
My Instagram account is: giorgos.xou ;) feel free to ask me anything
Donation
if you want me to continue support this library, learn and bring more functionality to it, it would be really helpfull if you would consider donating, even the least amount of 0.01$ would be really appreciated!
Bitcoin Address:
1N8niMND7SqE8B7rsE7Mx7gPctkx2rUMTf
(I need money, else I might gonna kill myself for real). stackedit.io have been used for the editing of this MD file