PyANN

Python wrapper for the libann neural network library with a Keras-like API.

Features

• Keras-style API: Familiar Sequential model with add(), fit(), predict()
• Layers: Dense, plus Conv2D, MaxPool2D and Flatten for convolutional networks
• Multiple optimizers: SGD, Momentum, RMSProp, AdaGrad, Adam
• Activation functions: Sigmoid, ReLU, LeakyReLU, Tanh, Softsign, Softmax, GELU
• Loss functions: MSE, Categorical Cross-Entropy
• Online training: train_on_batch() for streaming data, fine-tuning and custom training loops
• Learning rate schedulers: Step decay, Exponential decay, Cosine annealing
• Hyperparameter tuning: Grid search, Random search, Bayesian optimization, TPE (pure-Python searches built on Sequential; libann's native hypertuner is not exposed)
• NumPy optional: Works with Python lists, optimized for NumPy arrays
• Model persistence: Save/load in text or binary format, ONNX export/import (including CNNs and PyTorch-style ONNX JSON)
• Deploy to microcontrollers: Quantize a trained model to int8 and export it as a standalone C header (integer-only inference, no malloc), or export float32 weights
• Network visualization: Export architecture as PIKCHR diagrams (renders to SVG)
• Regularization: L1 and L2 (weight decay) regularization, dropout
• Evaluation: Accuracy scoring, confusion matrix with MCC
• Data utilities: CSV loading, normalization, one-hot encoding, train/val splitting
• Hardware acceleration: multithreaded, BLAS-accelerated CPU training everywhere (Accelerate on macOS, OpenBLAS on Linux and Windows), plus optional GPU inference and training via Metal (macOS) or CUDA for dense networks

Installation

pip install pylibann

The PyPI package is called pylibann, but you import it as pyann. (The unrelated pyann project on PyPI is a different package with the same module name, so don't install both in one environment.)

Pre-built wheels include the native library, so no compiler is needed:

On other platforms pip builds from source, which needs CMake and a C compiler (see below). NumPy is optional: pip install "pylibann[numpy]".

Building from source

git clone --recurse-submodules https://github.com/mseminatore/pyann.git
cd pyann
pip install -e ".[dev]"      # builds the native library with CMake
pytest

The build picks the same acceleration as the wheels. On Linux and Windows it downloads the prebuilt OpenBLAS from the scipy-openblas32 package. Choose other options by passing CMake flags through CMAKE_ARGS:

CMAKE_ARGS="-DUSE_CUDA=ON" pip install -e .                                 # NVIDIA CUDA (Linux/Windows)
CMAKE_ARGS="-DUSE_CBLAS=ON" pip install -e .                                # CBLAS from /opt/cblas
CMAKE_ARGS="-DUSE_BLAS=ON" pip install -e .                                 # system OpenBLAS from /opt/OpenBLAS
CMAKE_ARGS="-DUSE_METAL=OFF -DUSE_ACCELERATE=OFF" pip install -e .          # plain CPU build

python build_lib.py [--accelerate | --scipy-openblas | --cblas | --blas] [--metal] [--cuda] [--clean] builds just the shared library into src/pyann/lib/ without installing, which is handy when working on the C side. Set PYANN_LIB_PATH to load a library from another location.

Check what a given install was built with:

import pyann
pyann.backend_info()   # {'blas': 'accelerate', 'gpu': 'metal', 'gpu_active': None, ...}

Quick Start

from pyann import Sequential, Dense, Adam, Loss

# Create model
model = Sequential(optimizer=Adam(lr=0.001), loss=Loss.MSE)
model.add(Dense(128, activation='relu', input_shape=(784,)))
model.add(Dense(64, activation='relu'))
model.add(Dense(10, activation='softmax'))

# Train
model.fit(X_train, y_train, epochs=10, batch_size=32)

# Predict
predictions = model.predict(X_test)

# Evaluate
accuracy = model.evaluate(X_test, y_test)
print(f"Accuracy: {accuracy:.2%}")

Examples

XOR Problem

from pyann import Sequential, Dense

# XOR data
X = [[0, 0], [0, 1], [1, 0], [1, 1]]
y = [[0], [1], [1], [0]]

model = Sequential()
model.add(Dense(4, activation='relu', input_shape=(2,)))
model.add(Dense(1, activation='sigmoid'))

model.fit(X, y, epochs=1000, batch_size=4)
print(model.predict(X))

With NumPy

import numpy as np
from pyann import Sequential, Dense

X = np.random.randn(1000, 20).astype(np.float32)
y = (X.sum(axis=1) > 0).astype(np.float32).reshape(-1, 1)

model = Sequential()
model.add(Dense(32, activation='relu', input_shape=(20,)))
model.add(Dense(16, activation='relu'))
model.add(Dense(1, activation='sigmoid'))

model.fit(X, y, epochs=50)

