Gorgonia

https://gorgonia.org
Admin
Gorgonia is a machine learning library for Go that facilitates building and executing computation...
Tokens:26,615
Snippets:365
Trust Score:7.1
Update:6 months ago
Show doc for...
Tokens:
Raw
### CUDA Test Program Output Example

Source: https://gorgonia.org/cu

Example output from the `cudatest` program, showing detected CUDA version and device details. This confirms successful CUDA setup.

```text
CUDA version: 10020
CUDA devices: 1

Device 0
========
Name      : "TITAN RTX"
Clock Rate: 1770000 kHz
Memory    : 25393561600 bytes
Compute   : 	7.5


```

--------------------------------

### Install Gorgonia

Source: https://gorgonia.org/getting-started

Use this command to get the Gorgonia library and its dependencies.

```bash
$ go get gorgonia.org/gorgonia
```

--------------------------------

### CLI utility output example

Source: https://gorgonia.org/tutorials/iris

Example output format for the final CLI utility.

```text
./iris
sepal length: 5
sepal width: 3.5
petal length: 1.4
sepal length: 0.2

It is probably a setosa
```

--------------------------------

### Install CUDA Binding and Generate

Source: https://gorgonia.org/tutorials/mnist-cuda

Set GO111MODULE to off, install Gorgonia and the CUDA binding tool, and then generate the CUDA binding for your hardware. Ensure CUDA toolkit is in your PATH.

```bash
~ export GO111MODULE=off
~ go get gorgonia.org/gorgonia
~ export CGO_CFLAGS="-I/usr/local/cuda-10.0/include/"
~ export PATH=$PATH:/usr/local/cuda/bin/
~ go get gorgonia.org/cu
~ go install gorgonia.org/gorgonia/cmd/cudagen
~ $GOPATH/bin/cudagen
```

--------------------------------

### Install and Run CUDA Test Program

Source: https://gorgonia.org/cu

Installs the `cudatest` utility and runs it to verify CUDA installation and device detection. Ensure CUDA Toolkit is installed and configured.

```bash
go install gorgonia.org/cu/cmd/cudatest@latest
cudatest

```

--------------------------------

### Example execution output

Source: https://gorgonia.org/tutorials/hello-world

Shows the expected output when the complete Go program is executed.

```Shell
$ go run main.go
4.5
```

--------------------------------

### Install vecf64 Package

Source: https://gorgonia.org/vecf64

Use this command to install the vecf64 package. It only uses the standard library.

```bash
go get -u gorgonia.org/vecf64
```

--------------------------------

### String and Integer Addition Examples

Source: https://gorgonia.org/dtype

Demonstrates the concept of 'addition' for both string and integer types, illustrating their inclusion in an 'Addable' type class.

```Go
// string "addition".
s := "hello"
s += "world"

// int addition.
i := 1
i += 1

```

--------------------------------

### Generate SVG Graph from Dot Output

Source: https://gorgonia.org/how-to/dot

This command-line example shows how to pipe the output of the Go program into the `dot` command to generate an SVG image of the graph. Make sure Graphviz is installed.

```bash
$ go run main.go | dot -Tsvg > dot-example.svg
```

--------------------------------

### Compute float32 addition with Gorgonia

Source: https://gorgonia.org/reference/vm/gomachine

A complete example demonstrating graph creation, scalar definition, machine execution, and result retrieval.

```go
func main(){
    g := gorgonia.NewGraph()
    forty := gorgonia.F32(40.0)
    two := gorgonia.F32(2.0)
    n1 := gorgonia.NewScalar(g, gorgonia.Float32, gorgonia.WithValue(&forty), gorgonia.WithName("n1"))
    n2 := gorgonia.NewScalar(g, gorgonia.Float32, gorgonia.WithValue(&two), gorgonia.WithName("n2"))

    added, err := gorgonia.Add(n1, n2)
    if err != nil {
        log.Fatal(err)
    }
    machine := NewMachine(g)
    ctx, cancel := context.WithTimeout(context.Background(), 1000*time.Millisecond)
    defer cancel()
    defer machine.Close()
    err = machine.Run(ctx)
    if err != nil {
        log.Fatal(err)
    }
    fmt.Println(machine.GetResult(added.ID()))
}
```

--------------------------------

### Execution Output

Source: https://gorgonia.org/how-to/save-weights

Shows the console output when running the backup/restore example for the first time and subsequent times.

```bash
$  go run main.go
2019/10/28 08:07:26 cannot read backup, doing init open /tmp/example_gorgonia: no such file or directory
⎡0  2⎤
⎣4  6⎦
$  go run main.go
2019/10/28 08:07:29 decoding xT
2019/10/28 08:07:29 decoding yT
⎡0  2⎤
⎣4  6⎦
```

--------------------------------

### Install cudatest

Source: https://gorgonia.org/how-to/troubleshoot-gpu-issues

Installs the `cudatest` utility for troubleshooting CUDA issues. Ensure you have CUDA and cuDNN installed prior to running this command.

```bash
go install gorgonia.org/cu/cmd/cudatest
```

--------------------------------

### Run CUDA-enabled Convnet Example

Source: https://gorgonia.org/tutorials/mnist-cuda

Execute the convnet example with CUDA support enabled using the '-tags="CUDA"' flag. This command times the execution and redirects stderr to null.

```bash
time go run -tags='CUDA'  main.go -epochs 1 2> /dev/null
Epoch 0 599 / 600 [====================================================]  99.83%
```

