netpy 0.0.0.1

Library for ML

# NetPy

[](https://github.com/maclinchy/netpy/blob/master/LICENSE)

## Installation

### We recommended to install NetPy in virtual env.
To get started, you should install ```virtualenv``` tool with ```pip```.

```sh
pip install virtualenv
```
New directory to work with:
```sh
mkdir python-virtual-environments && cd python-virtual-environments
```
Create a new virtual environment inside the directory:
```sh
python -m venv netpy_env
```
You need to "activate" it.
```sh
source netpy_env/bin/activate
```
If your system is windows:
```sh
netpy_env\Scripts\activate
```

Then you need to install NetPy. There is two ways to install:

- **Install NetPy from PyPI (recommended):**

```sh
sudo pip install netpy
```

- **Install NetPy from GitHub source:**

```sh
git clone https://github.com/maclinchy/netpy.git
cd netpy
sudo python setup.py install
```

------------------
## Getting started

Let's create the net:
```sh
nhm newnet myFirstNet
```

`netpy-manager` will create folder `myFirstNet`, which contains a directory structure similar to this:

```sh
├── data_set
│ ├── data_X.txt
│ └── data_Y.txt
├── net_data
│ └── arch.json
├── conf.json
├── network.py
├── test.py
└── train.py

```

`data_set` is folder which contains data for learning.

`net_data` is folder which contains data about our net.

`conf.json` is file with settings.

`network.py` is main file for init the net.

`train.py` is file for train the net

`test.py` is file for test the net.

Let's create net with 3 layers.
------------------
`network.py`
```python
from netpy.nets import FeedForwardNet
from netpy.modules import LinearLayer,SigmoidLayer, FullRelation
from netpy.teachers import BackPropTeacher
import numpy as np

# Import file with settings
import conf

net = FeedForwardNet()

# Add your layers here
input_layer = LinearLayer()
hidden_layer = SigmoidLayer()
output_layer = SigmoidLayer()

# Add your layers to the net here
net.add_Input_Layer(input_layer)
net.add_Layer(hidden_layer)
net.add_Output_Layer(output_layer)

# Add your connections here
rel_in_hid = FullRelation(input_layer, hidden_layer)
rel_hid_out = FullRelation(hidden_layer, output_layer)

# Add your connections to the net here
net.add_Relation(rel_in_hid)
net.add_Relation(rel_hid_out)
```

`train.py`
```python
from netpy.teachers import BackPropTeacher
from network import net
import numpy as np
import conf

num_of_epoch = 8

# Your data set
train_X = np.loadtxt(conf.data_set_dir_X)
train_Y = np.loadtxt(conf.data_set_dir_Y)

# Config your teacher
teacher = BackPropTeacher(net,
error = 'MSE',
learning_rate = 1e-5,
alpha = 0.0001)


# Teach your net here
teacher.train(X, Y, num_of_epoch, load_data = False)
```

`test.py`
```python

from network import net

test = []

net.load()

print(net.forward(test))
```

We need to add neurons to the net:
```python
input_layer = LinearLayer(3)
hidden_layer = SigmoidLayer(20)
output_layer = SigmoidLayer(1)
```

Then we need to add the data:
`data_X.txt`
```sh
0 0 0
0 0 1
0 1 0
0 1 1
1 0 0
1 0 1
1 1 0
1 1 1
```
`data_Y.txt`
```sh
0
0
0
1
0
1
1
1
```

Let's train the net!
```sh
python train.py
```

In `net_data` two files - `weights.h5` and `arch.json`.

- In `weights.h5` we have save our weights.
- In `arch.json` we have save data about the net.
- In `myFirstNet.log` we have history of net's actions.

If you want to continue learning you should change parameter of train function:
```python
teacher.train(X, Y, 100, load_data = True)
```

Let's test the net!

```sh
python test.py
```

`test.py`
```python

from network import net

test = [0,0,0]

net.load()

print(net.forward(test))
```

You can fork your net:
```sh
nhm fork myFirstNet mySecondNet
```

You can rename your net:
```sh
nhm rename mySecondNet myAwesomeNet
```