NanoNet

A lightweight neural-network framework built from scratch with Python and NumPy, designed around transparent internals and built-in observability.

NanoNet started as a way to understand what frameworks such as PyTorch are actually doing behind calls like:

loss.backward()
optimizer.step()

Instead of wrapping an existing machine-learning library, NanoNet implements the important pieces directly — tensors, reverse-mode autodiff, layers, optimizers, training utilities — and adds first-class tools to inspect models, trace execution, explore autograd graphs, and diagnose numerical issues:

model.inspect()
model.trace(x)
output.graph()
model.diagnose(x)

The goal is not to compete with PyTorch on performance.

The goal is to make neural-network training mechanics small enough to read, understand, and debug.

NanoNet is currently pre-1.0; public APIs may evolve as the framework matures.

---

Installation

NanoNet is the project name. The PyPI distribution is nanonet-ml, and the Python import package is nanonet_ml.

Install from PyPI:

pip install nanonet-ml

import nanonet_ml as nn

The core library only requires NumPy.

From source

git clone https://github.com/ariv-chaudhry/NanoNet.git
cd NanoNet
python -m venv .venv

Activate the virtual environment:

# Linux / macOS
source .venv/bin/activate

# Windows PowerShell
.\.venv\Scripts\Activate.ps1

python -m pip install -e ".[dev]"

See RELEASING.md for the release process.

---

Quickstart

import nanonet_ml as nn
import numpy as np

nn.manual_seed(0)

model = nn.Sequential(
    nn.Linear(4, 8),
    nn.ReLU(),
    nn.Linear(8, 2),
)

x = nn.Tensor(np.random.randn(8, 4))
y = model(x)

model.inspect()
model.trace(x)
y.graph()
model.diagnose(x)

A familiar training-style workflow looks like:

import nanonet_ml as nn

nn.manual_seed(42)

model = nn.Sequential(
    nn.Dense(784, 128),
    nn.ReLU(),
    nn.Dropout(0.2),
    nn.Dense(128, 64),
    nn.ReLU(),
    nn.Dense(64, 10),
)

optimizer = nn.Adam(model.parameters(), lr=0.001)
loss_fn = nn.CrossEntropyLoss()

model.fit(
    X_train,
    y_train,
    loss_fn=loss_fn,
    optimizer=optimizer,
    epochs=10,
    batch_size=64,
)

accuracy = model.evaluate(X_test, y_test)
print(accuracy)

The API is intentionally familiar if you've used frameworks such as PyTorch or Keras, but the underlying implementation is NanoNet's own.

---

Public API

Core
  Tensor, no_grad, manual_seed, __version__

Modules
  Module, Parameter, Sequential

Layers
  Dense, Linear, Dropout, Flatten
  ReLU, Sigmoid, Tanh, Softmax

Optimization
  SGD, Adam, Optimizer

Losses
  MSELoss, CrossEntropyLoss

Observability (methods)
  Module.inspect, Module.trace, Module.diagnose
  Tensor.graph

Data (`nanonet_ml.data`)
  Dataset, TensorDataset, DataLoader, LogDataset
  load_mnist, download_mnist

Linear is an alias of Dense. Report types live under nanonet_ml.inspection for programmatic use; everyday workflows use the methods above.

---

How It Works

A neural-network training step has two main stages.

Forward Pass

Input → Dense → ReLU → Dense → Loss

NanoNet calculates the output while recording the operations that produced it.

Backward Pass

Input ← Dense ← ReLU ← Dense ← Loss

Calling:

loss.backward()

walks the computational graph in reverse and applies the chain rule to calculate gradients.

Those gradients are then used by an optimizer:

optimizer.step()

to update the trainable parameters.

---

Features

NanoNet currently includes:

Automatic Differentiation

• reverse-mode automatic differentiation
• computational graph construction
• topological backward traversal
• gradient accumulation
• repeated backward passes
• NumPy-style broadcasting gradients
• numerical gradient checking

Tensor Operations

• addition
• subtraction
• multiplication
• division
• powers
• negation
• sum
• mean
• reshape
• transpose
• indexing
• exponential
• logarithm
• element-wise maximum
• NumPy-style matmul

Matrix multiplication supports:

• vector @ vector
• matrix @ vector
• vector @ matrix
• matrix @ matrix
• batched matrix multiplication
• broadcast batch dimensions

Neural-Network Components

• Parameter
• Module
• Sequential
• Dense
• ReLU
• Sigmoid
• Tanh
• Softmax
• Dropout
• Flatten

Loss Functions

• Mean Squared Error
• numerically stable Cross Entropy

Optimizers

• SGD
• momentum
• L2 weight decay
• Adam

NanoNet's Adam weight decay is implemented as coupled L2 regularization, not AdamW-style decoupled weight decay.