--------------------------------

### Node Communication Example

Source: https://gorgonia.org/reference/vm/gomachine

Illustrates how nodes in a computation graph are wired together using channels for communication. This example shows sending values between nodes for operations like multiplication and addition.

```go
var aTimesX *node{op: mul}
var aTimesXPlusB *node{op: sum}

var a,b,c gorgonia.Value

aTimesX.inputC <- a
aTimesX.inputC <- x
aTimesXPlusB.inputC <- <- aTimesX.outputC 
aTimesXPlusB.inputC <- <- b
```

--------------------------------

### Install cudagen Tool

Source: https://gorgonia.org/reference/cuda

Install the `cudagen` utility, which is necessary for generating CUDA-related code for Gorgonia. This tool is part of the Gorgonia project.

```bash
go install gorgonia.org/gorgonia/cmd/cudagen
```

--------------------------------

### Build Gorgonia with CUDA Support

Source: https://gorgonia.org/reference/cuda

To enable CUDA support, build your Go application using the 'cuda' build tag. Ensure the CUDA toolkit is installed and `cudagen` is available.

```bash
go build -tags='cuda' .
```

--------------------------------

### Primitive Shape Construction Examples

Source: https://gorgonia.org/shapes

Demonstrates the primitive ways to construct shapes according to the BNF, including unit, integer-based, and variable-based shapes.

```text
()  // unit
(10,) // integer-based shape
(8, (10,)) // nested integer-based shape
((8, (10,)), 20) // deeply nested integer-based shape

```

--------------------------------

### Run MNIST Example

Source: https://gorgonia.org/tutorials/mnist

Executes the main program to output the label vector for the processed image.

```bash
$ go run main.go
[0.1 0.1 0.1 0.1 0.1 0.9 0.1 0.1 0.1 0.1]
```

--------------------------------

### Computation output

Source: https://gorgonia.org/reference/vm/gomachine

The expected output of the float32 addition example.

```text
42
```

--------------------------------

### Implement state function context handling

Source: https://gorgonia.org/reference/vm/gomachine

Example of handling context cancellation within a state function.

```go
func mystate(ctx context.Context, *node) stateFn { 
    // ...
    select {
        // ...
        case <- ctx.Done():
            n.err = ctx.Error()
            return nil
    }
}
```

--------------------------------

### Tensor Data Type Mismatch Example

Source: https://gorgonia.org/tensor

This example demonstrates a panic that occurs when attempting to set data of an incorrect type to a tensor. Ensure the data type passed to `SetAt` matches the tensor's underlying data type.

```go
b := New(WithBacking(Range(Float32, 0, 24)), WithShape(2, 3, 4))
x, _ := b.At(0, 1, 2)
fmt.Printf("x: %v\n", x)

// Setting data
b.SetAt(1000, 0, 1, 2)
fmt.Printf("b:\n%v", b)
```

--------------------------------

### Evolving Shape Constraints

Source: https://gorgonia.org/shapes

Example of applying a shape expression to a known input to evolve constraints.

```text
(a, b) → (b, c) → (a, c) @ (2, 3)
```

```text
(3, c) → (2, c)
```

--------------------------------

### Helper Function to Get User Input

Source: https://gorgonia.org/tutorials/iris

A utility function to prompt the user for input, read a line from stdin, and parse it as a float64. Handles newline characters and potential parsing errors.

```go
func getInput(s string) float64 {
        reader := bufio.NewReader(os.Stdin)
        fmt.Printf("%v: ", s)
        text, _ := reader.ReadString('\n')
        text = strings.Replace(text, "\n", "", -1)

        input, err := strconv.ParseFloat(text, 64)
        if err != nil {
                log.Fatal(err)
        }
        return input
}
```

--------------------------------

### Initialize Training Components

Source: https://gorgonia.org/tutorials/mnist

Sets up the VM, solver, and training loop structure.

```go
vm := gorgonia.NewTapeMachine(g, gorgonia.BindDualValues(m.learnables()...))
solver := gorgonia.NewRMSPropSolver(gorgonia.WithBatchSize(float64(bs)))
defer vm.Close()
```

```go
batches := numExamples / bs
```

```go
for i := 0; i < *epochs; i++ {
    for b := 0; b < batches; b++ {
        // ...
    }
}
```

--------------------------------

### Initialize Tape Machine

Source: https://gorgonia.org/tutorials/iris

Sets up a Tape Machine to execute the graph and compute gradients. It's crucial to bind dual values for parameters that require gradient tracking, like 'theta'.

```go
machine := gorgonia.NewTapeMachine(g, gorgonia.BindDualValues(theta))
defer machine.Close()
```

--------------------------------

### Initialize Vanilla Solver

Source: https://gorgonia.org/tutorials/iris

Creates a Vanilla Solver instance for gradient descent with a specified learning rate. Use this to update model parameters iteratively.

```go
solver := gorgonia.NewVanillaSolver(gorgonia.WithLearnRate(0.001))
```

