Lightweight and easy to use ml library for small projects, create a neural network in minutes.
Project description
POF - pile of feather
This library is not an alternative to big ml library like pytorch or tensorflow, it lacks features and optimization, such as gpu support. The goal is to create a lightweight library of about 100 lines of code that is easy to use and quick to implement for creating small projects or experiments.
Install module
pip install pileoffeather
Examples
examples/simple_image_classifier - Handwritten digit image classifier, dataset is made out of 500 images of ones and 500 images of zeros taken from the mnist dataset. The first code snippet is defining the neural network model, uploading the dataset and than training the model
from pileoffeather import pof, pod
import numpy as np
#Define neural network model
model = pof.neuralNetwork(layers = [[400,"input"],[30,"relu"],[10,"relu"],[1,"sigmoid"]], name = "test1")
#Load the images for the dataset, 500 ones images and 500 zeros images
ones = pod.load(data_type = "image", color = "grayscale", folder = "ones", resize = (20,20))
zeros = pod.load(data_type = "image", color = "grayscale", folder = "zeros", resize = (20,20))
input = pod.merge(ones, zeros)
#Generate expected output, first 500 images should output 1, the other 500 0
output = np.concatenate((np.ones(500), np.zeros(500)))
#Train the neural network using backpropagation
pof.train(model, input, output, batch_size = 16, epoch_number = 100, rate = 0.6)
The second code snippet is testing the neural network on some examples
from pileoffeather import pof, pod
#Define neural network model
model = pof.neuralNetwork(load = "test1")
#Run model
input = pod.loadImage("example_image_one.png", (20,20), "grayscale")
output = model.run(input)
#Print result
print(round(output[0],3))
examples/mnist - Training script for the mnist dataset, 2 epochs -> 20~ seconds on 12600k -> 96%+ accuracy on 10k-test dataset
from pileoffeather import pof, pod
#Define neural network model
model = pof.neuralNetwork(layers = [[784,""],[128,"relu"],[10,"sigmoid"]], name = "mnist")
#Upload mnist dataset
input_dataset = pod.load(data_type = "gz", path = "train-images-idx3-ubyte.gz", start_index = 16, input_number = 784, divide = 255)
output_dataset = pod.load(data_type = "gz", path = "train-labels-idx1-ubyte.gz", start_index = 8, one_hot = 10)
#Train the neural network using backpropagation
pof.train(model, input_dataset, output_dataset, batch_size = 12, epoch_number = 2, rate = 1)
Documentation
The library is structured in 2 files, pof.py (pile of feather) is used to create and train neural networks, pod.py (pile of data) is used import your own data to generate a training dataset.
pof.py - neural network module
Install module
pip install pileoffeather
Import module
from pileoffeather import pof
Define neural network model, the available activation functions are "sigmoid","relu","leakyRelu"
model = pof.neuralNetwork(layers = [[400,""],[50,"relu"],[10,"relu"],[1,"sigmoid"]], name = "test1")
Save the model
model.save()
Load an exsisting model
model = pof.neuralNetwork(load = "test1")
Use the neural network
output = model.run(input)
Compute backpropagation for a single batch, model.computeBatch(batch_input, batch_output, batch_size, learning_rate)
model.computeBatch(batch_input, batch_output, 16, 0.3)
Method that contains model.computeBatch() in a loop so that it iterates over all the dataset for n epochs
pof.train(model, input_matrix, output_matrix, batch_size = 16, epoch_number = 100, rate = 0.03)
pod.py - data load module
Import module
from pileoffeather import pod
Load dataset of images from local folder
dataset = pod.load(data_type = "image", color = "grayscale", folder = "folder_name_containing_all_images", resize = (20,20))
Load dataset from gz file like mnist
#Load training input data of mnist, normalize input from 0 to 1 using divide = 255
dataset = pod.load(data_type = "gz", path = "train-images-idx3-ubyte.gz", start_index = 16, input_number = 784, divide = 255)
#Load training output data of mnist, use one_hot encoding to convert a decimal number to an array (pass total number of classes as parameter)
#4 -> [0,0,0,0,1,0,0,0,0,0] 0 -> [1,0,0,0,0,0,0,0,0,0]
dataset = pod.load(data_type = "gz", path = "train-labels-idx1-ubyte.gz", start_index = 8, one_hot = 10)
Load a single image to feed the neural network loadImage(name, resize, color)
input_vector = pod.loadImage("example_image.png", (20,20), "grayscale")
Convert neural network output to image and save, saveImage(neural_network_output, image_path, resize, color)
pod.saveImage(neural_network_output, "image_path_and_name", (20,20), "grayscale")
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for pileoffeather-0.3.1-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 85feb9d8375a65505f4023f39ecd19b3e37d2dc073d36eec58e0a0a124679053 |
|
MD5 | 96d28f928708361b51ca3e5e07b3614e |
|
BLAKE2b-256 | dc261a23824ab1908ef8c4b5e6b2c4a70d7ce28f1e442798306edab4c1d7a9b6 |