https://github.com/mikhailkravets/spytnik
Deep learning field of experiments
https://github.com/mikhailkravets/spytnik
deep-learning learning mathematics neuralnetworks
Last synced: 26 days ago
JSON representation
Deep learning field of experiments
- Host: GitHub
- URL: https://github.com/mikhailkravets/spytnik
- Owner: MikhailKravets
- License: mit
- Created: 2017-11-28T18:46:54.000Z (almost 8 years ago)
- Default Branch: master
- Last Pushed: 2017-12-25T15:28:51.000Z (almost 8 years ago)
- Last Synced: 2025-08-18T01:37:18.276Z (3 months ago)
- Topics: deep-learning, learning, mathematics, neuralnetworks
- Language: Python
- Homepage:
- Size: 313 KB
- Stars: 0
- Watchers: 0
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.rst
- License: LICENSE
Awesome Lists containing this project
README
.. raw:: html
.. image:: https://travis-ci.org/MikhailKravets/Spytnik.svg?branch=master
:target: https://travis-ci.org/MikhailKravets/Spytnik
.. image:: https://codecov.io/gh/MikhailKravets/Spytnik/branch/master/graph/badge.svg
:target: https://codecov.io/gh/MikhailKravets/Spytnik
Spytnik is the Python project that gathers modern deep learning algorithms; the field of experiments
and place for application of new knowledge.
Installation
************
Clone (download) the project and run the following command
.. code::
python setup.py install
Then you can import the project
.. code:: python
import spytnik
Usage example
*************
Let's try to approximate very simple function `0.5*sin(e^x) - cos(e^(-x))`.
.. code:: python
# import all we need
import matplotlib.pyplot as plot
import random
import numpy
# that is our project's packages. Mehh... it's not documented yet :(
import layers
from core import FeedForward, separate_data
from core.estimators import cv
# the very simple function
def f(x):
return 0.5 * numpy.sin(numpy.exp(x)) - numpy.cos(numpy.exp(-1 * x))
# creating new FeedForward instance.
# FeedForward is our neural netrowk which we'll be trained
nn = FeedForward(momentum=0.2, learn_rate=0.05, weight_decay=0.2)
# Append layers to neural networks
# That's an interesting moment:
# You can combine any layers you want. There is only one constraint: the first layer must
# have the amount of input vector equal to input vector length; such as output layer must
# have the amount of neurons equal to output vector length.
#
# Such way, here we create input layer (arguments (1, 10) means that the layer contains
# input vector of length 1 to each neuron and 10 synapsis are outputed from the each neuron in layer);
#
# four hidden layers with Tanh activation function;
# and one linear output layer
nn += layers.Tanh(1, 10)
nn += layers.Tanh(10, 10)
nn += layers.Tanh(10, 10)
nn += layers.Tanh(10, 10)
nn += layers.Linear(10, 1)
# generate input data and desired output to this data
data = [([x], [f(x)]) for x in numpy.linspace(-2.2, 2.5, 150)]
# separate it on the training set and validation ses
ts, vs = separate_data(data, 0.15)
# duplicate x and y for easier plotting
x = numpy.linspace(-2.2, 2.5, 150)
y = f(x)
# train the neural networks feeding it the elements from training set randomly
error = []
v_error = []
for i in range(50_000):
r = random.randint(0, len(ts) - 1)
# here we train it... step by step...
nn.fit(ts[r][0], ts[r][1])
# just to see on the errors further
error.append(nn.error)
if i % 300 == 0:
v_error.append(cv(nn, vs))
# use our trained neural network for approximation of our simple function
y_trained = []
for v in x:
y_trained.append(nn.get([v])[0])
# just plot the result
plot.subplot(211)
plot.title("f(x) and its approximation")
plot.plot(x, y)
plot.plot(x, y_trained)
plot.subplot(212)
plot.title("Learning error")
plot.plot(error)
plot.plot([i * 300 for i in range(len(v_error))], v_error)
plot.show()
Plotted charts after running of the code above should be similar to this one ↓
.. raw:: html
Awesome, yeah?
There will be some more awesome information in the readme or even personal website but some later, wait a little.
Documentation and so on
***********************
I will write documentation with mathematical background to it, I promise... just believe me
License
*******
MIT License
Attribution
***********
The red sputnik from logo is made by `Freepik `_