--------------------------------

### Basic Graph Creation and Execution in Gorgonia

Source: https://gorgonia.org/gorgonia

Demonstrates how to define a simple mathematical expression z = x + y as a computation graph, compile it into a program, and execute it using a TapeMachine. Requires manual value setting with Let.

```go
package gorgonia_test

import (
	"fmt"
	"log"

	. "gorgonia.org/gorgonia"
)

// Basic example of representing mathematical equations as graphs.
//
// In this example, we want to represent the following equation
// 		z = x + y
func Example_basic() {
	g := NewGraph()

	var x, y, z *Node
	var err error

	// define the expression
	x = NewScalar(g, Float64, WithName("x"))
	y = NewScalar(g, Float64, WithName("y"))
	if z, err = Add(x, y); err != nil {
		log.Fatal(err)
	}

	// create a VM to run the program on

machine := NewTapeMachine(g)
	defer machine.Close()

	// set initial values then run
	Let(x, 2.0)
	Let(y, 2.5)
	if err = machine.RunAll(); err != nil {
		log.Fatal(err)
	}

	fmt.Printf("%v", z.Value())
	// Output: 4.5
}

```

--------------------------------

### Simple Computation with Gorgonia

Source: https://gorgonia.org/getting-started

This program demonstrates creating a computation graph, defining scalar nodes, performing addition, and executing the computation using a tape machine. Ensure you have the necessary imports. If you encounter a GC-related panic, set the environment variable `ASSUME_NO_MOVING_GC_UNSAFE_RISK_IT_WITH=go1.19`.

```go
package main

import (
        "fmt"
        "log"

        "gorgonia.org/gorgonia"
)

func main() {
        g := gorgonia.NewGraph()

        var x, y, z *gorgonia.Node
        var err error

        // define the expression
        x = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("x"))
        y = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("y"))
        if z, err = gorgonia.Add(x, y); err != nil {
                log.Fatal(err)
        }

        // create a VM to run the program on
        machine := gorgonia.NewTapeMachine(g)
        defer machine.Close()

        // set initial values then run
        gorgonia.Let(x, 2.0)
        gorgonia.Let(y, 2.5)
        if err = machine.RunAll(); err != nil {
                log.Fatal(err)
        }

        fmt.Printf("%v", z.Value())
}
```

--------------------------------

### Instantiate and run TapeMachine

Source: https://gorgonia.org/tutorials/hello-world

Creates a TapeMachine to execute the computation graph and runs all operations. Ensures the machine is closed afterwards using defer.

```Go
machine := gorgonia.NewTapeMachine(g)
defer machine.Close()

if err = machine.RunAll(); err != nil {
        log.Fatal(err)
}
```

--------------------------------

### Gorgonia Graph Addition

Source: https://gorgonia.org/about/computation-graph

Demonstrates creating a computation graph, defining nodes, performing addition, and executing the graph using a TapeMachine.

```go
func main() {
	// Create a graph.
	g := gorgonia.NewGraph()

	// Create a node called "x" with the value 1.
	x := gorgonia.NodeFromAny(g, 1, gorgonia.WithName("x"))

	// Create a node called "y" with the value 1.
	y := gorgonia.NodeFromAny(g, 1, gorgonia.WithName("y"))

	// z := x + y
	z := gorgonia.Must(gorgonia.Add(x, y))

	// Create a VM to execute the graph.
	vm := gorgonia.NewTapeMachine(g)

	// Run the VM. Errors are not checked.
	vm.RunAll()

	// Print the value of z.
	fmt.Printf("%v", z.Value())
}
```

--------------------------------

### Build with AVX Support

Source: https://gorgonia.org/vecf64

To enable SIMD acceleration using AVX instructions, use this build command.

```bash
go build -tags='avx' ...
```

--------------------------------

### Execute Training Step

Source: https://gorgonia.org/tutorials/mnist

Extracts batch data, assigns values to the graph, and runs the solver.

```go
var xVal, yVal tensor.Tensor
xVal, _ = inputs.Slice(sli{start, end})

yVal, _ = targets.Slice(sli{start, end})

xVal.(*tensor.Dense).Reshape(bs, 1, 28, 28)
```

```go
gorgonia.Let(x, xVal)
gorgonia.Let(y, yVal)
```

```go
vm.RunAll()
solver.Step(gorgonia.NodesToValueGrads(m.learnables()))
vm.Reset()
```

--------------------------------

### Build with SSE Support

Source: https://gorgonia.org/vecf64

To enable Single Instruction, Multiple Data (SIMD) acceleration using SSE instructions, use this build command.

```bash
go build -tags='sse' ...
```

--------------------------------

### Create CUDA Shared Object Symlink on Linux

Source: https://gorgonia.org/cu

Creates a symbolic link for `libcuda.so` on Linux systems, typically required when the CUDA shared object is installed in a non-standard location or not automatically linked by `ld`. Adjust paths as necessary.

```bash
sudo ln -s /PATH/TO/libcuda.so.1 /PATH/TO/libcuda.so

```

--------------------------------

### Create a new Gorgonia graph

Source: https://gorgonia.org/tutorials/hello-world

Initializes an empty expression graph. This is the first step in building any computation with Gorgonia.

```Go
g := gorgonia.NewGraph()
```

