private static Layer parse_crop(KeyValuePair[] options, SizeParams parameters)
{
int cropHeight = OptionList.option_find_int(options, "crop_height", 1);
int cropWidth = OptionList.option_find_int(options, "crop_width", 1);
bool flip = OptionList.option_find_int(options, "flip", 0) != 0;
float angle = OptionList.option_find_float(options, "angle", 0);
float saturation = OptionList.option_find_float(options, "saturation", 1);
float exposure = OptionList.option_find_float(options, "exposure", 1);
int batch, h, w, c;
h = parameters.H;
w = parameters.W;
c = parameters.C;
batch = parameters.Batch;
if (!(h != 0 && w != 0 && c != 0))
{
Utils.Error("Layer before crop Layer must output image.");
}
bool noadjust = OptionList.option_find_int_quiet(options, "noadjust", 0) != 0;
Layer l = Layer.make_crop_layer(batch, h, w, c, cropHeight, cropWidth, flip, angle, saturation, exposure);
l.Shift = OptionList.option_find_float(options, "shift", 0);
l.Noadjust = noadjust;
return(l);
}
private static Layer parse_connected(KeyValuePair[] options, SizeParams parameters)
{
int output = OptionList.option_find_int(options, "output", 1);
string activationS = OptionList.option_find_str(options, "activation", "logistic");
Activation activation = ActivationsHelper.Get_activation(activationS);
bool batchNormalize = OptionList.option_find_int_quiet(options, "batch_normalize", 0) != 0;
return(Layer.make_connected_layer(parameters.Batch, parameters.Inputs, output, activation, batchNormalize));
}
private static Layer parse_gru(KeyValuePair[] options, SizeParams parameters)
{
int output = OptionList.option_find_int(options, "output", 1);
bool batchNormalize = OptionList.option_find_int_quiet(options, "batch_normalize", 0) != 0;
Layer l = Layer.make_gru_layer(parameters.Batch, parameters.Inputs, output, parameters.TimeSteps, batchNormalize);
return(l);
}
private static Layer parse_region(KeyValuePair[] options, SizeParams parameters)
{
int coords = OptionList.option_find_int(options, "coords", 4);
int classes = OptionList.option_find_int(options, "classes", 20);
int num = OptionList.option_find_int(options, "num", 1);
Layer l = Layer.make_region_layer(parameters.Batch, parameters.W, parameters.H, num, classes, coords);
l.Log = OptionList.option_find_int_quiet(options, "log", 0);
l.Sqrt = OptionList.option_find_int_quiet(options, "sqrt", 0) != 0;
l.Softmax = OptionList.option_find_int(options, "softmax", 0) != 0;
l.MaxBoxes = OptionList.option_find_int_quiet(options, "max", 30);
l.Jitter = OptionList.option_find_float(options, "jitter", .2f);
l.Rescore = OptionList.option_find_int_quiet(options, "rescore", 0) != 0;
l.Thresh = OptionList.option_find_float(options, "thresh", .5f);
l.Classfix = OptionList.option_find_int_quiet(options, "classfix", 0);
l.Absolute = OptionList.option_find_int_quiet(options, "absolute", 0);
l.Random = OptionList.option_find_int_quiet(options, "random", 0) != 0;
l.CoordScale = OptionList.option_find_float(options, "coord_scale", 1);
l.ObjectScale = OptionList.option_find_float(options, "object_scale", 1);
l.NoobjectScale = OptionList.option_find_float(options, "noobject_scale", 1);
l.ClassScale = OptionList.option_find_float(options, "class_scale", 1);
l.BiasMatch = OptionList.option_find_int_quiet(options, "bias_match", 0) != 0;
string treeFile = OptionList.option_find_str(options, "tree", "");
if (!string.IsNullOrEmpty(treeFile))
{
l.SoftmaxTree = new Tree(treeFile);
}
string mapFile = OptionList.option_find_str(options, "map", "");
if (!string.IsNullOrEmpty(mapFile))
{
l.Map = Utils.read_map(mapFile);
}
string a = OptionList.option_find_str(options, "anchors", null);
if (!string.IsNullOrEmpty(a))
