feat: refactor to proper app + write tests

This commit is contained in:
Stevan Freeborn
2025-02-13 17:24:38 -06:00
parent baf4035616
commit 42435ad915
38 changed files with 1462 additions and 253 deletions
+1 -2
View File
@@ -4,6 +4,5 @@ internal interface IImage
{
int Width { get; }
int Height { get; }
void Resize(int width, int height);
IPixel GetPixel(int x, int y);
Stream Data { get; }
}
@@ -0,0 +1,6 @@
namespace BGR.Console.Removal;
internal interface IInferenceRunner
{
ITensor<float> Run(byte[] model, ITensor<float> inputTensor);
}
-8
View File
@@ -1,8 +0,0 @@
namespace BGR.Console.Removal;
internal interface IPixel
{
float R { get; }
float G { get; }
float B { get; }
}
+2 -1
View File
@@ -6,4 +6,5 @@ internal interface ITensor<T>
int Width { get; }
void SetValue(int batch, int channel, int y, int x, T value);
float GetValue(int batch, int channel, int y, int x);
}
Tensor<T> ToTensor();
}
+5 -1
View File
@@ -2,12 +2,16 @@ namespace BGR.Console.Removal;
internal abstract class ImageProcessor
{
public abstract Task<IImage> LoadImageAsync(string path);
public abstract Task<ITensor<float>> CreateTensorInputAsync(Stream image, Model model);
public abstract Task<Stream> GenerateMaskAsync(OnnxTensor maskTensor, int width, int height);
public abstract Task<Stream> GenerateMaskAsync(ITensor<float> maskTensor, int width, int height);
public abstract Task<Stream> RemoveBackgroundAsync(Stream image, Stream mask);
public abstract Task SaveImageAsync(Stream image, string path);
protected static void WalkImage(int height, int width, Action<int, int> action)
{
for (var y = 0; y < height; y++)
@@ -2,9 +2,26 @@ namespace BGR.Console.Removal.ImageSharp;
internal class ImageSharpProcessor : ImageProcessor
{
public override async Task<IImage> LoadImageAsync(string path)
{
var image = await Image.LoadAsync<Rgba32>(path);
if (image.Metadata.DecodedImageFormat is null)
{
throw new InvalidOperationException("Image format is not supported.");
}
var stream = new MemoryStream();
await image.SaveAsync(stream, image.Metadata.DecodedImageFormat);
stream.Position = 0;
return new SharpImage(image.Width, image.Height, stream);
}
public override async Task<ITensor<float>> CreateTensorInputAsync(Stream image, Model model)
{
using var resized = await Image.LoadAsync<Rgba32>(image);
resized.Mutate(x => x.Resize(model.InputWidth, model.InputHeight));
const int batchSize = 1;
@@ -22,7 +39,7 @@ internal class ImageSharpProcessor : ImageProcessor
return tensor;
}
public override async Task<Stream> GenerateMaskAsync(OnnxTensor maskTensor, int width, int height)
public override async Task<Stream> GenerateMaskAsync(ITensor<float> maskTensor, int width, int height)
{
using var mask = new Image<Rgba32>(width, height);
@@ -45,12 +62,16 @@ internal class ImageSharpProcessor : ImageProcessor
var stream = new MemoryStream();
await mask.SaveAsync(stream, new PngEncoder());
stream.Position = 0;
return stream;
}
public override async Task<Stream> RemoveBackgroundAsync(Stream image, Stream mask)
{
image.Position = 0;
mask.Position = 0;
var imageWithBg = await Image.LoadAsync<Rgba32>(image);
var maskImage = await Image.LoadAsync<Rgba32>(mask);
using var imageWithBgRemoved = new Image<Rgba32>(imageWithBg.Width, imageWithBg.Height);
@@ -72,9 +93,18 @@ internal class ImageSharpProcessor : ImageProcessor
var result = new MemoryStream();
await imageWithBgRemoved.SaveAsync(result, new PngEncoder());
result.Position = 0;
return result;
}
public override async Task SaveImageAsync(Stream image, string path)
{
image.Position = 0;
var img = await Image.LoadAsync<Rgba32>(image);
await img.SaveAsync(path, new PngEncoder());
}
private static float Normalize(float value)
{
const float binarizationThreshold = 0.5f;
@@ -1,19 +1,25 @@
namespace BGR.Console.Removal.ImageSharp;
internal class SharpImage(Image<Rgba32> image) : IImage
internal class SharpImage : IImage
{
private readonly Image<Rgba32> _image = image;
public int Width { get; }
public int Height { get; }
public Stream Data { get; }
public int Width => _image.Width;
public int Height => _image.Height;
public void Resize(int width, int height)
public SharpImage(int width, int height, Stream data)
{
_image.Mutate(x => x.Resize(width, height));
}
if (width <= 0)
{
throw new ArgumentOutOfRangeException(nameof(width), "must be greater than 0");