--------------------------------

### Initialize Convolutional Network Weights

Source: https://gorgonia.org/tutorials/mnist

Creates the network structure and initializes learnable parameters using Glorot initialization.

```go
// Note: gorgonia is abbreviated G in this example for clarity
func newConvNet(g *G.ExprGraph) *convnet {
	w0 := G.NewTensor(g, dt, 4, G.WithShape(32, 1, 3, 3), G.WithName("w0"), G.WithInit(G.GlorotN(1.0)))
	w1 := G.NewTensor(g, dt, 4, G.WithShape(64, 32, 3, 3), G.WithName("w1"), G.WithInit(G.GlorotN(1.0)))
	w2 := G.NewTensor(g, dt, 4, G.WithShape(128, 64, 3, 3), G.WithName("w2"), G.WithInit(G.GlorotN(1.0)))
	w3 := G.NewMatrix(g, dt, G.WithShape(128*3*3, 625), G.WithName("w3"), G.WithInit(G.GlorotN(1.0)))
	w4 := G.NewMatrix(g, dt, G.WithShape(625, 10), G.WithName("w4"), G.WithInit(G.GlorotN(1.0)))
	return &convnet{
		g:  g,
		w0: w0,
		w1: w1,
		w2: w2,
		w3: w3,
		w4: w4,

		d0: 0.2,
		d1: 0.2,
		d2: 0.2,
		d3: 0.55,
	}
}
```

--------------------------------

### Create a (2,2) Dense Tensor

Source: https://gorgonia.org/tensor

Demonstrates how to create a 2x2 dense tensor with integer backing data. Ensure the backing data matches the shape.

```go
a := New(WithShape(2, 2), WithBacking([]int{1, 2, 3, 4}))
fmt.Printf("a:\n%v\n", a)
```

--------------------------------

### Instantiate Neural Network Graph

Source: https://gorgonia.org/tutorials/mnist

Initializes the computation graph and defines input/output nodes.

```go
g := gorgonia.NewGraph()
x := gorgonia.NewTensor(g, dt, 4, gorgonia.WithShape(bs, 1, 28, 28), gorgonia.WithName("x"))
y := gorgonia.NewMatrix(g, dt, gorgonia.WithShape(bs, 10), gorgonia.WithName("y"))
m := newConvNet(g)
m.fwd(x)
```

--------------------------------

### Standard Go Addition

Source: https://gorgonia.org/about/computation-graph

A basic Go program performing addition without using the Gorgonia library.

```go
func main() {
	fmt.Printf("%v", 1+1)
}
```

--------------------------------

### Define the VM interface

Source: https://gorgonia.org/reference/vm

The VM interface defines the core methods required for executing and managing computational resources in Gorgonia.

```go
type VM interface {
    RunAll() error
    Reset()

    // Close closes all the machine resources (CUDA, if any, loggers if any)
    Close() error
}
```

--------------------------------

### Run Prediction Loop in CLI

Source: https://gorgonia.org/tutorials/iris

Initializes a tape machine, sets up input processing, and runs the prediction loop. It continuously takes input, performs computation, and displays the predicted iris species. Ensure `getInput` function is defined.

```go
machine := gorgonia.NewTapeMachine(g)
values[4] = 1.0
for {
        values[0] = getInput("sepal length")
        values[1] = getInput("sepal width")
        values[2] = getInput("petal length")
        values[3] = getInput("petal width")

        if err = machine.RunAll(); err != nil {
                log.Fatal(err)
        }
        switch math.Round(y.Value().Data().(float64)) {
        case 1:
                fmt.Println("It is probably a setosa")
        case 2:
                fmt.Println("It is probably a virginica")
        case 3:
                fmt.Println("It is probably a versicolor")
        default:
                fmt.Println("unknown iris")
        }
        machine.Reset()
}
```

--------------------------------

### Execute with TapeMachine

Source: https://gorgonia.org/how-to/autodiff

Run the graph using the TapeMachine and access gradients via nodes or the Grad method.

```go
machine := gorgonia.NewTapeMachine(g)
defer machine.Close()
if err = machine.RunAll(); err != nil {
        log.Fatal(err)
}

fmt.Printf("result: %v\n", result.Value())
if zgrad, err := z.Grad(); err == nil {
        fmt.Printf("dz/dx: %v | %v\n", zgrad, grads[0].Value())
}

if xgrad, err := x.Grad(); err == nil {
        fmt.Printf("dz/dx: %v | %v\n", xgrad, grads[1].Value())
}

if ygrad, err := y.Grad(); err == nil {
        fmt.Printf("dz/dy: %v | %v\n", ygrad, grads[2].Value())
}
```

--------------------------------

### Create an expression graph

Source: https://gorgonia.org/how-to/autodiff

Initialize a new graph and define the mathematical expression f(x,y,z)=(x+y)×z.

```go
g := gorgonia.NewGraph()

var x, y, z *gorgonia.Node
var err error

// define the expression
x = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("x"))
y = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("y"))
z = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("z"))
q, err := gorgonia.Add(x, y)
if err != nil {
    log.Fatal(err)
}
result, err := gorgonia.Mul(z, q)
if err != nil {
    log.Fatal(err)
}
```

