vladk-neural-network 0.1.7

Mini-framework that implements a simple version of a feedforward neural network (FNN) and convolutional neural network (CNN) from scratch using Python and PyTorch

Custom Neural Network Framework

This repository is a mini-framework that implements a simple version of a feedforward neural network (FNN) and convolutional neural network (CNN) from scratch using Python and PyTorch. PyTorch is used solely for mathematical and element-wise operations on tensors (without using autograd), and for speeding up computations by utilizing the GPU. Simply put, PyTorch is used as a replacement for NumPy but with GPU acceleration. The framework provides basic functionalities for building, training, and evaluating custom neural network models for both regression and classification tasks.

Features

• Feedforward Neural Network (FNN): A fully connected neural network suitable for regression and classification tasks.
• Convolutional Neural Network (CNN): A simplified CNN implementation with convolutional and max-pooling layers.
• Activation Functions: Includes ReLU, Leaky ReLU, Sigmoid, and Linear activations.
• Loss Functions: Support for mean squared error (MSE), binary cross-entropy (BCE) and categorical cross-entropy (CCE) losses.
• Optimizers: Implementations of stochastic gradient descent (SGD) and Adam optimizers.
• Metrics: Accuracy metric for classification tasks, especially useful for one-hot encoded data, and R2 score for regression tasks.

Installation

pip install vladk_neural_network

Usage

Data Format examples:

Example for regression:

# data shape (2, 1) - 2 input valus, 1 output value
dataset = [
    {
        "input": [0.1, 0.2],
        "output": [0.15],
    },
    {
        "input": [0.8, 0.9],
        "output": [0.7],
    },
]

Example for classification, output values one-hot encoded:

# data shape (4, 2) - 4 input valus, 2 output one-hot encoded values
dataset = [
    {
        "input": [0.13, 0.22, 0.37, 0.41],
        "output": [1.0, 0.0],
    },
    {
        "input": [0.76, 0.87, 0.91, 0.93],
        "output": [0.0, 1.0],
    },
]

Model creation examples:

Feedforward Neural Network for regression

from vladk_neural_network.model.activation import Linear, Relu
from vladk_neural_network.model.base import NeuralNetwork
from vladk_neural_network.model.layer import FullyConnected, Input
from vladk_neural_network.model.loss import MeanSquaredError
from vladk_neural_network.model.metric import R2Score
from vladk_neural_network.model.optimizer import SGD

# Build model
layers = [
    FullyConnected(64, Relu()),
    FullyConnected(64, Relu()),
    FullyConnected(1, Linear()),
]
nn = NeuralNetwork(
    Input(2),
    layers,
    optimizer=SGD(),
    loss=MeanSquaredError(),
    metric=R2Score()
)

# Train model
history = nn.fit(train_dataset, test_dataset, epochs=20, batch_size=1, verbose=True)

# Using model for prediction
prediction = nn.predict(test_dataset)

Convolutional Neural Network for classification

from vladk_neural_network.model.activation import LeakyRelu, Linear
from vladk_neural_network.model.base import NeuralNetwork
from vladk_neural_network.model.layer import (
    Convolutional,
    Flatten,
    FullyConnected,
    Input3D,
    MaxPool2D,
)
from vladk_neural_network.model.loss import CategoricalCrossEntropy
from vladk_neural_network.model.metric import AccuracyOneHot
from vladk_neural_network.model.optimizer import Adam

# Build model using gpu acceleration and applying argmax convert to raw prediction probabilities
layers = [
    Convolutional(LeakyRelu(), filters_num=4, kernel_size=3, padding_type="same"),
    Convolutional(LeakyRelu(), filters_num=8, kernel_size=3),
    Convolutional(LeakyRelu(), filters_num=16, kernel_size=3),
    MaxPool2D(),
    Flatten(),
    FullyConnected(64, LeakyRelu()),
    FullyConnected(10, Linear()),
]
cnn = NeuralNetwork(
    Input3D((1, 28, 28)),
    layers,
    optimizer=Adam(),
    loss=CategoricalCrossEntropy(),
    metric=AccuracyOneHot(),
    convert_prediction='argmax',
    use_gpu=True
)

# Train model
cnn.fit(train_dataset, test_dataset, epochs=10, batch_size=1, verbose=True)

# Using model for prediction
prediction = cnn.predict(test_dataset)

Several examples, including training feedforward and convolutional neural networks, are available in the form of Jupyter notebooks in the notebooks/ folder. You can view and run these examples to understand how to use the framework for different tasks.

License

This project is licensed under the MIT License. See the LICENSE file for more details.