Convolutional Networks

Images are channels-first, as in PyTorch and ONNX: a batch has shape (N, channels, height, width). Channels-last arrays (Keras' default) need x.transpose(0, 3, 1, 2) first.

import numpy as np
from pyann import Sequential, Input, Conv2D, MaxPool2D, Flatten, Dense

# x: (N, 1, 28, 28) float32 images, y: (N, 10) one-hot labels
model = Sequential([
    Input((1, 28, 28)),
    Conv2D(8, 3, activation='relu'),                  # -> (8, 26, 26)
    MaxPool2D(2),                                     # -> (8, 13, 13)
    Conv2D(16, 3, activation='relu'),                 # -> (16, 11, 11)
    MaxPool2D(2),                                     # -> (16, 5, 5)
    Flatten(),                                        # -> (400,)
    Dense(64, activation='relu'),
    Dense(10, activation='softmax'),
])
model.compile(optimizer='adam', loss='cce')
model.fit(x, y, epochs=10, batch_size=32)
print(model.summary())

On Fashion-MNIST this network reaches about 90% test accuracy. predict(), evaluate() and train_on_batch() take image batches, a single image, or flattened rows. Convolutional networks train on the CPU (BLAS-accelerated); enable_gpu() only speeds up dense networks for now.

Learning Rate Scheduling

from pyann import Sequential, Dense
from pyann.callbacks import CosineAnnealing

model = Sequential()
model.add(Dense(64, activation='relu', input_shape=(10,)))
model.add(Dense(1, activation='sigmoid'))

scheduler = CosineAnnealing(T_max=100, min_lr=0.0001)
model.fit(X, y, epochs=100, lr_scheduler=scheduler)

Early Stopping

from pyann.callbacks import EarlyStopping

stopping = EarlyStopping(patience=10, min_delta=0.0001, restore_best=True)
history = model.fit(X, y, epochs=500, validation_data=(X_val, y_val), callbacks=[stopping])
print(history["best_epoch"], history["val_loss"])

libann measures the validation loss after every epoch and stops once it has not improved for patience epochs; with restore_best=True the model keeps the best epoch's weights. Without validation_data it monitors the training loss.

Online Training and Fine-Tuning

train_on_batch() runs one training step on a mini-batch and returns its loss. Unlike fit() it keeps the weights and the optimizer state (such as Adam's momentum) between calls, so it works for streaming data, custom training loops, and fine-tuning a loaded model.

model = Sequential.load('model.nna')        # or a model you just built and compiled

for x_batch, y_batch in stream():           # any source of (samples, features) batches
    loss = model.train_on_batch(x_batch, y_batch)

model.end_online_training()                 # optional: dropout off, GPU weights refreshed
model.save('model_finetuned.nna')

Batches can vary in size, predict() and evaluate() work between steps, and fit() ends any open session itself. The learning rate scheduler and callbacks only apply to fit(). Shapes are checked up front: x must be (samples, input_size) and y (samples, output_size), and a mismatch raises ValueError.

Hyperparameter Tuning

from pyann.hypertune import RandomSearch, HyperparamSpace
from pyann.utils import split_data

# Split data
X_train, y_train, X_val, y_val = split_data(X, y, train_ratio=0.8)

# Define search space
space = HyperparamSpace(
    learning_rate=(0.0001, 0.01),
    batch_sizes=[32, 64, 128],
    hidden_layers=(1, 3),
    layer_sizes=[64, 128, 256],
)

# Run search
search = RandomSearch(space, n_trials=20)
best, results = search.run(
    X_train, y_train, X_val, y_val,
    input_size=784, output_size=10
)

# Create best model
model = search.create_model(best, input_size=784, output_size=10)

TPE Hyperparameter Search

from pyann.hypertune import TPESearch, HyperparamSpace

space = HyperparamSpace(
    learning_rate=(0.0001, 0.01),
    batch_sizes=[32, 64, 128],
    hidden_layers=(1, 3),
    layer_sizes=[64, 128, 256],
)

# TPE uses KDE to model good vs bad configs
search = TPESearch(space, n_trials=50, n_startup=10, gamma=0.25)
best, results = search.run(
    X_train, y_train, X_val, y_val,
    input_size=784, output_size=10
)

Save and Load

# Save
model.save('model.nna')  # Text format
model.save('model.nnb', format='binary')  # Binary format
model.export_onnx('model.onnx.json')  # ONNX format