--------------------------------

### Create Gorgonia Expression Graph

Source: https://gorgonia.org/tutorials/iris

Initializes a Gorgonia graph and defines nodes for input data (x, y) and model parameters (theta). Use this to represent your model's structure.

```go
func getXY() (*tensor.Dense, *tensor.Dense) {
	x, y := getXYMat()

	xT := tensor.FromMat64(x)
	yT := tensor.FromMat64(y)
	// Get rid of the last dimension to create a vector
	s := yT.Shape()
	yT.Reshape(s[0])
	return xT, yT
}

func main() {
	xT, yT := getXY()
	g := gorgonia.NewGraph()
	x := gorgonia.NodeFromAny(g, xT, gorgonia.WithName("x"))
	y := gorgonia.NodeFromAny(g, yT, gorgonia.WithName("y"))
	heta := gorgonia.NewVector(
		g,
		gorgonia.Float64,
		gorgonia.WithName("theta"),
		gorgonia.WithShape(xT.Shape()[1]),
		gorgonia.WithInit(gorgonia.Uniform(0, 1)))

	pred := must(gorgonia.Mul(x, theta))
    // Saving the value for later use
    var predicted gorgonia.Value
    gorgonia.Read(pred, &predicted)
```

--------------------------------

### Numerical Stability Comparison

Source: https://gorgonia.org/about/computation-graph

Compares standard math.Log with math.Log1p to demonstrate numerical stability issues with floating-point arithmetic.

```go
func main() {
	fmt.Printf("%v\n", math.Log(1.0+10e-16))
	fmt.Printf("%v\n", math.Log1p(10e-16))
}
```

--------------------------------

### Print values and gradients

Source: https://gorgonia.org/how-to/autodiff

Display the results and partial derivatives after graph execution.

```go
fmt.Printf("x=%v;y=%v;z=%v\n", x.Value(), y.Value(), z.Value())
fmt.Printf("f(x,y,z) = %v\n", result.Value())

if xgrad, err := x.Grad(); err == nil {
    fmt.Printf("df/dx: %v\n", xgrad)
}
if ygrad, err := y.Grad(); err == nil {
    fmt.Printf("df/dy: %v\n", ygrad)
}
if xgrad, err := z.Grad(); err == nil {
    fmt.Printf("df/dx: %v\n", xgrad)
}
```

--------------------------------

### Run Gradient Descent Iterations

Source: https://gorgonia.org/tutorials/iris

Executes the graph, updates model parameters using the solver, and resets the machine for the next iteration. This loop performs the core gradient descent training.

```go
iter := 1000000
        var err error
        for i := 0; i < iter; i++ {
                if err = machine.RunAll(); err != nil {
                        fmt.Printf("Error during iteration: %v: %v\n", i, err)
                        break
                }

                if err = solver.Step(model); err != nil {
                        log.Fatal(err)
                }
                machine.Reset() // Reset is necessary in a loop like this
        }
```

--------------------------------

### Generic Tensor Allocation with String Backing

Source: https://gorgonia.org/tensor

Demonstrates generic tensor allocation using a backing slice of strings and specifying a shape. The *Dense structure allows for runtime type flexibility.

```go
x := New(WithBacking([]string{"hello", "world", "hello", "world"}), WithShape(2,2))
```

--------------------------------

### Display Directory Structure

Source: https://gorgonia.org/tutorials/mnist

Expected file layout for the MNIST test data directory.

```bash
$ ls -alhg *
-rw-r--r--  1 staff   375B Nov 11 13:48 main.go

testdata:
total 107344
drwxr-xr-x  6 staff   192B Nov 11 13:48 .
drwxr-xr-x  4 staff   128B Nov 11 13:48 ..
-rw-r--r--  1 staff   7.5M Jul 21  2000 t10k-images.idx3-ubyte
-rw-r--r--  1 staff   9.8K Jul 21  2000 t10k-labels.idx1-ubyte
-rw-r--r--  1 staff    45M Jul 21  2000 train-images.idx3-ubyte
-rw-r--r--  1 staff    59K Jul 21  2000 train-labels.idx1-ubyte
```

--------------------------------

### Prepare Model Parameters for Solver

Source: https://gorgonia.org/tutorials/iris

Defines the model parameters (theta) that the solver will update. The solver operates on ValueGrad interfaces, and *Node fulfills this.

```go
model := []gorgonia.ValueGrad{theta}
```

--------------------------------

### Access and Set Tensor Data

Source: https://gorgonia.org/tensor

Illustrates how to access a specific element in a tensor using `At` and how to modify it using `SetAt`. Note that dimensions are 0-indexed.

```go
b := New(WithBacking(Range(Float32, 0, 24)), WithShape(2, 3, 4))
x, _ := b.At(0, 1, 2)
fmt.Printf("x: %v\n", x)

b.SetAt(float32(1000), 0, 1, 2)
fmt.Printf("b:\n%v", b)
```

--------------------------------

### Fork and Clone Gorgonia CUDA Repository

Source: https://gorgonia.org/cu

Steps to fork the Gorgonia CUDA repository on GitHub, clone it locally, and prepare for making contributions. This involves standard Git workflow.

```bash
git clone git@github.com:YOUR_USERNAME/cu.git

```

--------------------------------

### Marshal ExprGraph to Dot Language

Source: https://gorgonia.org/how-to/dot

Use this Go code to create a simple graph and marshal it into the dot language format. Ensure you have the Gorgonia library imported.

```go
package main

import (
        "fmt"
        "log"

        "gorgonia.org/gorgonia"
        "gorgonia.org/gorgonia/encoding/dot"
)

func main() {
        g := gorgonia.NewGraph()

        var x, y *gorgonia.Node

        // define the expression
        x = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("x"))
        y = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("y"))
        gorgonia.Add(x, y)
        b, err := dot.Marshal(g)
        if err != nil {
                log.Fatal(err)
        }
        fmt.Println(string(b))
}
```