Data

NanoNet provides a general data pipeline:

Dataset  →  DataLoader  →  NanoNet training

Concrete sources include:

• Dataset protocol (__len__ / __getitem__)
• TensorDataset for in-memory arrays
• mini-batch DataLoader (shuffle, deterministic seeding, NumPy collation)
• LogDataset for parser-driven line-oriented log files
• configurable log encoding and optional blank-line filtering
• contextual parsing diagnostics (file + physical line number)
• MNIST downloading and caching

LogDataset does not interpret log semantics automatically — you supply a parser that turns each line into features or (features, target). See docs/data.md and examples/log_anomaly_detection.py.

from nanonet_ml.data import DataLoader, LogDataset

def parse_log(line: str):
    level, status, latency = line.split()
    levels = {"INFO": 0.0, "WARNING": 1.0, "ERROR": 2.0}
    return [levels[level], float(status), float(latency)]

dataset = LogDataset(
    "server.log",
    parser=parse_log,
    skip_blank_lines=True,
)

loader = DataLoader(dataset, batch_size=32, shuffle=True)
for batch in loader:
    ...

Parsers may also return supervised samples:

def parse_log(line: str):
    ...
    return features, label

Training

• reusable Trainer
• model.fit(...)
• model.evaluate(...)
• validation metrics
• training history
• plotting support
• no_grad() inference

Model Utilities

• parameter counting
• model summaries
• state_dict()
• .npz serialization
• JSON parameter metadata

---

Why I Built NanoNet

Modern machine-learning libraries make it possible to build a neural network without knowing much about what happens internally.

For example:

loss.backward()

looks simple, but underneath it requires:

computational graphs
topological sorting
the chain rule
gradient accumulation
broadcast-gradient reduction
matrix derivatives

Similarly:

optimizer.step()

requires algorithms such as SGD or Adam to convert those gradients into parameter updates.

NanoNet was my way of implementing those mechanics myself instead of only using them through another framework.

---

Automatic Differentiation

The core of NanoNet is its automatic-differentiation engine.

Consider:

from nanonet_ml import Tensor

x = Tensor(
    3.0,
    requires_grad=True,
)

y = x**2 + 2*x

y.backward()

print(x.grad)

Output:

8.0

Mathematically:

y = x² + 2x

dy/dx = 2x + 2

At:

x = 3

we get:

dy/dx = 8

NanoNet creates the computational graph from the operations used to construct y, then traverses that graph backward to calculate the derivative.

---

Branching Graphs

A Tensor can contribute to the output through multiple paths.

For example:

x = Tensor(
    2.0,
    requires_grad=True,
)

y = x*x + 3*x

y.backward()

print(x.grad)

Output:

7.0

because:

dy/dx = 2x + 3

and:

2(2) + 3 = 7

NanoNet automatically combines the gradients from both paths.

---

Repeated Backward Calls

Gradients accumulate across calls to:

backward()

For example:

x = Tensor(
    2.0,
    requires_grad=True,
)

y = x * x

y.backward()

print(x.grad)

produces:

4.0

Calling:

y.backward()

again gives:

8.0

Each backward traversal propagates only the gradient generated by that particular call.

Previously accumulated .grad values are stored for the user but are not incorrectly propagated through the graph again.

---

Broadcasting

NanoNet supports NumPy-style broadcasting.

For example:

X: (32, 128)
b: (128,)

allows:

Y = X + b

to produce:

(32, 128)

During backward, NanoNet reduces the bias gradient back to:

(128,)

rather than incorrectly leaving it as:

(32, 128)

This is required for operations such as Dense-layer bias addition.

---

Matrix Multiplication

NanoNet's @ operator follows NumPy-style matrix multiplication.

Supported forms include:

vector @ vector
matrix @ vector
vector @ matrix
matrix @ matrix
batched matrix @ matrix
batched matrix @ batched matrix

For example:

from nanonet_ml import Tensor

a = Tensor(
    [1.0, 2.0, 3.0],
    requires_grad=True,
)

b = Tensor(
    [4.0, 5.0, 6.0],
    requires_grad=True,
)

y = a @ b

y.backward()

Both vectors receive the correct gradients.

