why?

1. I want to better learn PyTorch and how it works so what better way than to just re-implement some of its core features.

2. I write mainly in Go and haven't come across a lot of ML support in Go

3. I'd rather have a Go ML service instead of spinning up additional infrastructure to just support a python ML service in my Go projects

4. Go's static typing, native concurrency (avoid GIL problem in python), efficient memory management, single binary deployment and more make it a better interface compared to python IMO

Knowing the PyTorch is mainly implemented in C++ and C under the covers, I'm not expecting any performance gains by porting it to Go. But still, will be interesting to see how it compares.

Two other things:

1. This isn't meant to be go bindings for PyTorch. This is a full rewrite in go from the ground up. Not just a Go abstraction with C bindings.

2. This library only depends on the standard go library with no other dependencies

testing

Im trying to include as many tests as I can to test everything. I'm also testing things side-by-side with pytorch to verify logic. It's not always perfect (for many reasons) but trying to be reasonably close.

roadmap

• Tensor Operations: addition, subtraction, multiplication, division, and more advanced operations like dot products, transposition, and reshaping.

• way to randomly generate tensors

• a way to format a tensor in it's expanded view and not flattened (mainly for sanity checks)

• Automatic Differentiation: basic version of autograd (forward & back propogation)

• Neural Network Layers: some sort of basic neural network layers (fully connected (dense) layers, convolutional layers, and recurrent layers)

• Activation Functions: ReLU, Sigmoid,Tanh and softmax

• Loss Functions: mean squared error (MSE),and binary cross-entropy, categorical cross entropy loss

• Optimizers:Stochastic Gradient Descent (SGD)

• Data Handling: loading, save

• Model Definition and Training Loop: A framework to define models, compile them with loss functions and optimizers, and a training loop to fit models to data. (started but definitely not done)

• A way to sample a model

• probably want a way to manage model state internally to the model

• GPU Support ???????

• Serialization: saving and loading models and their weights. Probably just model state and weights and not the entire model. Reference