static ndarray <float> PrepareImage(byte[,,] image)
{
int height = image.GetLength(0);
int width = image.GetLength(1);
int channels = image.GetLength(2);
var normalized = SirenTests.NormalizeChannelValue(image.ToNumPyArray());
var flattened = normalized.reshape(new[] { height *width, channels }).astype(np.float32_fn);
return((ndarray <float>)flattened);
}
static void Render(Model siren, int width, int height, string path)
{
var renderCoords = SirenTests.Coord(height, width).ToNumPyArray();
ndarray <float> renderBytes = siren.predict(
renderCoords.reshape(new[] { height *width, 2 }),
batch_size: 1024);
const int channels = 4;
renderBytes = (ndarray <float>)renderBytes.reshape(new[] { height, width, channels });
using var bitmap = ToImage(RestoreImage(renderBytes));
bitmap.Save(path, ImageFormat.Png);
}
static void Main(string[] args)
{
GradientEngine.UseEnvironmentFromVariable();
TensorFlowSetup.Instance.EnsureInitialized();
// this allows SIREN to oversaturate channels without adding to the loss
var clampToValidChannelRange = PythonFunctionContainer.Of <Tensor, Tensor>(ClampToValidChannelValueRange);
var siren = new Sequential(new object[] {
new GaussianNoise(stddev: 1f / (128 * 1024)),
new Siren(2, Enumerable.Repeat(256, 5).ToArray()),
new Dense(units: 4, activation: clampToValidChannelRange),
new GaussianNoise(stddev: 1f / 128),
});
siren.compile(
// too slow to converge
//optimizer: new SGD(momentum: 0.5),
// lowered learning rate to avoid destabilization
optimizer: new Adam(learning_rate: 0.00032),
loss: new MeanSquaredError());
if (args.Length == 0)
{
siren.load_weights("sample.weights");
Render(siren, 1034 * 3, 1536 * 3, "sample6X.png");
return;
}
foreach (string imagePath in args)
{
using var original = new Bitmap(imagePath);
byte[,,] image = ToBytesHWC(original);
int height = image.GetLength(0);
int width = image.GetLength(1);
int channels = image.GetLength(2);
Debug.Assert(channels == 4);
var imageSamples = PrepareImage(image);
var coords = SirenTests.Coord(height, width).ToNumPyArray()
.reshape(new[] { width *height, 2 });
var upscaleCoords = SirenTests.Coord(height * 2, width * 2).ToNumPyArray();
var improved = new ImprovedCallback();
improved.OnLossImproved += (sender, eventArgs) => {
if (eventArgs.Epoch < 10)
{
return;
}
ndarray <float> upscaled = siren.predict(
upscaleCoords.reshape(new[] { height *width * 4, 2 }),
batch_size: 1024);
upscaled = (ndarray <float>)upscaled.reshape(new[] { height * 2, width * 2, channels });
using var bitmap = ToImage(RestoreImage(upscaled));
bitmap.Save("sample4X.png", ImageFormat.Png);
siren.save_weights("sample.weights");
Console.WriteLine();
Console.WriteLine("saved!");
};
siren.fit(coords, imageSamples, epochs: 100, batchSize: 64, stepsPerEpoch: 200,
shuffleMode: TrainingShuffleMode.Batch,
callbacks: new ICallback[] { improved });
}
}
static Tensor ClampToValidChannelValueRange(Tensor input)
=> tf.clip_by_value(input,
clip_value_min: SirenTests.NormalizeChannelValue(-0.01f),
clip_value_max: SirenTests.NormalizeChannelValue(255.01f));