The same autograd operation also supports the matrix multiplication required by Dense layers:

X @ W

---

Neural Networks

Models can be assembled with Sequential.

from nanonet_ml import Sequential
from nanonet_ml.layers import Dense, ReLU

model = Sequential([
    Dense(784, 128),
    ReLU(),
    Dense(128, 64),
    ReLU(),
    Dense(64, 10),
])

NanoNet automatically tracks Parameters contained within each layer.

You can inspect the model with:

model.summary(
    input_shape=(784,),
)

and count trainable parameters with:

print(
    model.num_parameters()
)

---

Training

NanoNet includes a reusable training loop.

from nanonet_ml.training import Trainer

trainer = Trainer(model)

history = trainer.fit(
    X_train,
    y_train,
    loss_fn=loss_fn,
    optimizer=optimizer,
    epochs=10,
    batch_size=64,
    validation_data=(
        X_test,
        y_test,
    ),
)

The same functionality is exposed through:

model.fit(...)

A training iteration performs:

forward
  ↓
loss
  ↓
clear gradients
  ↓
backward
  ↓
optimizer step

---

Inference and no_grad()

Training mode and gradient recording are separate concepts.

To disable layer behavior such as Dropout:

model.eval()

To disable autograd graph construction:

from nanonet_ml import no_grad

with no_grad():
    predictions = model(inputs)

model.evaluate(...) automatically uses both evaluation mode and no_grad().

This avoids building unnecessary computational graphs during inference.

---

Cross Entropy

Classification models should pass raw logits directly to:

CrossEntropyLoss()

For example:

logits = model(inputs)

loss = loss_fn(
    logits,
    targets,
)

Do not apply Softmax before Cross Entropy.

NanoNet combines log-softmax and negative log likelihood using a numerically stable log-sum-exp implementation.

Targets must contain valid integer class labels.

For example:

0
1
2

are valid.

A fractional target such as:

1.5

raises an error instead of being silently converted to class 1.

---

Optimizers

NanoNet includes SGD and Adam implementations written directly with NumPy.

SGD

Basic SGD performs:

θ ← θ − η∇L

Momentum can also be enabled.

Adam

Adam tracks exponential moving averages of:

gradients

and:

squared gradients

and uses them to adapt each parameter's update.

NanoNet's Adam weight_decay option is coupled L2 regularization:

g ← ∇L + λθ

The adjusted gradient then enters Adam's moment calculations.

This is different from AdamW, where weight decay is applied separately from the gradient-based Adam update.

See:

docs/optimizers.md

for the full derivation.

---

Gradient Clearing

Gradients accumulate unless they are cleared.

A typical training step therefore uses:

optimizer.zero_grad()

predictions = model(inputs)
loss = loss_fn(
    predictions,
    targets,
)

loss.backward()

optimizer.step()

You can also clear gradients with:

model.zero_grad()

NanoNet represents a cleared gradient as:

None

rather than allocating an array of zeros.

---

MNIST

NanoNet includes a fully connected MNIST example.

For a quick smoke test:

python examples/mnist_mlp.py \
    --epochs 1 \
    --train-limit 5000

During development, the model reached:

88.15% test accuracy

after one epoch using 5,000 training examples.

For a longer run:

python examples/mnist_mlp.py \
    --epochs 10 \
    --batch-size 64 \
    --lr 0.001

Results vary depending on:

• initialization
• random seed
• training configuration
• number of epochs

The downloaded MNIST files are cached under:

data/mnist/

and are not stored in the Git repository.

---

Nonlinear Regression

NanoNet can also train regression models.

The regression example approximates a nonlinear function using a small multilayer perceptron.

Run:

python examples/regression.py

---

Gradient Checking

One of the easiest ways to introduce bugs into an autodiff engine is to write a backward derivative that looks plausible but is slightly wrong.

NanoNet therefore includes numerical gradient checking.

Example:

from nanonet_ml import Tensor
from nanonet_ml.gradcheck import gradcheck

a = Tensor(
    [1.5, -2.0],
    requires_grad=True,
)

b = Tensor(
    [0.5, 3.0],
    requires_grad=True,
)

result = gradcheck(
    lambda x, y: (x * y).sum(),
    [a, b],
)

print(
    result.passed
)

print(
    result.max_abs_error
)

print(
    result.max_rel_error
)

The numerical derivative is calculated using central finite differences and compared against NanoNet's analytical gradient.