# Load
loaded = Sequential.load('model.nna')
loaded = Sequential.load_onnx('model.onnx.json')  # From ONNX

int8 Quantization and C Export

Quantize a trained model to int8, check its accuracy, and export it as a C header for a microcontroller or any C program. The header runs on ann_rt.h (copied next to it): integer-only math, no malloc, no libc needed for the int8 path. On Fashion-MNIST, int8 accuracy stays within 0.2 points of float32 with 4x less parameter memory.

q = model.quantize(x_train[:1000])        # calibrate on representative inputs
print(model.evaluate(x_test, y_test), q.evaluate(x_test, y_test))

q.export_c_header('model.h')              # writes model.h and ann_rt.h
model.export_c_header('model_f.h')        # float32 weights instead

On the device:

#include "model.h"

static int8_t workspace[MODEL_WORKSPACE_SIZE];
int8_t scores[MODEL_OUTPUT_SIZE];
model_predict(input_q, scores, workspace);  /* MODEL_INPUT_SIZE int8 values */
int label = model_argmax(scores);

q.quantize_input(x) gives the int8 inputs the device expects, and q.predict_int8(x_q) returns exactly what the header computes. See libann's C export guide for the scheme and generated API.

Network Visualization

# Export architecture as PIKCHR diagram
model.export_pikchr('network.pikchr')

# Then render to SVG with the pikchr CLI tool:
# pikchr network.pikchr > network.svg

Small networks (≤10 nodes per layer) get a detailed diagram with individual nodes and connections. Larger networks get a simplified box diagram.

Data Utilities

from pyann.utils import load_csv, split_data
from pyann.utils.data import normalize, one_hot_encode

# Load CSV data
data, rows, cols = load_csv('data.csv', has_header=True)

# Normalize features
X_norm, mean, std = normalize(X)

# One-hot encode labels
y_encoded = one_hot_encode(labels, num_classes=10)

# Split into train/validation sets
X_train, y_train, X_val, y_val = split_data(X, y, train_ratio=0.8)

API Reference

Sequential

Sequential(layers=None, optimizer=None, loss=Loss.MSE, name=None)

Methods:

• add(layer) - Add a layer
• compile(optimizer, loss, learning_rate, weight_decay, l1_regularization, gradient_clip, weight_init, convergence, dropout) - Configure training (see below)
• fit(x, y, epochs, batch_size, ...) - Train the model
• train_on_batch(x, y) - One training step on a mini-batch; returns the loss
• end_online_training() - End the online session started by train_on_batch()
• predict(x) - Generate predictions
• evaluate(x, y) - Compute accuracy
• evaluate_loss(x, y) - Compute the loss, averaged like the training loss
• confusion_matrix(x, y) - Compute TP/FP/TN/FN and MCC (binary classification)
• save(filepath, format) - Save model
• load(filepath, format) - Load model (classmethod)
• export_onnx(filepath) - Export to ONNX JSON format
• load_onnx(filepath) - Load from ONNX JSON (classmethod)
• quantize(x_calibration) - Make an int8 QuantizedModel
• export_c_header(filepath, prefix=None, runtime=True) - Export float32 weights as a C header (and copy ann_rt.h next to it)
• export_pikchr(filepath) - Export architecture as PIKCHR diagram
• export_learning_curve(filepath) - Export training history as CSV (adds val_loss when fit() had validation data)
• clear_history() - Clear training history to free memory
• set_weight_decay(lambda_) - Set L2 regularization coefficient
• set_l1_regularization(lambda_) - Set L1 regularization coefficient
• summary() - Get model summary

QuantizedModel

Made by Sequential.quantize(); independent of the model it came from.

• predict(x) - Real in and out, int8 inside (matches the header's <prefix>_predict_float)
• predict_int8(x_q) - int8 in and out (matches the header's <prefix>_predict bit for bit)
• quantize_input(x) / dequantize_output(y_q) - Convert between real and int8 values
• evaluate(x, y) - Accuracy of the int8 model
• export_c_header(filepath, prefix=None, runtime=True) - Export the int8 model as a C header
• input_scale, input_zero_point, output_scale, output_zero_point - Quantization: real = scale * (q - zero_point)

pyann.runtime_header_path() gives the path of the bundled ann_rt.h.