{
var lines = a.Split(',');
for (int i = 0; i < lines.Length; ++i)
{
l.BiasesComplete[l.BiasesIndex + i] = float.Parse(lines[i]);
}
}
return(l);
}
private static Layer parse_softmax(KeyValuePair[] options, SizeParams parameters)
{
int groups = OptionList.option_find_int_quiet(options, "groups", 1);
Layer layer = Layer.make_softmax_layer(parameters.Batch, parameters.Inputs, groups);
layer.Temperature = OptionList.option_find_float_quiet(options, "temperature", 1);
string treeFile = OptionList.option_find_str(options, "tree", "");
if (!string.IsNullOrEmpty(treeFile))
{
layer.SoftmaxTree = new Tree(treeFile);
}
return(layer);
}
private static Layer parse_crnn(KeyValuePair[] options, SizeParams parameters)
{
int outputFilters = OptionList.option_find_int(options, "output_filters", 1);
int hiddenFilters = OptionList.option_find_int(options, "hidden_filters", 1);
string activationS = OptionList.option_find_str(options, "activation", "logistic");
Activation activation = ActivationsHelper.Get_activation(activationS);
bool batchNormalize = OptionList.option_find_int_quiet(options, "batch_normalize", 0) != 0;
Layer l = Layer.make_crnn_layer(parameters.Batch, parameters.W, parameters.H, parameters.C, hiddenFilters, outputFilters, parameters.TimeSteps, activation, batchNormalize);
l.Shortcut = OptionList.option_find_int_quiet(options, "shortcut", 0) != 0;
return(l);
}
private static Layer parse_reorg(KeyValuePair[] options, SizeParams parameters)
{
int stride = OptionList.option_find_int(options, "stride", 1);
bool reverse = OptionList.option_find_int_quiet(options, "reverse", 0) != 0;
int batch, h, w, c;
h = parameters.H;
w = parameters.W;
c = parameters.C;
batch = parameters.Batch;
if (!(h != 0 && w != 0 && c != 0))
{
Utils.Error("Layer before reorg Layer must output image.");
}
return(Layer.make_reorg_layer(batch, w, h, c, stride, reverse));
}
private static Layer parse_maxpool(KeyValuePair[] options, SizeParams parameters)
{
int stride = OptionList.option_find_int(options, "stride", 1);
int size = OptionList.option_find_int(options, "size", stride);
int padding = OptionList.option_find_int_quiet(options, "padding", (size - 1) / 2);
int batch, h, w, c;
h = parameters.H;
w = parameters.W;
c = parameters.C;
batch = parameters.Batch;
if (!(h != 0 && w != 0 && c != 0))
{
Utils.Error("Layer before maxpool Layer must output image.");
}
return(Layer.make_maxpool_layer(batch, h, w, c, size, stride, padding));
}
private static Layer parse_convolutional(KeyValuePair[] options, SizeParams parameters)
{
int n = OptionList.option_find_int(options, "filters", 1);
int size = OptionList.option_find_int(options, "size", 1);
int stride = OptionList.option_find_int(options, "stride", 1);
int pad = OptionList.option_find_int_quiet(options, "pad", 0);
int padding = OptionList.option_find_int_quiet(options, "padding", 0);
if (pad != 0)
{
padding = size / 2;
}
string activationS = OptionList.option_find_str(options, "activation", "logistic");
Activation activation = ActivationsHelper.Get_activation(activationS);
int batch, h, w, c;
h = parameters.H;
w = parameters.W;
c = parameters.C;
batch = parameters.Batch;
if (!(h != 0 && w != 0 && c != 0))
{
Utils.Error("Layer before convolutional Layer must output image.");
}
bool batchNormalize = OptionList.option_find_int_quiet(options, "batch_normalize", 0) != 0;
bool binary = OptionList.option_find_int_quiet(options, "binary", 0) != 0;
bool xnor = OptionList.option_find_int_quiet(options, "xnor", 0) != 0;
Layer layer = Layer.make_convolutional_layer(batch, h, w, c, n, size, stride, padding, activation, batchNormalize, binary, xnor, parameters.Net.Adam);