Gradient checks cover operations including:

multiplication
matrix multiplication
vector matrix multiplication
batched matrix multiplication
small neural-network graphs
unused differentiable inputs

---

Model Saving and Loading

Models can be saved using:

model.save(
    "checkpoints/mnist"
)

NanoNet automatically normalizes the path and creates:

checkpoints/mnist.npz
checkpoints/mnist.npz.meta.json

You can then restore it using:

model.load(
    "checkpoints/mnist"
)

Using the suffix explicitly also works:

model.save(
    "checkpoints/mnist.npz"
)

NanoNet stores parameters using NumPy .npz files and a JSON metadata sidecar.

It does not rely on Python pickle for model parameters.

---

Model Summary

NanoNet can print layer and parameter information:

model.summary(
    input_shape=(784,),
)

Example structure:

Layer                        Output Shape         Parameters
------------------------------------------------------------
Dense(784,128)               ('?', 128)             ...
ReLU                         ('?', 128)               0
Dense(128,64)                ('?', 64)              ...
ReLU                         ('?', 64)                0
Dense(64,10)                 ('?', 10)              ...
------------------------------------------------------------
Total parameters:                                   ...

Shape inference is performed without constructing an autograd graph.

---

Observability

NanoNet makes neural-network internals observable through model inspection, execution tracing, autograd graph inspection, and evidence-based diagnostics.

API	Purpose
model.inspect()	Model structure, parameters, shapes, and statistics
model.trace(x)	Actual module execution order for an input
tensor.graph()	Autograd operation / dependency graph
model.diagnose(x)	Numerical and optimization warning checks

All four APIs share the same conventions:

• print by default; suppress with verbose=False
• return a structured report object
• support print(report) via __str__
• support report.to_dict() for JSON-compatible metadata

model.inspect(x)

trace = model.trace(x)

prediction = model(x)
loss = criterion(prediction, target)
loss.graph()

loss.backward()
model.diagnose(x)

Model Inspection

report = model.inspect()
report = model.inspect(x)  # runtime shapes / activations

Execution Tracing

trace = model.trace(x, verbose=False)
for step in trace.steps:
    print(step.module_name, step.outputs[0].shape)

Autograd stays enabled so trace.output can participate in backward(). Timings include instrumentation overhead (debugging only).

Computation Graphs

graph = loss.graph(verbose=False)
print(graph.root_id, [op.name for op in graph.operations])

Exposes lower-level autograd ops (for example MatMul/Add inside Dense), not merely module names. Graph IDs are local to each graph() call.

Diagnostics

report = model.diagnose(x, verbose=False)
for finding in report.findings:
    if finding.severity != "info":
        print(finding.code, finding.message)

Never calls backward(). NaN/Inf checks are definitive; vanishing gradients, dead ReLU, and saturation are conservative heuristics (DiagnosticThresholds).

Detailed notes: docs/observability.md. Examples: examples/observability_workflow.py and the focused scripts under examples/.

---

Benchmarks

NanoNet is not designed to outperform PyTorch.

PyTorch benefits from:

• optimized C/C++ kernels
• optimized BLAS implementations
• sophisticated memory management
• GPU acceleration
• years of production optimization

NanoNet intentionally prioritizes readability.

The repository includes equivalent synthetic benchmark workloads for NanoNet and PyTorch.

Run:

python benchmarks/benchmark_nanonet.py

and, if PyTorch is installed:

python benchmarks/benchmark_pytorch.py

Then compare them using multi-trial methodology:

python benchmarks/compare.py --samples 5000 --runs 5

Both frameworks are timed on CPU with float64. Because synthetic labels are random, these scripts measure performance rather than model quality.

For mathematical validation and MNIST learning comparison, see Empirical Evaluation below.

---

Empirical Evaluation

NanoNet has been compared against PyTorch for:

• forward-pass numerical agreement
• loss agreement
• gradient agreement
• SGD update agreement
• multi-trial training/inference runtime
• workload scaling
• matched MNIST learning performance

Full methodology, environment metadata, interpretation, and limitations: docs/evaluation.md.