--------------------------------

### Perform Automatic Differentiation with Gorgonia

Source: https://gorgonia.org/how-to/autodiff

Uses LispMachine to execute forward and backward mode differentiation on a scalar expression.

```go
func main() {
	g := gorgonia.NewGraph()

	var x, y, z *gorgonia.Node
	var err error

	// define the expression
	x = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("x"))
	y = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("y"))
	z = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("z"))
	q, err := gorgonia.Add(x, y)
	if err != nil {
		log.Fatal(err)
	}
	result, err := gorgonia.Mul(z, q)
	if err != nil {
		log.Fatal(err)
	}

	// set initial values then run
	gorgonia.Let(x, -2.0)
	gorgonia.Let(y, 5.0)
	gorgonia.Let(z, -4.0)

	// by default, lispmachine performs forward mode and backwards mode execution
	m := gorgonia.NewLispMachine(g)
	defer m.Close()
	if err = m.RunAll(); err != nil {
		log.fatal(err)
	}

	fmt.Printf("x=%v;y=%v;z=%v\n", x.Value(), y.Value(), z.Value())
	fmt.Printf("f(x,y,z)=(x+y)*z\n")
	fmt.Printf("f(x,y,z) = %v\n", result.Value())

	if xgrad, err := x.Grad(); err == nil {
		fmt.Printf("df/dx: %v\n", xgrad)
	}

	if ygrad, err := y.Grad(); err == nil {
		fmt.Printf("df/dy: %v\n", ygrad)
	}
	if xgrad, err := z.Grad(); err == nil {
		fmt.Printf("df/dz: %v\n", xgrad)
	}
}
```

```bash
$ go run main.go
x=-2;y=5;z=-4
f(x,y,z)=(x+y)*z
f(x,y,z) = -12
df/dx: -4
df/dy: -4
df/dz: 3
```

--------------------------------

### Visualize MNIST Image

Source: https://gorgonia.org/tutorials/mnist

Loads MNIST data, processes the first element into a grayscale image, and saves it as a PNG file.

```go
import (
        //...
        "image"
        "image/png"

        "gorgonia.org/gorgonia/examples/mnist"
        "gorgonia.org/tensor"
        "gorgonia.org/tensor/native"
)

func main() {
        inputs, targets, err := mnist.Load("train", "./testdata", tensor.Float64)
        if err != nil {
                log.Fatal(err)
        }
        cols := inputs.Shape()[1]
        imageBackend := make([]uint8, cols)
        for i := 0; i < cols; i++ {
                v, _ := inputs.At(0, i)
                imageBackend[i] = uint8((v.(float64) - 0.1) * 0.9 * 255)
        }
        img := &image.Gray{
                Pix:    imageBackend,
                Stride: 28,
                Rect:   image.Rect(0, 0, 28, 28),
        }
        w, _ := os.Create("output.png")
        vals, _ := native.MatrixF64(targets.(*tensor.Dense))
        fmt.Println(vals[0])
        err = png.Encode(w, img)
}
```

--------------------------------

### VM Interface Definition

Source: https://gorgonia.org/reference/vm

Defines the core interface for a Virtual Machine (VM) in Gorgonia, outlining the methods for computation and resource management.

```APIDOC
## VM Interface

### Description
A VM in Gorgonia is an object that understands the expression graph and has implemented the ability to perform computation with it. Technically, it is an `interface{}` with three methods.

### Methods

- **RunAll() error**: Executes all pending operations in the VM.
- **Reset()**: Resets the VM to its initial state.
- **Close() error**: Closes all machine resources (e.g., CUDA, loggers).

### Available VM Implementations

- Go Machine
- LispMachine
- Tapemachine

These VMs function differently and may take different inputs.
```

--------------------------------

### Perform Symbolic Differentiation with Gorgonia

Source: https://gorgonia.org/how-to/autodiff

Uses TapeMachine to compute gradients symbolically for a defined graph expression.

```go
func main() {
	g := gorgonia.NewGraph()

	var x, y, z *gorgonia.Node
	var err error

	// define the expression
	x = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("x"))
	y = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("y"))
	z = gorgonia.NewScalar(g, gorgonia.Float64, gorgonia.WithName("z"))
	q, err := gorgonia.Add(x, y)
	if err != nil {
		log.Fatal(err)
	}
	result, err := gorgonia.Mul(z, q)
	if err != nil {
		log.Fatal(err)
	}

	if grads, err = Grad(result,z, x, y); err != nil {
		log.Fatal(err)
	}

	// set initial values then run
	gorgonia.Let(x, -2.0)
	gorgonia.Let(y, 5.0)
	gorgonia.Let(z, -4.0)

	machine := gorgonia.NewTapeMachine(g)
	defer machine.Close()
	if err = machine.RunAll(); err != nil {
		log.Fatal(err)
	}

	fmt.Printf("x=%v;y=%v;z=%v\n", x.Value(), y.Value(), z.Value())
	fmt.Printf("f(x,y,z)=(x+y)*z\n")
	fmt.Printf("f(x,y,z) = %v\n", result.Value())

	if zgrad, err := z.Grad(); err == nil {
		fmt.Printf("dz/dx: %v | %v\n", zgrad, grads[0].Value())
	}

	if xgrad, err := x.Grad(); err == nil {
		fmt.Printf("dz/dx: %v | %v\n", xgrad, grads[1].Value())
	}

	if ygrad, err := y.Grad(); err == nil {
		fmt.Printf("dz/dy: %v | %v\n", ygrad, grads[2].Value())
	}
}
```

```bash
$ go run main.go
x=-2;y=5;z=-4
f(x,y,z)=(x+y)*z
f(x,y,z) = -12
df/dx: -4 | -4
df/dy: -4 | -4
df/dz: 3 | 3
```