}
public IPixel GetPixel(int x, int y)
{
return new SharpPixel(_image[x, y]);
if (height <= 0)
{
throw new ArgumentOutOfRangeException(nameof(height), "must be greater than 0");
}
Width = width;
Height = height;
Data = data ?? throw new ArgumentNullException(nameof(data));
}
}
@@ -1,10 +0,0 @@
namespace BGR.Console.Removal.ImageSharp;
internal class SharpPixel(Rgba32 pixel) : IPixel
{
private readonly Rgba32 _pixel = pixel;
public float R => _pixel.R;
public float G => _pixel.G;
public float B => _pixel.B;
}
@@ -0,0 +1,6 @@
namespace BGR.Console.Removal.Models;
internal interface IModelFactory
{
Model Create(string resourceName);
}
@@ -1,7 +1,8 @@
namespace BGR.Console.Removal.Models;
internal class ModNetModel : Model
internal class ModNetModel(byte[] modelBytes) : Model(modelBytes)
{
public const string Id = "modnet";
public override int InputWidth => 512;
public override int InputHeight => 512;
public override float RedNormalizationMean => 0.485f;
+13 -3
View File
@@ -11,18 +11,28 @@ internal abstract class Model
public abstract float RedNormalizationStd { get; }
public abstract float GreenNormalizationStd { get; }
public abstract float BlueNormalizationStd { get; }
public byte[] Bytes { get; } = [];
public float NormalizeRed(float value)
internal Model()
{
}
protected Model(byte[] modelBytes)
{
Bytes = modelBytes;
}
public virtual float NormalizeRed(float value)
{
return Normalize(value, RedNormalizationMean, RedNormalizationStd);
}
public float NormalizeGreen(float value)
public virtual float NormalizeGreen(float value)
{
return Normalize(value, GreenNormalizationMean, GreenNormalizationStd);
}
public float NormalizeBlue(float value)
public virtual float NormalizeBlue(float value)
{
return Normalize(value, BlueNormalizationMean, BlueNormalizationStd);
}
@@ -0,0 +1,21 @@
namespace BGR.Console.Removal.Models;
internal class ModelFactory(IResourceManager resourceManager) : IModelFactory
{
private readonly IResourceManager _resourceManager = resourceManager;
public Model Create(string resourceName)
{
var resource = _resourceManager.GetResource(resourceName);
var model = new byte[resource.Length];
resource.ReadExactly(model);
return resourceName switch
{
$"{U2NetModel.Id}.onnx" => new U2NetModel(model),
$"{RmbgModel.Id}.onnx" => new RmbgModel(model),
$"{ModNetModel.Id}.onnx" => new ModNetModel(model),
_ => throw new ArgumentException($"Unknown model name: {resourceName}")
};
}
}
+2 -2
View File
@@ -1,8 +1,8 @@
namespace BGR.Console.Removal.Models;
internal class RmbgModel : Model
internal class RmbgModel(byte[] modelBytes) : Model(modelBytes)
{
public const string Id = "rmbg";
public override int InputWidth => 1024;
public override int InputHeight => 1024;
public override float RedNormalizationMean => 0.485f;
+3 -2
View File
@@ -1,7 +1,8 @@
namespace BGR.Console.Removal.Models;
internal class U2NetModel : Model
internal class U2NetModel(byte[] modelBytes) : Model(modelBytes)
{
public const string Id = "u2net";
public override int InputWidth => 320;
public override int InputHeight => 320;
public override float RedNormalizationMean => 0.485f;
@@ -10,4 +11,4 @@ internal class U2NetModel : Model
public override float RedNormalizationStd => 0.229f;
public override float GreenNormalizationStd => 0.224f;
public override float BlueNormalizationStd => 0.225f;
}
}
@@ -0,0 +1,18 @@
namespace BGR.Console.Removal.Onnx;
internal class OnnxInferenceRunner : IInferenceRunner
{
public ITensor<float> Run(byte[] model, ITensor<float> inputTensor)
{
using var options = new SessionOptions() { LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_ERROR };
using var session = new InferenceSession(model, options);
var inputs = new List<NamedOnnxValue>()
{
NamedOnnxValue.CreateFromTensor(session.InputNames[0], inputTensor.ToTensor()),
};
var results = session.Run(inputs);
var outputTensor = results[0].AsTensor<float>();
return new OnnxTensor(outputTensor);
}
}
+22 -8
View File
@@ -1,19 +1,28 @@
namespace BGR.Console.Removal.Onnx;
public class OnnxTensor(
int batchSize,
int channels,
int height,
int width
) : ITensor<float>
public class OnnxTensor : ITensor<float>
{
private readonly DenseTensor<float> _tensor =
new([batchSize, channels, height, width]);
private readonly Tensor<float> _tensor;
public int Height => _tensor.Dimensions[2];
