Connect the maths to code.
Trace backward(), inspect tensor operations, and change a layer or optimizer. Make the implementation part of your understanding.
Follow your first gradientWHY EVERYO / FOR THE DEVELOPER WHO ASKS WHY
Build the model.
Understand the mechanics.
EveryO is a readable Python framework for learning how neural networks work. Use its NumPy-based stack to trace a gradient, teach a concept, or run a small experiment you can explain.
THE SHORT ANSWER
Why should I
use EveryO?
Use EveryO when understanding the implementation is part of the goal. Follow a tensor through the forward pass, inspect how gradients are calculated, and see how an optimizer updates the parameters.
You get a complete learning workflow in Python: data, models, training, charts, and saved experiments. Start on CPU, read the source as you go, and change one thing at a time.
See what the framework includes01 / WHO EVERYO IS FOR
A different starting point.
The same curiosity.
Choose the path that matches what you want to understand or build.
Make the lesson executable.
Use seeded, locally generated datasets to connect a lecture or notebook to runnable Python examples. Installation downloads dependencies; experiments can then run locally.
Choose a teaching exampleKeep the experiment inspectable.
Train small CPU models, compare architectures, inspect learning curves, and save a model for the next run.
Build a complete experiment02 / WHAT YOU CAN USE IT FOR
Turn a concept
into a working experiment.
These are practical paths through the existing framework and its example scripts.
01
02
03
04
05
Understand automatic differentiation
Compute the derivative of x², then follow the graph traversal and gradient accumulation in the source.
Train and evaluate a classifier
Generate data, split it, train a network, and inspect held-out results and learning curves.
Explore what a CNN learns
Compare a convolutional network with a dense baseline and inspect learned filters.
Compare recurrence and attention
Try RNN, LSTM, GRU, and transformer experiments on a long-range recall task.
Save and revisit your work
Store a supported model in a pickle-free .evo archive, restore it, and run inference.
03 / WHAT YOU GAIN WHILE USING IT
More than a prediction.
A clearer explanation.
Each part of the workflow gives you something concrete to inspect.
| When you… | You can inspect… | That helps you… |
|---|---|---|
| Call backward() | The recorded operations and accumulated gradients | Connect the chain rule to its implementation |
| Compose a model | Layer shapes, parameters, and the forward pass | Reason about how inputs become predictions |
| Run a training loop | Loss, evaluation metrics, callbacks, and history | Understand the effect of a model or optimizer change |
| Read numerical tests | Finite-difference checks and optional reference comparisons | Evaluate evidence for numerical correctness |
| Change an example | Seeds, data, code, and output in one local workflow | Explain and reproduce a small experiment |
04 / CHOOSE IT FOR THE RIGHT JOB
Use it to learn.
Know where it fits.
EveryO is built for learning, teaching, numerical inspection, and small-scale experiments.
Already using another framework?
Keep your existing workflow and use EveryO to study the mechanics in a smaller NumPy implementation. It can be a companion for understanding concepts you use elsewhere.
Planning production or large-scale training?
Use a mature framework and validate your deployment stack. EveryO does not claim production readiness, GPU performance leadership, or validated multi-machine training.
Read the validation limits05 / DIRECT ANSWERS
Good questions.
Clear starting points.
The essentials before your first experiment.
What is EveryO?
EveryO is an open-source Python neural computing framework built from scratch on NumPy. It implements tensors, automatic differentiation, neural networks, optimizers, and training for learning, teaching, and small-scale experiments.
Who should use EveryO?
EveryO is for developers learning neural-network internals, educators teaching numerical computing, and experimenters who want readable implementations and small, reproducible CPU workflows.
Why use EveryO if I already use PyTorch or JAX?
Use EveryO as a learning companion when you want to trace how gradients, layers, and parameter updates are implemented. Keep using your established framework for large-scale training, optimized accelerator execution, or its broader model ecosystem.
What can I build with EveryO?
You can build small classifiers, regression models, convolutional networks, recurrent networks, and transformer experiments. The repository includes runnable examples with locally generated datasets, visualization, and model saving.
What do I gain while using EveryO?
You can connect mathematical concepts to readable Python, inspect intermediate tensors and gradients, compare numerical results, and modify small experiments. Those workflows help you understand why a model behaves as it does.
Do I need a GPU, an account, or an API key?
No GPU, account, or API key is required for the core NumPy CPU workflow. Install the Python dependencies first; the bundled datasets are generated locally. CUDA, TensorFlow reference checks, and ONNX export are optional integrations.
Is EveryO suitable for production or large-scale training?
EveryO is designed for learning, teaching, and small-scale experimentation. It is not a production training runtime. GPU runtime behavior needs hardware validation, multi-machine training remains unverified, and supported ONNX exports must be checked with the target runtime.