Euclid ML is my personal, lightweight machine learning framework built entirely in Python with NumPy and CuPy. It's designed to make neural networks easy to understand, experiment with, and build from the ground up. I've developed several projects using Euclid ML to demonstrate its capabilities and explore what can be achieved with a machine learning framework built from scratch.
Euclid isn't trying to replace PyTorch or TensorFlow.
Quite frankly there is no reason to use it, just use PyTorch.
It was built as a learning project and is written in pure Python with NumPy and CuPy. Building it has taught me a lot about how machine learning frameworks actually work.
TinyGPT is a language model I built using Euclid. The model consists of 12 Transformer blocks, with 12 attention heads per block and a feed-forward dimension of 3,072. It uses a model dimension of 768 and supports a sequence length of 256 tokens.
TinyGPT was trained on the FineWeb-Edu dataset from Hugging Face for 520,000 iterations, using a learning rate of 1 × 10⁻⁴.
The final model contains 134,436,608 parameters.
My neural image compressor uses a convolutional encoder to compress a 2^n×2^n RGB image into a 48x times smaller latent representation.
A convolutional decoder then reconstructs the original image from this compact latent. It was trained on 10,000 ImageNet-10 images for 20 epochs using MSE loss and Adam with a learning rate of 1e-4.
I compare its reconstructions against JPEG at the same 48× compression ratio, with JPEG typically achieving around 2× better MSE.