MorphosML v0.3.0

MorphosML is a high-performance machine learning, differentiable computing, and out-of-core data ingestion library engineered in modern C++17 with zero-overhead Python bindings.

Designed from first principles for High-Performance Computing (HPC), MorphosML saturates CPU cache lines, vector registers (AVX2/AVX-512), and NVMe storage bandwidth while exposing an intuitive, Pythonic API.

---

📚 Documentation & Guides

Explore the comprehensive MorphosML documentation library:

• 🏗️ Architectural Blueprint: Deep dive into contiguous memory layouts, zero-copy buffer protocol, POSIX mmap kernel caching, and asynchronous double buffering.
• 📖 API Reference: Exhaustive class and method specifications for Matrix, Vector, TensorView, MMapDataset, DataLoader, calculus, and models.
• 🚀 Quickstart Tutorial: Step-by-step introduction from installation to model evaluation.
• ⚡ High-Performance Ingestion Tutorial: Out-of-core .mldat streaming, idempotent sampling, and cursor checkpoint recovery.
• 🧮 Calculus & Differentiable Programming: $\mathcal{O}(h^4)$ Taylor stencils, gradients, Jacobians, Hessians, quadrature, and limits.
• 📐 LaTeX Visualization & Jupyter Integration: LaTeX mathematical typesetting and native interactive notebook rendering.
• 🤝 Contributing Guide: Development setup, building C++ extensions, formatting standards, and pull request workflows.

---

Table of Contents

• Core Principles
• Architecture & Performance
• Installation
• Quickstart Guide
• Memory & Linear Algebra Layer
• Out-of-Core Data Ingestion Engine
• Calculus & Differentiable Programming Core
• Machine Learning Models
• Latexify & Jupyter Notebook Integration
• Verification & CI/CD Pipeline
• License & Author

---

Core Principles

MorphosML is governed by four core architectural tenets:

Principle	Engineering Implementation
High Performance (HPC)	Contiguous 1D flat buffers, zero-copy Python Buffer Protocol, $i$-$k$-$j$ cache-line matrix multiplication, and POSIX mmap with MADV_SEQUENTIAL.
Idempotency	Reproducible batch sequencing and shuffling across epochs and workers using deterministic SplitMix64 PRNG: $\text{state} = \text{seed} \oplus (\text{epoch} \cdot C_1 + C_2)$.
Fault-Tolerance	Sub-millisecond cursor tokens (epoch:sample_offset:checksum). Training pipelines resume mid-epoch after crashes without re-streaming from line 0.
Resiliency	FNV-1a 64-bit binary checksum validation, shape verification, and poison record isolation preventing memory faults or silent data corruption.

---

Architecture & Performance

MorphosML System Topology
├── Core Memory & Linear Algebra Layer (C++17)
│   ├── Vector              (Contiguous double-precision vector)
│   ├── Matrix              (Contiguous 1D flat row-major buffer with pybind11 buffer protocol)
│   └── TensorView          (Non-owning, zero-copy contiguous 2D view with read-only buffer protocol)
├── High-Throughput Storage & Ingestion Engine
│   ├── MMapBuffer          (POSIX mmap reader, MADV_SEQUENTIAL line-rate streaming)
│   ├── Native Binary .mldat (64-byte structured header + raw payload + FNV-1a checksum)
│   ├── IdempotentSampler   (Deterministic SplitMix64 Fisher-Yates generator)
│   ├── IngestionCursor     (16-byte state checkpoint: {epoch, sample_offset, checksum})
│   └── DataLoader          (Double-buffered async prefetch queue for zero I/O bubbles)
├── Numerical Calculus Engine
│   ├── Differentiation     (5-point stencils O(h^4) derivatives, gradients, Jacobians, Hessians)
│   ├── Integration         (Trapezoidal, Simpson's 1/3 O(h^4), Simpson's 3/8, Gauss-Legendre, 2D)
│   └── Limits              (Left, right, two-sided limits with asymptotic divergence detection)
├── Machine Learning Models & Metrics
│   ├── LinearRegression    (C++ gradient descent, MSE cost, R^2 score)
│   ├── LogisticRegression  (C++ sigmoid, BCE loss, probability predictions)
│   └── KNN Classifier      (C++ k-nearest neighbors classification)
└── Latexify & Presentation Layer
    ├── Formatter           (Matrix, Vector, Model, and Calculus formula to LaTeX string)
    └── Jupyter Hooks       (Rich _repr_latex_ typography in notebooks)

---

Installation

From Source (Recommended for Local Dev)

Requirements: C++17 compatible compiler (g++ or clang++), CMake >= 3.15, Python >= 3.10.