public int Width => _tensor.Dimensions[3];
public OnnxTensor(
int batchSize,
int channels,
int height,
int width
)
{
_tensor = new DenseTensor<float>([batchSize, channels, height, width]);
}
public OnnxTensor(Tensor<float> tensor)
{
_tensor = tensor;
}
public void SetValue(
int batch,
int channel,
@@ -34,4 +43,9 @@ public class OnnxTensor(
{
return _tensor[batch, channel, y, x];
}
public Tensor<float> ToTensor()
{
return _tensor;
}
}
+139
View File
@@ -0,0 +1,139 @@
using System.Diagnostics;
namespace BGR.Console.Removal;
internal class RemovalCommand(
IModelFactory modelFactory,
ImageProcessor imageProcessor,
IInferenceRunner inferenceRunner,
IAnsiConsole console,
ILogger<RemovalCommand> logger
) : AsyncCommand<RemovalCommand.Settings>
{
private readonly IModelFactory _modelFactory = modelFactory;
private readonly ImageProcessor _imageProcessor = imageProcessor;
private readonly IInferenceRunner _inferenceRunner = inferenceRunner;
private readonly IAnsiConsole _console = console;
private readonly ILogger<RemovalCommand> _logger = logger;
public override async Task<int> ExecuteAsync(CommandContext context, Settings settings)
{
await _console.Status()
.Spinner(Spinner.Known.Dots)
.SpinnerStyle(Style.Parse("green"))
.StartAsync("Removing background...", async ctx =>
{
ctx.Status("Loading model...");
var model = _logger.TimeAndLogAction(
"Loading model",
() => _modelFactory.Create(settings.ResourceName)
);
ctx.Status("Loading image...");
var image = await _logger.TimeAndLogActionAsync(
"Loading image",
async () => await _imageProcessor.LoadImageAsync(settings.Image)
);
ctx.Status("Creating tensor input...");
var inputTensor = await _logger.TimeAndLogActionAsync(
"Creating tensor input",
async () => await _imageProcessor.CreateTensorInputAsync(image.Data, model)
);
ctx.Status("Running inference...");
var outputTensor = _logger.TimeAndLogAction(
"Running inference",
() => _inferenceRunner.Run(model.Bytes, inputTensor)
);
ctx.Status("Generating mask...");
var mask = await _logger.TimeAndLogActionAsync(
"Generating mask",
async () => await _imageProcessor.GenerateMaskAsync(outputTensor, image.Width, image.Height)
);
ctx.Status("Removing background...");
var output = await _logger.TimeAndLogActionAsync(
"Removing background",
async () => await _imageProcessor.RemoveBackgroundAsync(image.Data, mask)
);
if (settings.IncludeMask)
{
await _logger.TimeAndLogActionAsync(
"Saving mask",
async () => await _imageProcessor.SaveImageAsync(mask, settings.MaskPath)
);
_console.MarkupLine($"[bold]Mask saved to:[/] [blue]{settings.OutputPath}[/]");
}
await _logger.TimeAndLogActionAsync(
"Saving output",
async () => await _imageProcessor.SaveImageAsync(output, settings.OutputPath)
);
_console.MarkupLine($"[bold]Output saved to:[/] [green]{settings.OutputPath}[/]");
});
return 0;
}
internal class Settings : CommandSettings
{
private static readonly Dictionary<string, string> Models = new()
{
{ RmbgModel.Id, "rmbg.onnx" },
{ ModNetModel.Id, "modnet.onnx" },
{ U2NetModel.Id, "u2net.onnx" },
};
[CommandArgument(0, "<image>")]
[Description("Path to the image file whose background you want to remove")]
public string Image { get; init; } = string.Empty;
[CommandOption("--model|-m")]
[Description("The model to use for background removal")]
public string Model { get; init; } = "rmbg";
[CommandOption("--include-mask|-i")]
[Description("Generate and output the mask used for background removal")]
public bool IncludeMask { get; init; } = false;
[CommandOption("--output|-o")]
[Description("Path to output image without background to. File extension will always be .png")]
public string Output { get; init; } = string.Empty;
public string ResourceName => Models[Model];
public string MaskPath => GetOutputPath("_mask");
public string OutputPath => GetOutputPath("_no_bg");
public override ValidationResult Validate()
{
if (File.Exists(Image) is false)
{
return ValidationResult.Error($"The image file '{Image}' does not exist.");
}
if (Models.ContainsKey(Model) is false)
{
return ValidationResult.Error($"The model '{Model}' is not supported.");
}
return ValidationResult.Success();
}
private string GetOutputPath(string modifier)
{
if (string.IsNullOrWhiteSpace(Output))
{
return Path.ChangeExtension(Image, null) + modifier + ".png";
}
return Path.ChangeExtension(Output, ".png");
}
}
}