layer.Flipped = OptionList.option_find_int_quiet(options, "flipped", 0);
layer.Dot = OptionList.option_find_float_quiet(options, "dot", 0);
if (parameters.Net.Adam)
{
layer.B1 = parameters.Net.B1;
layer.B2 = parameters.Net.B2;
layer.Eps = parameters.Net.Eps;
}
return(layer);
}
private static Layer parse_detection(KeyValuePair[] options, SizeParams parameters)
{
int coords = OptionList.option_find_int(options, "coords", 1);
int classes = OptionList.option_find_int(options, "classes", 1);
bool rescore = OptionList.option_find_int(options, "rescore", 0) != 0;
int num = OptionList.option_find_int(options, "num", 1);
int side = OptionList.option_find_int(options, "side", 7);
Layer layer = Layer.make_detection_layer(parameters.Batch, parameters.Inputs, num, side, classes, coords, rescore);
layer.Softmax = OptionList.option_find_int(options, "softmax", 0) != 0;
layer.Sqrt = OptionList.option_find_int(options, "sqrt", 0) != 0;
layer.MaxBoxes = OptionList.option_find_int_quiet(options, "max", 30);
layer.CoordScale = OptionList.option_find_float(options, "coord_scale", 1);
layer.Forced = OptionList.option_find_int(options, "forced", 0);
layer.ObjectScale = OptionList.option_find_float(options, "object_scale", 1);
layer.NoobjectScale = OptionList.option_find_float(options, "noobject_scale", 1);
layer.ClassScale = OptionList.option_find_float(options, "class_scale", 1);
layer.Jitter = OptionList.option_find_float(options, "jitter", .2f);
layer.Random = OptionList.option_find_int_quiet(options, "random", 0) != 0;
layer.Reorg = OptionList.option_find_int_quiet(options, "reorg", 0);
return(layer);
}
public static Network parse_network_cfg(string filename)
{
Section[] sections = read_cfg(filename);
if (sections.Length < 1)
{
Utils.Error("Config file has no Sections");
}
var n = sections[0];
Network net = new Network(sections.Length - 1);
SizeParams parameters = new SizeParams();
var s = new Section(n);
var options = s.Options;
if (is_network(s))
{
Utils.Error("First Section must be [net] or [Network]");
}
parse_net_options(options, net);
parameters.H = net.H;
parameters.W = net.W;
parameters.C = net.C;
parameters.Inputs = net.Inputs;
parameters.Batch = net.Batch;
parameters.TimeSteps = net.TimeSteps;
parameters.Net = net;
ulong workspaceSize = 0;
var index = 1;
int count = 0;
Console.Error.Write("Layer filters size input output\n");
while (index < sections.Length)
{
n = sections[index];
index++;
parameters.Index = count;
Console.Error.Write($"{count:5} ");
s = new Section(n);
options = s.Options;
Layer l = new Layer();
LayerType lt = string_to_layer_type(s.Type);
if (lt == LayerType.Convolutional)
{
l = parse_convolutional(options, parameters);
}
else if (lt == LayerType.Local)
{
l = parse_local(options, parameters);
}
else if (lt == LayerType.Avgpool)
{
l = parse_activation(options, parameters);
}
else if (lt == LayerType.Rnn)
{
l = parse_rnn(options, parameters);
}
else if (lt == LayerType.Gru)
{
l = parse_gru(options, parameters);
}
else if (lt == LayerType.Crnn)
{
l = parse_crnn(options, parameters);
}
else if (lt == LayerType.Connected)
{
l = parse_connected(options, parameters);
}
else if (lt == LayerType.Crop)
{
l = parse_crop(options, parameters);
}
else if (lt == LayerType.Cost)
{
l = parse_cost(options, parameters);
}
else if (lt == LayerType.Region)
{
l = parse_region(options, parameters);
}
else if (lt == LayerType.Detection)
{
l = parse_detection(options, parameters);
}
else if (lt == LayerType.Softmax)
{
l = parse_softmax(options, parameters);
net.Hierarchy = l.SoftmaxTree;
}
else if (lt == LayerType.Normalization)
{
l = parse_normalization(options, parameters);
}
else if (lt == LayerType.Batchnorm)
