Deep learning framework built from the ground up in pure Go

go-torch is an open-source deep learning framework built from the ground up in pure Go. It provides a modular, PyTorch-like API for building and training neural networks with a stable auto-differentiation engine.

mail - [email protected]

blog - https://abinesh-mathivanan.vercel.app/en/posts/post-5/

• dynamic computation graph: tensors track their history, allowing for automatic gradient calculation during the backward pass.
• extensible module system (nn.Layer, nn.Sequential): build complex model architectures with a flexible, Keras-like sequential API.
• layer and function library: includes Conv2D, Linear, MaxPooling2D, Flatten, ReLU, CrossEntropyLoss, and SGD
• real-time TUI dashboard: live graphs for batch-wise loss and epoch-wise validation accuracy, monitoring of memory usage (Heap/Total Alloc), GC cycles, and active goroutines along with keras-like summary.
• optimized performance: using BLAS, go-routines and topological autograd + grad accumulation

TUI Dashboard

• add support for RNN, LSTM, Transformers
• implement Adam with Ga-lore and LORA techniques, RMSProp etc...
• model.load() and model.save() without gob
• support building Transformers

• Go 1.18 or later.
• system-installed BLAS library is recommended for maximum performance but not required.
• some todo's are written inside the files. use 'better comments' extension for best experience.

run the mnist training file to test out the features.