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 - abineshmathivanan31@gmail.com
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.
git clone https://github.com/abinesh-mathivanan/go-torch.git
cd go-torchgo mod tidyrun the mnist training file to test out the features.
go run ./cnn_benchmark/go_bench.go| Benchmark Detail | 128x128 | 512x512 | 1024x1024 |
|---|---|---|---|
| Matrix Multiply | 510.33 µs | 13.54 ms | 130.50 ms |
| Element-wise Add | 71.72 µs | 1.29 ms | 4.13 ms |
| Element-wise Mul | 47.83 µs | 1.63 ms | 3.91 ms |
| ReLU Activation | 121.18 µs | 1.75 ms | 6.45 ms |
| Linear Layer Forward (B32,I128,O10) | 71.93 µs | ||
| CrossEntropyLoss (B32,C10) | 11.16 µs | ||
| Full Fwd-Bwd (Net:128-256-10, B32) | 4.02 ms |