--------------------------------

### Perform Backpropagation

Source: https://gorgonia.org/tutorials/mnist

Executes symbolic backpropagation and defines the learnable nodes.

```go
gorgonia.Grad(cost, m.learnables()...)
```

```go
func (m *convnet) learnables() gorgonia.Nodes {
    return gorgonia.Nodes{m.w0, m.w1, m.w2, m.w3, m.w4}
}
```

--------------------------------

### Log Training Progress

Source: https://gorgonia.org/tutorials/iris

Prints the current value of theta, iteration count, cost, and accuracy during training. This helps monitor the learning process.

```go
fmt.Printf("theta: %2.2f  Iter: %v Cost: %2.3f Accuracy: %2.2f \r",
        theta.Value(),
        i,
        cost.Value(),
        accuracy(predicted.Data().([]float64), y.Value().Data().([]float64)))
```

--------------------------------

### Assign values to input nodes

Source: https://gorgonia.org/tutorials/hello-world

Uses the `Let` function to assign concrete values (2.0 for x, 2.5 for y) to the previously defined input nodes.

```Go
gorgonia.Let(x, 2.0)
gorgonia.Let(y, 2.5)
```

--------------------------------

### Create a new Hugo content page

Source: https://gorgonia.org/getting-started/contributing-doc

Use this Hugo command to create a new Markdown file for documentation content. Ensure you specify the correct directory based on the content type.

```bash
hugo new content/about/mypage.md
```

--------------------------------

### PubSub Run Method Signature

Source: https://gorgonia.org/reference/vm/gomachine

Signature for the `run` method of the `pubsub` struct, which initiates the broadcast and merge operations for all registered publishers and subscribers.

```go
func (p *pubsub) run(ctx context.Context) (context.CancelFunc, *sync.WaitGroup) { ... }
```

--------------------------------

### Run automatic differentiation with LispMachine

Source: https://gorgonia.org/how-to/autodiff

Execute the graph using the LispMachine to automatically compute gradients.

```go
m := gorgonia.NewLispMachine(g)
defer m.Close()
if err = m.RunAll(); err != nil {
    log.fatal(err)
}
```

--------------------------------

### Visualize Dataset with Gonum Plotter

Source: https://gorgonia.org/tutorials/iris

Generates a PNG image visualizing dataset points, plotting sepal length against sepal width, categorized by species. Requires `gonum.org/v1/plot` libraries. The output is returned as a byte slice and also saved to 'out.png'.

```go
import (
    "gonum.org/v1/plot"
    "gonum.org/v1/plot/plotter"
    "gonum.org/v1/plot/plotutil"
    "gonum.org/v1/plot/vg"
    "gonum.org/v1/plot/vg/draw"
)

func plotData(x []float64, a []float64) []byte {
	p, err := plot.New()
	if err != nil {
		log.Fatal(err)
	}

	p.Title.Text = "sepal length & width"
	p.X.Label.Text = "length"
	p.Y.Label.Text = "width"
	p.Add(plotter.NewGrid())

	l := len(x) / len(a)
	for k := 1; k <= 3; k++ {
		data0 := make(plotter.XYs, 0)
		for i := 0; i < len(a); i++ {
			if k != int(a[i]) {
				continue
			}
			x1 := x[i*l+0] // sepal_length
			y1 := x[i*l+1] // sepal_width
			data0 = append(data0, plotter.XY{X: x1, Y: y1})
		}
		data, err := plotter.NewScatter(data0)
		if err != nil {
			log.Fatal(err)
		}
		data.GlyphStyle.Color = plotutil.Color(k - 1)
		data.Shape = &draw.PyramidGlyph{}
		p.Add(data)
		p.Legend.Add(fmt.Sprint(k), data)
	}

	w, err := p.WriterTo(4*vg.Inch, 4*vg.Inch, "png")
	if err != nil {
		panic(err)
	}
	var b bytes.Buffer
	writer := bufio.NewWriter(&b)
	w.WriteTo(writer)
	ioutil.WriteFile("out.png", b.Bytes(), 0644)
	return b.Bytes()
}
```

--------------------------------

### Benchmark Comparison: CUDA vs. Non-CUDA

Source: https://gorgonia.org/reference/cuda

This benchmark demonstrates the performance difference when using CUDA for operations like 'sigmoid' compared to standard Go implementations. Note that the CUDA kernel used is a basic implementation.

```go
BenchmarkOneMilCUDA-8   	     300	   3348711 ns/op
BenchmarkOneMil-8       	      50	  33169036 ns/op
```

--------------------------------

### Create Gorgonia Model for Prediction

Source: https://gorgonia.org/tutorials/iris

Constructs the computational graph (exprgraph) for predictions using loaded weights and input nodes. This code assumes weights have been loaded.

```go
g := gorgonia.NewGraph()
theta := gorgonia.NodeFromAny(g, thetaT, gorgonia.WithName("theta"))
values := make([]float64, 5)
xT := tensor.New(tensor.WithBacking(values))
x := gorgonia.NodeFromAny(g, xT, gorgonia.WithName("x"))
y, err := gorgonia.Mul(x, theta)
```