Backend

pyann.backend_info()   # dict: blas, gpu (compiled in), gpu_active, plus CBLAS details when present
pyann.enable_gpu()     # initialize the GPU backend; returns True if one is active
pyann.gpu_enabled()    # whether a GPU backend is active
pyann.set_num_threads(4)   # CPU threads for training/batched inference (0 = default, 1 = single-threaded)
pyann.get_num_threads()    # current CPU thread count

Training options

All are optional keyword arguments to compile(); unset options keep libann's defaults.

model.compile(optimizer='adam', loss='mse', gradient_clip=5.0, weight_init='he', convergence=0.0)

Layers

Dense(units, activation=None, input_shape=None, dropout=0.0)
Input(shape)              # (features,) or (channels, height, width)
Conv2D(filters, kernel_size=3, strides=1, padding=0, activation=None, input_shape=None, dropout=0.0)
MaxPool2D(pool_size=2, strides=None)   # strides default to pool_size; alias MaxPooling2D
Flatten()                 # optional: Dense layers read image outputs as vectors

Conv2D uses square kernels; padding is an int, 'valid' (0) or 'same' (keeps the size; odd kernel sizes with stride 1). Output size is (size + 2 * padding - kernel_size) // strides + 1. A model whose layers don't fit together raises NetworkError when it is built.

Optimizers

SGD(lr=0.05)
Momentum(lr=0.01)
RMSProp(lr=0.001)
AdaGrad(lr=0.01)
Adam(lr=0.001)  # Recommended

Loss Functions

Loss.MSE                        # Mean Squared Error (regression: use a linear output)
Loss.CATEGORICAL_CROSS_ENTROPY  # For classification
Loss.BINARY_CROSS_ENTROPY       # Alias of CATEGORICAL_CROSS_ENTROPY, for sigmoid outputs ('bce')

CATEGORICAL_CROSS_ENTROPY needs a softmax output layer (one class per sample) or a sigmoid output layer (binary cross-entropy per output). Training with any other output activation fails.

Activations

Activation.NONE      # Linear
Activation.SIGMOID
Activation.RELU
Activation.LEAKY_RELU
Activation.TANH
Activation.SOFTSIGN
Activation.SOFTMAX
Activation.GELU      # tanh approximation; ONNX export uses opset 20

Callbacks

StepDecay(step_size, gamma=0.5)          # Multiply LR by gamma every step_size epochs
ExponentialDecay(gamma=0.95)             # Multiply LR by gamma each epoch
CosineAnnealing(T_max, min_lr=0.0001)   # Smooth decay to min_lr over T_max epochs
EarlyStopping(patience=10, min_delta=0.0001, restore_best=True)  # Stop when (validation) loss stops improving

Hyperparameter Tuning

GridSearch(space, lr_steps=3)                          # Exhaustive grid search
RandomSearch(space, n_trials=50)                       # Random sampling
BayesianSearch(space, n_trials=50, n_initial=10)       # Bayesian optimization
TPESearch(space, n_trials=50, n_startup=10, gamma=0.25)  # Tree-structured Parzen Estimator

All accept a HyperparamSpace and return (best_result, all_results) from run().

What's Not Exposed

pyann covers libann's training, inference, persistence, regularization and GPU features. These parts of libann's C API have no Python equivalent yet:

• libann's native hypertuner (ann_hypertune.h). pyann.hypertune is a separate Python implementation, so its search spaces and results differ from the C one (for example topology names, per-layer activation search, and the Gaussian-process settings).
• Error-log callbacks: errors surface as Python exceptions instead.
• Tensor math (gemm, elementwise ops and so on): use NumPy.

tests/test_binding_coverage.py keeps this list honest. It compares libann's headers with the bindings and fails when upstream adds, removes or changes a function without a matching decision recorded there.

Hardware Acceleration

The CPU path uses the BLAS the library was built with (see backend_info()["blas"]). GPU use is opt-in:

import pyann

if pyann.enable_gpu():          # True when a GPU backend is compiled in and a device is available
    print("Using", pyann.backend_info()["gpu_active"])   # 'metal' or 'cuda'

model.fit(X, y, epochs=10)      # trains on the GPU when enabled
model.predict(X)                # batch inference on the GPU when enabled

If no GPU backend is built in, or no device is found, enable_gpu() returns False and everything keeps running on the CPU with BLAS, so the same code works everywhere.

Notes:

• enable_gpu() applies to the whole process and cannot be undone.
• GPU speedups depend on problem size. In an informal test on Apple silicon (batch size 128) Metal sped up batch inference but was several times slower than the CPU for training, which is why it is not enabled automatically. Measure your own workload.
• The macOS wheels include Metal. CUDA is only available when building from source with USE_CUDA=ON, and has had little testing.
• predict() runs the whole input as one batch call.

License

MIT License - see LICENSE for details.

The Linux and Windows wheels also bundle OpenBLAS, LAPACK and the GCC runtime libraries (libgfortran, libquadmath) under their own licenses, included in the wheel as pyann/lib/LICENSE-scipy-openblas32.txt.