TensorNN is a Python machine learning library built from scratch. It provides fundamental building blocks for creating and training neural networks.
Note: This library is currently under heavy development and is considered unusable for production purposes.
You can install TensorNN via pip (once it's published):
pip install tensornnTensorNN requires the following Python packages:
- numpy
- tqdm
For documentation, additional packages are required:
- sphinx
- sphinx_rtd_theme
TensorNN provides a range of modules to facilitate the creation and training of neural networks:
- Tensor Operations (
tensornn.tensor): CustomTensorclass based on NumPy for multi-dimensional array operations. - Neural Network Core (
tensornn.nn): TheNeuralNetworkclass to define and manage network architecture. - Layers (
tensornn.layers):Input: Specifies the input shape to the network.Dense: Fully connected layer.
- Activation Functions (
tensornn.activation): A variety of activation functions to introduce non-linearity, including:ReLUSigmoidSoftmaxTanhLeakyReLUELUSwishLecunTanhNoActivation(Linear)
- Loss Functions (
tensornn.loss): Various functions to evaluate model performance:MSE(Mean Squared Error)CategoricalCrossEntropyBinaryCrossEntropyRMSE(Root Mean Squared Error)MAE(Mean Absolute Error)MSLE(Mean Squared Logarithmic Error)PoissonSquaredHingeRSS(Residual Sum of Squares)
- Optimizers (
tensornn.optimizers): Algorithms to update network weights:SGD(Stochastic Gradient Descent)AdamRMSProp
- Utilities (
tensornn.utils): Helper functions likeone_hotencoding, datanormalize,set_seedfor reproducibility,flatten, andatleast_2d. - Debugging (
tensornn.debug): Tools to help inspect and debug your network's behavior, including progress bars and logging. - Custom Errors (
tensornn.errors): Specific error types for easier debugging (e.g.,NotRegisteredError,InitializationError,InputDimError).
Here's a simple example of how to define and train a neural network to solve the XOR problem:
from tensornn import Tensor, NeuralNetwork, Input, Dense, Tanh, Softmax, MSE, CategoricalCrossEntropy, SGD, Adam, set_seed
# For reproducibility
set_seed(0)
# 1. Define the data
inputs = Tensor([
[0, 0],
[0, 1],
[1, 0],
[1, 1],
])
# For regression (outputting a single value)
# outputs_reg = Tensor([[0], [1], [1], [0]])
# For classification (outputting probabilities for each class)
outputs_class = Tensor([
[1, 0], # Class 0
[0, 1], # Class 1
[0, 1], # Class 1
[1, 0], # Class 0
])
# 2. Create the Neural Network
# Example for classification using Softmax and CategoricalCrossEntropy
nn_classifier = NeuralNetwork([
Input(2), # Input layer with 2 features
Dense(2, activation=Tanh()), # Hidden layer with 2 neurons and Tanh activation
Dense(2, activation=Softmax()) # Output layer with 2 neurons (for 2 classes) and Softmax activation
])
# 3. Register the network with a loss function and an optimizer
nn_classifier.register(CategoricalCrossEntropy(), Adam(learning_rate=0.01))
# 4. Train the network
print("Training XOR classifier...")
nn_classifier.train(inputs, outputs_class, epochs=2000, batch_size=4)
# 5. Make predictions
predictions = nn_classifier.forward(inputs)
predicted_classes = predictions.argmax(axis=1)
print("\nClassifier Predictions:")
for i in range(len(inputs)):
print(f"Input: {inputs[i]}, Predicted probabilities: {predictions[i]}, Predicted class: {predicted_classes[i]}, Expected class: {outputs_class[i].argmax()}")You can find more examples in the examples/ directory, such as sine_vis.py which demonstrates training a network to learn the sine function with real-time visualization.
The official documentation for TensorNN is hosted on ReadTheDocs:
https://tensornn.readthedocs.io/en/latest/
The documentation is built using Sphinx from the files in the docs/ directory. Key configuration files include .readthedocs.yaml and docs/conf.py.
TensorNN uses pytest for testing. Tests are located in the tests/ directory. Examples include:
test_and.py: Tests the AND gate logic.test_xor_class.py: Tests XOR classification.test_xor_reg.py: Tests XOR regression.test_linear_regression.py: Tests simple linear regression.test_sine.py: Tests learning the sine function.
To run the tests, you would typically navigate to the root directory of the project and run:
pytest(Note: Ensure you have pytest installed: pip install pytest)
This project is licensed under the GNU General Public License v3.0. See the LICENSE file for more details.
- Arjun Sahlot (iarjun.sahlot@gmail.com)
- GitHub: ArjunSahlot