sahandsafizadeh/qeep: Deep Learning framework in Go with Tensors, AutoGrad, and GPU acceleration


Welcome to qeep (pronounced /kēp/)! This project implements a deep learning framework in Go. It allows you to define neural networks in a declarative way while being able to control operations at tensor level.

  • 📐 Multi-Dimensional Tensors with a wide range of linear algebra and statistical operations.
  • 🔁 Automatic differentiation (AutoGrad) for tensors.
  • ⚡ GPU acceleration via CUDA for high-performance large tensor computations.
  • 🧱 A variety of neural network components, such as fully connected (FC) layer.
  • 📝A declarative API for defining neural networks using stream package.

Navigate to your project directory and clone the qeep repository:

git clone https://github.com/sahandsafizadeh/qeep

Ensure that you have Go installed. Then, Link the local qeep package to your project using Go modules:

go mod edit -require=github.com/sahandsafizadeh/qeep@v0.0.0
go mod edit -replace=github.com/sahandsafizadeh/qeep=./qeep
go mod tidy

💡 This setup assumes that qeep is cloned into a local folder named ./qeep relative to your Go project.

Here is an example of defining a classification model:

package main

import (
	"github.com/sahandsafizadeh/qeep/component/layers"
	"github.com/sahandsafizadeh/qeep/component/layers/activations"
	"github.com/sahandsafizadeh/qeep/component/losses"
	"github.com/sahandsafizadeh/qeep/component/optimizers"
	"github.com/sahandsafizadeh/qeep/model"
	"github.com/sahandsafizadeh/qeep/model/stream"
	"github.com/sahandsafizadeh/qeep/tensor"
)

const dev = tensor.CPU

func prepareModel() (m *model.Model, err error) {
	input := stream.Input()

	x := stream.FC(&layers.FCConfig{Inputs: 4, Outputs: 16, Device: dev})(input)
	x = stream.Relu()(x)
	x = stream.Dropout(&layers.DropoutConfig{Rate: 0.3})(x)

	x = stream.FC(&layers.FCConfig{Inputs: 16, Outputs: 3, Device: dev})(x)
	output := stream.Softmax(&activations.SoftmaxConfig{Dim: 1})(x)

	/* -------------------- */

	loss := losses.NewCE()

	optimizer, err := optimizers.NewAdamW(&optimizers.AdamWConfig{
		LearningRate: 1e-5,
		WeightDecay:  optimizers.AdamWDefaultWeightDecay,
		Beta1:        optimizers.AdamWDefaultBeta1,
		Beta2:        optimizers.AdamWDefaultBeta2,
		Eps:          optimizers.AdamWDefaultEps,
	})
	if err != nil {
		return
	}

	m, err = model.NewModel(input, output, &model.ModelConfig{
		Loss:      loss,
		Optimizer: optimizer,
	})
	if err != nil {
		return
	}

	return m, nil
}

📂 More working examples are provided. You can download their dataset and run them like the following for Iris Classification:

cd ./qeep/examples/03-Iris/
curl -o data.csv https://archive.ics.uci.edu/ml/machine-learning-databases/iris/iris.data
go run .
  1. An accessible CUDA-capable GPU.
  2. CUDA Toolkit installed.
  3. A working C Toolchain like gcc.
  4. CGO enabled and configured (setup guide).

Build the necessary CUDA libraries once:

Now in your Go code, you can set devices to tensor.CUDA.

🔥 Finally, run your program with the cuda build tag:

The qeep project is licensed under the MIT License. See the LICENSE file for details.



Source link