Snapshot results (measured, CPU, float64)

Evaluation	Result
Forward max abs. error	2.22e-16
Loss abs. error	2.22e-16
Gradient max abs. error	1.11e-16
Post-SGD parameter max error	8.67e-19
NanoNet MNIST accuracy (5k/1k, 1 epoch, SGD)	66.50%
PyTorch MNIST accuracy (matched)	66.50%
Train slowdown @ 5k samples (mean of 5 runs)	~3.1×
Infer slowdown @ 256-batch (mean of 5 runs)	~9.5×

Runtime values are machine-dependent (recorded on Windows 11 / Intel CPU; see JSON under results/).

pip install -e ".[benchmark]"
python benchmarks/numerical_parity.py
python benchmarks/scaling_benchmark.py --sizes 1000 5000 10000 --runs 5
python benchmarks/mnist_comparison.py
python benchmarks/run_evaluation.py --quick   # or --full

---

Testing

Run the complete test suite with:

pytest -v

Run coverage with:

pytest \
    --cov=nanonet_ml \
    --cov-report=term-missing

Run Ruff with:

ruff check nanonet_ml tests examples benchmarks

NanoNet's tests cover areas such as:

• Tensor operations and reverse-mode autodiff
• neural-network modules, layers, and activations
• losses and optimizers
• gradient checking and model serialization
• training workflows and graph-free evaluation
• Dataset / TensorDataset / DataLoader behavior
• LogDataset parsing, encodings, blank-line handling, and diagnostics
• DataLoader batching with log-backed samples
• end-to-end log anomaly-detection smoke coverage

---

Examples

Several examples are included.

Automatic Differentiation

python examples/autodiff_demo.py

XOR

python examples/xor.py

Regression

python examples/regression.py

MNIST

python examples/mnist_mlp.py

Log anomaly classification

End-to-end parser-driven log anomaly classification using LogDataset:

python examples/log_anomaly_detection.py

Model Inspection

python examples/model_inspection.py

Execution Tracing

python examples/execution_trace.py

Computation Graph Inspection

python examples/computation_graph.py

Model Diagnostics

python examples/model_diagnostics.py

Observability Workflow

python examples/observability_workflow.py

---

Project Structure

NanoNet/
├── nanonet_ml/
│   ├── __init__.py
│   ├── tensor.py
│   ├── autograd.py
│   ├── gradcheck.py
│   ├── serialization.py
│   ├── utils.py
│   │
│   ├── nn/
│   ├── layers/
│   ├── losses/
│   ├── optimizers/
│   ├── data/
│   ├── metrics/
│   ├── training/
│   └── inspection/
│
├── examples/
│   ├── autodiff_demo.py
│   ├── xor.py
│   ├── regression.py
│   ├── mnist_mlp.py
│   ├── log_anomaly_detection.py
│   ├── data/
│   ├── model_inspection.py
│   ├── execution_trace.py
│   ├── computation_graph.py
│   ├── model_diagnostics.py
│   └── observability_workflow.py
│
├── benchmarks/
│   ├── benchmark_nanonet.py
│   ├── benchmark_pytorch.py
│   ├── compare.py
│   └── observability_overhead.py
│
├── tests/
├── docs/
├── scripts/
├── README.md
├── LICENSE
├── pyproject.toml
└── .gitignore

---

Architecture

The main relationship between NanoNet's components is:

Tensor
  │
  ▼
Automatic Differentiation
  │
  ▼
Parameter
  │
  ▼
Module / Layer
  │
  ▼
Sequential Model
  │
  ▼
Loss
  │
  ▼
Backward
  │
  ▼
Optimizer

Data and training utilities surround this core:

Dataset
   │
   ▼
DataLoader
   │
   ▼
Trainer
   │
   ├──── Model
   ├──── Loss
   └──── Optimizer

More detailed explanations are available in:

• docs/architecture.md
• docs/autodiff.md
• docs/backpropagation.md
• docs/data.md
• docs/optimizers.md

---

Limitations

NanoNet is intentionally small.

Current limitations include:

• CPU / NumPy computation only
• no CUDA backend
• no convolutional layers
• no pooling layers
• no batch normalization
• no multiprocessing DataLoader
• no distributed training
• no mixed-precision training
• no explicit computational-graph freeing API

The project currently focuses on fully connected neural networks and the mechanics behind their training.

These limitations are deliberate.

Adding fewer features with understandable implementations is more valuable to the project's goal than attempting to recreate all of PyTorch.

---

Roadmap

Possible future additions include:

• Conv2D
• MaxPool2D
• BatchNorm
• learning-rate schedulers
• additional datasets
• additional examples
• optional GPU experiments
• mixed-precision experiments

The goal is to add features when they introduce an interesting implementation or mathematical concept rather than simply increasing the feature count.

---

License

NanoNet is source-available, not open source.

The source code is publicly accessible for viewing, educational reference, and evaluation purposes. Public availability does not grant permission to copy, modify, redistribute, republish, sublicense, sell, or incorporate substantial portions of the project into another project without prior written permission.

Copyright © 2026 Ariv Chaudhry. All rights reserved.

See LICENSE for the complete license terms.