.TH ML_CONFIG "5" "September 2024" "ml 0.5.0" "Configuration file"

.SH NAME
ml_config - ml configuration file

.SH DESCRIPTION
.PP
ml configuration file format consists of lines of key-value options grouped in
multiple sections as is shown below in the demo configuration file:
.PP

.EX
.RS 4
; comments
[net]
loss = square ; options (square)
epochs = 500 ; comment
batch = 32
alpha = 1
weights_path = utils/weights.bin
inputs = sepal_length,sepal_width,petal_length,petal_width
labels = species


[categorical_fields]
species=setosa,versicolor,virginica

[preprocessing]
onehot=species


[layer]
neurons=10
; options (relu, sigmoid, softplus, leaky_relu, linear, tanh)
activation=sigmoid

[outlayer]
activation = sigmoid
.RE
.EE

.SS [net]
.PP
In this section you describe general network configurations more specifically
its training parameter, weights filepath and input and output columns.

.TS
box nospaces center tab(|);
L L L
Lb L L.
Key | Description | Type
_
alpha           | learning rate     | decimal
loss            | loss function     | option (string)
epochs          | training epochs   | integer
batch           | batch size        | integer
weights_path    | weights filepath  | string
inputs          | input fields      | list (string)
labels          | label fields      | list (string)
.TE

.SS [preprocessing]
Indicate preprocessing operations for input or label fields


.TS
box nospaces center tab(|);
L L L
Lb L L.
Key | Description | Type
_
onehot  | onehot columns | list (string)
.TE

.PP
Categorical preprocessing like operations
.B onehot
require to have their fields specified in
.B [categorical_fields].
otherwise the program will prompt you an error message.


.SS [categorical_fields]
.PP
This section is kind of special, here you must set the possible values a field
can take. taking the first example as reference:

.EX
.RS
[categorical_fields]
species=setosa,versicolor,virginica
.RE
.EE

the column
.B species
can only take the values
.B setosa,versicolor
and
.B virginica,
if the program encounters a different value it will stop its execution.

.SS [layer]
.SS [outlayer]
.PP
Create a layer component.
.B [outlayer]
section does not require
.B neurons
parameter.
.TS
box nospaces center tab(:);
L L L
Lb L L.
Key : Description : Type
_
neurons     :  number of neurons            : integer
activation  :  activation function          : option (string)
.TE

.PP
The topology of the network depends of the order in which
.B [layer]
, is put on the file, the program will load each
.B [layer]
section sequentially. For instance the following example:
.PP

.EX
.RS
[layer]
neurons=10
activation=relu

[layer]
neurons=20
activation=sigmoid

[outlayer]
activation = sigmoid
.RE
.EE

will produce a network with the following topology
.TS
nospaces center tab(:);
Lb | Lb
L | L.
Activation : Dimension
_
ReLu        : in \[mu] 10
Sigmoid     : 10 \[mu] 20
outlayers   : 20 \[mu] out
.TE

.SH FILES
.PP
The configuration file is searched in the following order:
.PP
.PD 0
.IP \(bu 4
CLI filepath
.IP \(bu 4
ML_CONFIG_PATH env variable
.IP \(bu 4 
$HOME/.\&config/ml/ml.cfg
.PP

.SH AUTHOR
Written by jvech

.SH COPYRIGHT
Copyright \(co 2024  jvech

.PP
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.