CVNN-Jamie 0.3.0

A neural network framework supporting complex and real-valued neural networks

CVNN_Jamie: Complex-Valued Neural Network Framework

CVNN_Jamie is a Python library for building and training complex-valued neural networks. It provides modular layers, a flexible Sequential model, a wide range of complex activation functions, and custom initialisation methods. Designed for research and experimentation with complex-valued data and models.

Features

• Complex-valued layers (Dense, Activation, etc.)
• Modular and extensible design
• Custom activation functions and derivatives
• Multiple initialisation methods (including custom/phase-constrained)
• Easy integration with NumPy
• Simple Sequential API for stacking layers and activations
• Full backpropagation and training support

Installation

Install from PyPI:

pip install CVNN_Jamie

Or from source:

pip install -r requirements.txt

---

Model API: Sequential

Model Construction: Real and Complex Networks

Complex-Valued Network (default)

from cvnn import Sequential
from cvnn.layers import Dense
from cvnn.activations import complex_relu
import numpy as np

# Each Dense layer is complex by default
model = Sequential([
	Dense(input_dim=4, output_dim=2),
	complex_relu
])
x = np.random.randn(1, 4) + 1j * np.random.randn(1, 4)
out = model.forward(x)
print("Single-layer output (complex):", out)

Real-Valued Network

from cvnn import Sequential
from cvnn.layers import Dense
from cvnn.activations import complex_relu
import numpy as np

# Use real=True for Dense, and pass real=True to activations
model = Sequential([
	Dense(input_dim=4, output_dim=2, real=True),
	lambda x: complex_relu(x, real=True)
], complex=False)
x = np.random.randn(1, 4)
out = model.forward(x)
print("Single-layer output (real):", out)

Multilayer Real or Complex Network

from cvnn import Sequential
from cvnn.layers import Dense
from cvnn.activations import complex_tanh, complex_sigmoid
import numpy as np

# Real-valued multilayer
model = Sequential([
	Dense(2, 2, real=True),
	(lambda x: complex_tanh(x, real=True), lambda z, g: complex_tanh_backward(z, g, real=True)),
	Dense(2, 1, real=True),
	(lambda x: complex_sigmoid(x, real=True), lambda z, g: complex_sigmoid_backward(z, g, real=True))
], complex=False)

# Complex-valued multilayer
model_c = Sequential([
	Dense(2, 2),
	(complex_tanh, complex_tanh_backward),
	Dense(2, 1),
	(complex_sigmoid, complex_sigmoid_backward)
])

Training (Demo: 1-layer, MSE loss, SGD)

from cvnn import Sequential
from cvnn.layers import Dense
import numpy as np

# Dummy data: learn identity mapping (complex)
x = np.random.randn(10, 2) + 1j * np.random.randn(10, 2)
y = x.copy()
model = Sequential([
	Dense(input_dim=2, output_dim=2)
])
model.fit(x, y, epochs=50, lr=0.01)

Complex-Valued Neural Network (CVNN) Framework

This library provides a framework for building complex-valued neural networks in Python. It includes core modules for layers, activations, and operations that support complex numbers.

Features

• Complex-valued layers (Dense, Activation, etc.)
• Modular and extensible design
• Easy integration with NumPy

Getting Started

Install requirements:

pip install -r requirements.txt

Example Usage

ComplexDense Layer with Custom Initialisation

from cvnn.layers import ComplexDense
from cvnn.activations import complex_glorot_uniform, jamie
import numpy as np

# Use Glorot uniform for weights, jamie for bias
layer = ComplexDense(input_dim=4, output_dim=2, weight_init=complex_glorot_uniform, bias_init=jamie)
x = np.random.randn(1, 4) + 1j * np.random.randn(1, 4)
out = layer.forward(x)
print(out)

Using Activation Functions

from cvnn.activations import complex_relu, complex_sigmoid, complex_tanh, modrelu

z = np.array([1+2j, -1-2j, 0+0j, -3+4j])
print("ReLU:", complex_relu(z))
print("Sigmoid (separable):", complex_sigmoid(z))
print("Sigmoid (fully complex):", complex_sigmoid(z, fully_complex=True))
print("Tanh (separable):", complex_tanh(z))
print("Tanh (fully complex):", complex_tanh(z, fully_complex=True))
print("modReLU:", modrelu(z, bias=0.5))

Using Initialisation Methods Directly

from cvnn.activations import complex_zeros, complex_ones, complex_normal, complex_glorot_uniform, complex_he_normal, jamie

w = complex_zeros((3, 2))
b = jamie((1, 2))
print("Zeros init:", w)
print("Jamie init:", b)

License

MIT