kronyx 1.0.1

A lightweight deep learning framework built from scratch using NumPy

Kronyx

A lightweight deep learning framework built from first principles using NumPy. Designed for education, research, and production use with a clean Keras-like API.

Features

Feature	Description
Pure NumPy	No external ML dependencies, just NumPy
Clean API	Keras-like Sequential model interface
Educational	Built for learning with visualization and inspection tools
Layers	Dense, Conv2D, Flatten, Dropout, BatchNormalization
Activations	ReLU, Sigmoid, Tanh, Softmax
Optimizers	SGD, Adam with full state management
Callbacks	EarlyStopping, ModelCheckpoint, CSVLogger, ReduceLROnPlateau
Serialization	Save/load models with .krx format
Visualization	history.plot(), model.visualize(), model.summary()

Installation

pip install kronyx

For development:

git clone https://github.com/Kronyx/kronyx.git
cd kronyx
pip install -e ".[dev]"

Quick Start

import numpy as np
from kronyx import Sequential, Dense, ReLU, Sigmoid, BinaryCrossEntropy, Adam, Accuracy

# XOR problem - binary classification
X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([[0], [1], [1], [0]])

model = Sequential()
model.add(Dense(2, 8))  # input_size=2, output_size=8
model.add(ReLU())
model.add(Dense(8, 1))
model.add(Sigmoid())

model.compile(
    loss=BinaryCrossEntropy(),
    optimizer=Adam(learning_rate=0.1),
    metric=Accuracy()
)

model.fit(X, y, epochs=1000)
predictions = model.predict(X)
print(f"Accuracy: {(predictions.round() == y).mean():.2%}")

Binary Classification Example

import numpy as np
from kronyx import Sequential, Dense, ReLU, Sigmoid, BinaryCrossEntropy, Adam, Accuracy

X_train = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y_train = np.array([[0], [1], [1], [0]])

model = Sequential()
model.add(Dense(2, 16))
model.add(ReLU())
model.add(Dense(16, 1))
model.add(Sigmoid())

model.compile(
    loss=BinaryCrossEntropy(),
    optimizer=Adam(learning_rate=0.1),
    metric=Accuracy()
)

history = model.fit(X_train, y_train, epochs=500)
model.summary()

Multi-class Classification Example

import numpy as np
from kronyx import Sequential, Dense, ReLU, SoftmaxCategoricalCrossEntropy, Adam, Accuracy

# One-hot encoded labels
X = np.random.randn(100, 4)
y = np.eye(3)[np.random.randint(0, 3, 100)]

model = Sequential()
model.add(Dense(4, 32))
model.add(ReLU())
model.add(Dense(32, 3))
model.add(Softmax())

model.compile(
    loss=SoftmaxCategoricalCrossEntropy(),
    optimizer=Adam(learning_rate=0.01),
    metric=Accuracy()
)

model.fit(X, y, epochs=100)

Convolutional Neural Network Example

import numpy as np
from kronyx import Sequential, Conv2D, ReLU, Flatten, Dense, Softmax

# Simple image input (batch, height, width, channels)
X = np.random.randn(10, 8, 8, 1)
y = np.eye(2)[np.random.randint(0, 2, 10)]

model = Sequential()
model.add(Conv2D(filters=8, kernel_size=3, padding='same'))
model.add(ReLU())
model.add(Flatten())
model.add(Dense(64, 2))
model.add(Softmax())

model.compile(
    loss=SoftmaxCategoricalCrossEntropy(),
    optimizer=Adam(learning_rate=0.01),
    metric=Accuracy()
)

model.fit(X, y, epochs=10)

Saving and Loading Models

# Save complete model with architecture, weights, and configuration
model.save('model.krx')

# Load complete model
loaded = load_model('model.krx')

# Continue training or make predictions
loaded.predict(X_test)

Saving and Loading Weights

# Save only trainable weights
model.save_weights('weights.npz')

# Create a new model with matching architecture
new_model = Sequential()
new_model.add(Dense(2, 8))
new_model.add(ReLU())
new_model.add(Dense(8, 1))
new_model.add(Sigmoid())

# Load weights into the new model
new_model.load_weights('weights.npz')

Exporting JSON Architecture

# Export architecture to JSON string
json_str = model.to_json()

# Create model from JSON (weights initialized randomly)
model = Sequential.from_json(json_str)

.krx Archive Format

The .krx format is a zip archive containing:

model.krx
├── metadata.json      # Framework version, Python/numpy versions, timestamp
├── architecture.json  # Layer configuration, loss, optimizer settings
├── weights.npz        # Trainable weights and biases (numpy archive)
└── optimizer.npz      # Optional: optimizer state for resumable training

Examples

• examples/xor.py - XOR problem with binary classification
• examples/iris_classification.py - Iris dataset multi-class classification
• examples/iris_dropout.py - Dropout regularization example
• examples/iris_l2.py - L2 weight regularization
• examples/batchnorm_iris.py - Batch normalization demonstration
• examples/flatten_demo.py - Multi-dimensional input handling
• examples/conv2d_demo.py - Convolutional neural network example
• examples/mnist_classifier.py - MNIST digit classification

Documentation

kronyx/
├── activations.py    # ReLU, Sigmoid, Tanh, Softmax
├── layers.py         # Dense, Conv2D, Flatten, Dropout, BatchNormalization
├── model.py          # Sequential, History
├── optimizers.py     # SGD, Adam
├── losses.py         # BinaryCrossEntropy, CategoricalCrossEntropy
├── metrics.py        # Accuracy
├── callbacks.py      # Callback base, EarlyStopping, etc.
├── regularizers.py   # L2
├── initializers.py   # he_normal, xavier_uniform, lecun_normal
├── serialization.py  # Save/load model (.krx format)
├── utils.py          # Utility functions
└── exceptions.py     # Error types

Contributing

See CONTRIBUTING.md for development guidelines.

Roadmap

See ROADMAP.md for planned features.

License

MIT License - see LICENSE