public void load(string file_path_name)
{
FileStream stream = new FileStream(file_path_name, FileMode.Open);
BinaryReader reader = new BinaryReader(stream);
WeightLayerPool.load(reader);
int count = reader.ReadInt32();
for (int i = 0; i < count; i++)
{
string id = reader.ReadString();
int depth = reader.ReadInt32();
int width = reader.ReadInt32();
int height = reader.ReadInt32();
UnitLayer layer = new UnitLayer(id, width, height, depth);
UnitLayers.Add(id, layer);
}
reader.Close();
stream.Close();
}
private void backward_convolute_on_depth(UnitLayer to, int stride, int depth, int length)
{
int to_depth = depth + length;
if (to_depth >= Depth)
{
to_depth = Depth;
}
for (int d = depth; d < to_depth; d++)
{
WeightLayer weights = WeightLayerPool.find(to.Id, Id, d);
int width_of_weights = weights.Width, height_of_weights = weights.Height;
for (int y = 0; y < Height; y++)
{
int top = y * stride;
int bottom = top + height_of_weights;
for (int x = 0; x < Width; x++)
{
int left = x * stride;
to.diff_convolute_product(left, top, left + width_of_weights, bottom, Units[x, y, d].gradient_at_inport, weights);
}
}
weights.diff_weights();
}
}
private void forward_convolute_on_depth(UnitLayer to, int stride, int depth, int length)
{
int to_depth = depth + length;
if (to_depth >= to.Depth)
{
to_depth = to.Depth;
}
for (int d = depth; d < to_depth; d++)
{
WeightLayer weights = WeightLayerPool.find(Id, to.Id, d);
int width_of_weights = weights.Width, height_of_weights = weights.Height;
int width_of_units = to.Width, height_of_units = to.Height;
for (int y = 0; y < height_of_units; y++)
{
int top = y * stride;
int bottom = top + height_of_weights;
for (int x = 0; x < width_of_units; x++)
{
int left = x * stride;
to.set_value_at_inport(x, y, d, weights.convolute_product(left, top, left + width_of_weights, bottom, Units));
}
}
}
}
static public void create_convolution(UnitLayer from, UnitLayer to, int width, int height)
{
for (int d = 0; d < to.Depth; d++)
{
string id = string.Format("C/{0}/{1}/{2}", from.Id, to.Id, d);
WeightLayer layer = new WeightLayer(id, width, height, from.Depth);
layer.fill_weights(from.Width * from.Height * from.Depth);
WeightLayers.Add(id, layer);
}
}
public void backward_convolute(UnitLayer to, int stride)
{
int index = 0, length = Depth / Config.MAX_NUMBER_OF_TASKS + Config.MIN_LENGTH_OF_TASKS;
Task[] tasks = new Task[Config.MAX_NUMBER_OF_TASKS];
for (int d = 0; d < Depth; d += length)
{
tasks[index++] = Task.Factory.StartNew((depth) => backward_convolute_on_depth(to, stride, (int)depth, length), d);
}
Array.Resize(ref tasks, index);
Task.WaitAll(tasks);
}
public void forward_fully_connect(UnitLayer to)
{
int index = 0, length = to.Depth / Config.MAX_NUMBER_OF_TASKS + Config.MIN_LENGTH_OF_TASKS;
Task[] tasks = new Task[Config.MAX_NUMBER_OF_TASKS];
for (int d = 0; d < to.Depth; d += length)
{
tasks[index++] = Task.Factory.StartNew((depth) => forward_fully_connect_on_depth(to, (int)depth, length), d);
}
Array.Resize(ref tasks, index);
Task.WaitAll(tasks);
}
public double[] get_location_at_forward_convolution(UnitLayer units, int width_of_kernel, int height_of_kernel, int stride, double x, double y)
{
int width_of_layer = units.Width;
int height_of_layer = units.Height;
int width = (width_of_layer - width_of_kernel) / stride + 1;
int height = (height_of_layer - height_of_kernel) / stride + 1;
x = x * width / width_of_layer;
y = y * height / height_of_layer;
return(new double[] { x, y });
}
static public void create_fully_connection(UnitLayer from, UnitLayer to)
{
for (int d = 0; d < to.Depth; d++)
{
for (int y = 0; y < to.Height; y++)
{
for (int x = 0; x < to.Width; x++)
{
string id = string.Format("F/{0}/{1}/{2}/{3}/{4}", from.Id, to.Id, x, y, d);
WeightLayer layer = new WeightLayer(id, from.Width, from.Height, from.Depth);
layer.fill_weights(from.Width * from.Height * from.Depth);
WeightLayers.Add(id, layer);
}
}
}
}
public void backward_average_pool(UnitLayer to, int width, int height, int stride)
{
for (int d = 0; d < Depth; d++)
{
for (int y = 0; y < Height; y++)
{
int top = y * stride;
int bottom = top + height;
for (int x = 0; x < Width; x++)
{
int left = x * stride;
to.diff_average_pool(left, top, left + width, bottom, d, Units[x, y, d].gradient_at_inport);
}
}
}
}
public void forward_average_pool(UnitLayer to, int width, int height, int stride)
{
int width_of_units = to.Width, height_of_units = to.Height;
for (int d = 0; d < Depth; d++)
{
for (int y = 0; y < height_of_units; y++)
{
int top = y * stride;
int bottom = top + height;
for (int x = 0; x < width_of_units; x++)
{
int left = x * stride;
to.set_value_at_inport(x, y, d, average_pool(left, top, left + width, bottom, d));
}
}
}
}
private void backward_fully_connect_on_depth(UnitLayer to, int depth, int length)
{
int to_depth = depth + length;
if (to_depth >= Depth)
{
to_depth = Depth;
}
for (int d = depth; d < to_depth; d++)
{
for (int y = 0; y < Height; y++)
{
for (int x = 0; x < Width; x++)
{
WeightLayer weights = WeightLayerPool.find(to.Id, Id, x, y, d);
to.diff_fully_product(Units[x, y, d].gradient_at_inport, weights);
weights.diff_weights();
}
}
}
}
private void forward_fully_connect_on_depth(UnitLayer to, int depth, int length)
{
int width = to.Width, height = to.Height;
int to_depth = depth + length;
if (to_depth >= to.Depth)
{
to_depth = to.Depth;
}
for (int d = depth; d < to_depth; d++)
{
for (int y = 0; y < height; y++)
{
for (int x = 0; x < width; x++)
{
WeightLayer weights = WeightLayerPool.find(Id, to.Id, x, y, d);
to.set_value_at_inport(x, y, d, weights.fully_product(Units));
}
}
}
}
public void create_unit_layer(string id, int width, int height, int depth)
{
UnitLayer layer = new UnitLayer(id, width, height, depth);
UnitLayers.Add(id, layer);
}
public void create_convolution(UnitLayer to, int width, int height)
{
WeightLayerPool.create_convolution(this, to, width, height);
}
public void create_fully_connection(UnitLayer to)
{
WeightLayerPool.create_fully_connection(this, to);
}