--------------------------------

### Define Tensor Preparation Functions

Source: https://gorgonia.org/tutorials/mnist

Functions for converting raw MNIST data into Gorgonia tensors.

```go
func prepareX(M []RawImage, dt tensor.Dtype) (retVal tensor.Tensor)

func prepareY(N []Label, dt tensor.Dtype) (retVal tensor.Tensor)
```

--------------------------------

### Verify CUDA device with cudatest

Source: https://gorgonia.org/how-to/troubleshoot-gpu-issues

Executes `cudatest` with a specified GPU to verify CUDA initialization and retrieve device information. This confirms that the selected GPU is recognized and functional by CUDA.

```bash
$ CUDA_VISIBLE_DEVICES=0 cudatest
CUDA version: 11000
CUDA devices: 1
Device 0
========
Name      :     "Tesla K20Xm"
Clock Rate:     732000 kHz
Memory    :     5977800704 bytes
Compute   :     3.5
```

--------------------------------

### Execute state machine

Source: https://gorgonia.org/reference/vm/gomachine

Method to run the node state machine until completion.

```go
func (n *node) Compute(ctx context.Context) error {
	for state := defaultState; state != nil; {
		state = state(ctx, n)
	}
	return n.err
}
```

--------------------------------

### PubSub Structure Definition

Source: https://gorgonia.org/reference/vm/gomachine

Defines the `pubsub` struct, which aggregates publishers and subscribers and is responsible for setting up the channel network for communication.

```go
type pubsub struct {
	publishers  []*publisher
	subscribers []*subscriber
}
```

--------------------------------

### Identify GPUs with nvidia-smi

Source: https://gorgonia.org/how-to/troubleshoot-gpu-issues

Use `nvidia-smi` to list all detected GPUs and their CUDA compatibility. This helps in identifying which GPUs can be used with CUDA.

```bash
Thu Jul 16 17:41:10 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.05    Driver Version: 450.51.05    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla K20Xm         On   | 00000000:06:00.0 Off |                    0 |
| N/A   33C    P8    16W / 235W |      0MiB /  5700MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GT 1030     On   | 00000000:07:00.0  On |                  N/A |
| 35%   33C    P0    N/A /  30W |    656MiB /  1994MiB |     51%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    1   N/A  N/A      XXXX      G   /usr/lib/xorg/Xorg                270MiB |
|    1   N/A  N/A      XXXX      G   /usr/bin/PROGRAMNAME               77MiB |
|    1   N/A  N/A      XXXX      G   /usr/bin/PROGRAMNAME               68MiB |
|    1   N/A  N/A      XXXX      G   ...AAAAAAAAA= --shared-files      221MiB |
+-----------------------------------------------------------------------------+
```

--------------------------------

### Monitor CUDA Usage with nvidia-smi

Source: https://gorgonia.org/tutorials/mnist-cuda

Run this command in a separate window to monitor GPU utilization and memory usage while the CUDA application is running. It displays driver and CUDA versions.

```bash
~  nvidia-smi
Sun Feb 16 22:05:15 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.87.00    Driver Version: 418.87.00    CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------|
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla M60           On   | 00000000:00:1E.0 Off |                    0 |
| N/A   54C    P0    73W / 150W |    841MiB /  7618MiB |     76%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0     18614      C   /tmp/go-build629284435/b001/exe/main         372MiB |
+-----------------------------------------------------------------------------+
```

--------------------------------

### Create a (2,3,4) 3-Tensor

Source: https://gorgonia.org/tensor

Shows how to create a 3-dimensional tensor with float32 data. The `Range` function is used to populate the backing data.

```go
b := New(WithBacking(Range(Float32, 0, 24)), WithShape(2, 3, 4))
fmt.Printf("b:\n%1.1f\n", b)
```

--------------------------------

### List AWS Deep Learning AMIs

Source: https://gorgonia.org/tutorials/mnist-cuda

Use this AWS CLI command to find AMIs with the CUDA toolkit pre-installed. These are tested on g3s.xlarge instances.

```bash
~ aws ec2 describe-images --owners amazon --filters 'Name=state,Values=available' 'Name=name,Values=Deep Learning AMI (Ubuntu)*' --query 'sort_by(Images, &CreationDate)[].Name'
```

--------------------------------

### Define state machine actions

Source: https://gorgonia.org/reference/vm/gomachine

Required state functions for node operations.

```go
func defaultState(context.Context, *node) stateFn { ... }

func receiveInput(context.Context, *node) stateFn { ... }

func computeFwd(context.Context, *node) stateFn { ... }

func emitOutput(context.Context, *node) stateFn { ... }
```

--------------------------------

### Implement Matrix Interface Wrapper

Source: https://gorgonia.org/how-to/dataframe

Defines a wrapper struct to allow a gota DataFrame to satisfy the gonum Matrix interface.

```go
type matrix struct {
	dataframe.DataFrame
}

func (m matrix) At(i, j int) float64 {
	return m.Elem(i, j).Float()
}

func (m matrix) T() mat.Matrix {
	return mat.Transpose{Matrix: m}
}
```
Tokens:
Raw