# Clone the repository
git clone https://github.com/MorphosML/MorphosML.git
cd MorphosML

# Build and install in editable mode
pip install -e . --no-build-isolation --no-deps

Pre-Flight Automated Validation Script

MorphosML includes a dedicated CI validation and deployment script:

# Run local compilation, Black, Ruff, and 100% test suite
./scripts/ci_deploy.sh --check-only

---

Quickstart Guide

import numpy as np
import morphosml as mml

# 1. HPC Linear Algebra (Zero-Copy Interoperability)
mat = mml.Matrix.random(1000, 1000, seed=42)
np_arr = np.asarray(mat)  # Zero-copy view! 0 memory allocated

# 2. Calculus & Differentiable Programming
@mml.differentiable
def loss_fn(v):
    x, y = v[0], v[1]
    return (1.0 - x)**2 + 100.0 * (y - x**2)**2

grad = loss_fn.grad([1.0, 1.0])       # [0.0, 0.0]
hess = loss_fn.hessian([1.0, 1.0])    # 2x2 curvature Matrix

# 3. Machine Learning with C++ Core
X = np.random.randn(200, 5)
y = X @ np.array([1.5, -2.0, 0.5, 3.0, -1.0]) + 0.25

X_train, X_test, y_train, y_test = mml.train_test_split(X, y, test_size=0.2, random_state=42)
model = mml.LinearRegression(learning_rate=0.01, epochs=1000)
model.fit(X_train, y_train)

print(f"R2 Score: {model.score(X_test, y_test):.4f}")

# 4. Rich LaTeX Math Export
print(mml.to_latex(model))
# Output: \hat{y} = 1.5 x_{1} - 2 x_{2} + 0.5 x_{3} + 3 x_{4} - 1 x_{5} + 0.25

---

Memory & Linear Algebra Layer

Contiguous 1D Flat Matrix

Unlike naive implementations that allocate vector<vector<double>> (causing cache-line invalidations and pointer chasing), MorphosML allocates a single contiguous buffer vector<double> data_ of length rows * cols.

• Element Access: mat(i, j), mat[i, j], mat.row_ptr(i).
• $i$-$k$-$j$ Loop Ordering: Matrix multiplication uses unit memory strides in the inner loop, unlocking compiler auto-vectorization (AVX2/AVX-512) and 100% cache-line utilization.
• Buffer Protocol: Pass matrices directly to NumPy or Polars via np.asarray(matrix) without duplicating memory.

Non-Owning TensorView

Wraps arbitrary 2D pointers with row stride:

from morphosml import TensorView, tensor_view_from_numpy

# Wrap a NumPy array without copying
arr = np.ones((500, 20), dtype=np.float64)
view = tensor_view_from_numpy(arr)

assert view.rows() == 500
assert view.cols() == 20

---

Out-of-Core Data Ingestion Engine

Native Binary .mldat Format

Parsing text datasets (e.g. CSV) wastes up to 80% of CPU cycles in strtod. MorphosML provides a native binary format (.mldat) with a 64-byte header and FNV-1a checksum:

from morphosml.data import MMapDataset, DataLoader

# 1. Export NumPy array directly to binary .mldat
data = np.random.randn(100000, 64).astype(np.float64)
MMapDataset.dump("dataset_100k.mldat", data)

# 2. Map dataset instantly (0 ms load time)
dataset = MMapDataset("dataset_100k.mldat")
print(f"Dataset Shape: {dataset.shape}, Checksum: {dataset.checksum}")

# 3. Create high-performance, fault-tolerant DataLoader
loader = DataLoader(
    dataset,
    batch_size=64,
    shuffle=True,
    seed=1337,
    prefetch_batches=2  # Asynchronous background double-buffering
)

Idempotency & Fault-Tolerant Checkpointing

# Checkpoint mid-epoch
for i, batch in enumerate(loader):
    if i == 500:
        cursor = loader.get_cursor()
        saved_token = cursor.to_string() # e.g. "0:32000:14695981039346656037"
        break

# Resume from exact sample offset in a new process without re-reading:
new_loader = DataLoader(dataset, batch_size=64, shuffle=True, seed=1337)
new_loader.resume_from(saved_token)

---

Calculus & Differentiable Programming Core

MorphosML provides a high-precision numerical calculus engine implemented in C++:

Numerical Differentiation & Derivatives

import math
from morphosml.calculus import derivative, gradient, jacobian, hessian

# 1st, 2nd, 3rd derivatives with O(h^4) 5-point stencils
f = lambda x: math.sin(x)
print(derivative(f, 0.0, order=1))  # 1.0 (cos(0))
print(derivative(f, 0.0, order=2))  # 0.0 (-sin(0))

# Multivariate Gradient
f_multi = lambda v: v[0]**2 + 3.0 * v[1]**2
print(gradient(f_multi, [2.0, 1.0]))  # [4.0, 6.0]

# Jacobian Matrix
f_vec = lambda v: [v[0]**2 + v[1], 3.0 * v[0] - v[1]**2]
J = jacobian(f_vec, [2.0, 3.0])  # Shape (2, 2)

# Hessian Matrix (Curvature)
H = hessian(f_multi, [1.0, 1.0])  # [[2.0, 0.0], [0.0, 6.0]]

Numerical Integration (Quadrature)

from morphosml.calculus import integrate, integrate_2d

# 1D Integrals: 'simpson' (O(h^4)), 'trapezoidal', 'simpson_38', 'gauss_legendre'
val = integrate(math.sin, 0.0, math.pi, method="simpson")
print(val)  # 2.0

# 2D Double Integral: integral_0^1 integral_0^1 (x + y) dx dy
val_2d = integrate_2d(lambda x, y: x + y, 0.0, 1.0, 0.0, 1.0)
print(val_2d)  # 1.0

Numerical Limits

from morphosml.calculus import limit, evaluate_limit

# Removable singularity: lim_{x -> 0} sin(x)/x = 1.0
print(limit(lambda x: math.sin(x) / x, 0.0))  # 1.0

# One-sided infinite limits
diag = evaluate_limit(lambda x: 1.0 / x, 0.0, direction="right")
print(diag.is_infinite, diag.value)  # True, +inf

---

Machine Learning Models

Model	Core Language	Optimization	Features
LinearRegression	C++17	Gradient Descent	Contiguous memory, $R^2$, MSE cost, fit, predict, score
LogisticRegression	C++17	Gradient Descent + Sigmoid	Binary Cross-Entropy loss, predict_proba, score
KNN	C++17	Vectorized Euclidean distance	Configurable k, single & bulk classification

from morphosml import LogisticRegression

clf = LogisticRegression(learning_rate=0.05, epochs=500)
clf.fit(X_train, y_train)

y_pred = clf.predict(X_test, threshold=0.5)
probabilities = clf.predict_proba(X_test)
print(f"Accuracy: {clf.score(X_test, y_test):.2%}")

---

Latexify & Jupyter Notebook Integration

MorphosML makes presenting and publishing mathematical results effortless:

import morphosml as mml

# LaTeX conversion
mat = mml.Matrix.identity(3)
print(mml.to_latex(mat))
# Output: \begin{pmatrix}1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1\end{pmatrix}

# In Jupyter Notebooks, simply evaluating any Matrix, Vector, or fitted Model
# automatically renders KaTeX math typography via _repr_latex_():
model = mml.LinearRegression().fit(X, y)
mml.latexify(model)

---

Verification & CI/CD Pipeline

The test suite validates correctness across all linear algebra, calculus, data ingestion, and model components:

pytest tests/ -v
# ============================== 61 passed in 0.22s ==============================

Code quality and style are strictly maintained:

• Black Formatter: Targeted for Python 3.12 (line-length = 88).
• Ruff Linter: PEP 8 compliance, zero warnings.
• GitHub Actions Matrix: Automated multi-OS workflows running tests across Python 3.10, 3.11, and 3.12 with Docker volume isolation.

---

License & Author

• Author: Gabriel Carmona
• License: MIT License