{
l = parse_batchnorm(options, parameters);
}
else if (lt == LayerType.Maxpool)
{
l = parse_maxpool(options, parameters);
}
else if (lt == LayerType.Reorg)
{
l = parse_reorg(options, parameters);
}
else if (lt == LayerType.Avgpool)
{
l = parse_avgpool(options, parameters);
}
else if (lt == LayerType.Route)
{
l = parse_route(options, parameters, net);
}
else if (lt == LayerType.Shortcut)
{
l = parse_shortcut(options, parameters, net);
}
else if (lt == LayerType.Dropout)
{
l = parse_dropout(options, parameters);
l.Output = net.Layers[count - 1].Output;
l.Delta = net.Layers[count - 1].Delta;
l.OutputGpu = net.Layers[count - 1].OutputGpu;
l.DeltaGpu = net.Layers[count - 1].DeltaGpu;
}
else
{
Console.Error.Write($"LayerType not recognized: {s.Type}\n");
}
l.Dontload = OptionList.option_find_int_quiet(options, "dontload", 0) != 0;
l.Dontloadscales = OptionList.option_find_int_quiet(options, "dontloadscales", 0) != 0;
OptionList.option_unused(options);
net.Layers[count] = l;
if (l.WorkspaceSize > workspaceSize)
{
workspaceSize = l.WorkspaceSize;
}
++count;
if (index + 1 < sections.Length)
{
parameters.H = l.OutH;
parameters.W = l.OutW;
parameters.C = l.OutC;
parameters.Inputs = l.Outputs;
}
}
net.Outputs = Network.get_network_output_size(net);
net.Output = Network.get_network_output(net);
if (workspaceSize != 0)
{
if (CudaUtils.UseGpu)
{
net.Workspace = new float[(workspaceSize - 1) / sizeof(float) + 1];
}
else
{
net.Workspace = new float[1];
}
}
return(net);
}
private static void parse_net_options(KeyValuePair[] options, Network net)
{
net.Batch = OptionList.option_find_int(options, "batch", 1);
net.LearningRate = OptionList.option_find_float(options, "learning_rate", .001f);
net.Momentum = OptionList.option_find_float(options, "momentum", .9f);
net.Decay = OptionList.option_find_float(options, "decay", .0001f);
int subdivs = OptionList.option_find_int(options, "subdivisions", 1);
net.TimeSteps = OptionList.option_find_int_quiet(options, "time_steps", 1);
net.Batch /= subdivs;
net.Batch *= net.TimeSteps;
net.Subdivisions = subdivs;
net.Adam = OptionList.option_find_int_quiet(options, "adam", 0) != 0;
if (net.Adam)
{
net.B1 = OptionList.option_find_float(options, "B1", .9f);
net.B2 = OptionList.option_find_float(options, "B2", .999f);
net.Eps = OptionList.option_find_float(options, "eps", .000001f);
}
net.H = OptionList.option_find_int_quiet(options, "height", 0);
net.W = OptionList.option_find_int_quiet(options, "width", 0);
net.C = OptionList.option_find_int_quiet(options, "channels", 0);
net.Inputs = OptionList.option_find_int_quiet(options, "inputs", net.H * net.W * net.C);
net.MaxCrop = OptionList.option_find_int_quiet(options, "max_crop", net.W * 2);
net.MinCrop = OptionList.option_find_int_quiet(options, "min_crop", net.W);
net.Angle = OptionList.option_find_float_quiet(options, "angle", 0);
net.Aspect = OptionList.option_find_float_quiet(options, "aspect", 1);
net.Saturation = OptionList.option_find_float_quiet(options, "saturation", 1);
net.Exposure = OptionList.option_find_float_quiet(options, "exposure", 1);
net.Hue = OptionList.option_find_float_quiet(options, "hue", 0);
if (net.Inputs == 0 && !(net.H != 0 && net.W != 0 && net.C != 0))
{
Utils.Error("No input parameters supplied");
}
string policyS = OptionList.option_find_str(options, "policy", "constant");
net.Policy = get_policy(policyS);
net.BurnIn = OptionList.option_find_int_quiet(options, "burn_in", 0);
if (net.Policy == LearningRatePolicy.Step)
{
net.Step = OptionList.option_find_int(options, "step", 1);
net.Scale = OptionList.option_find_float(options, "scale", 1);
}
else if (net.Policy == LearningRatePolicy.Steps)
{
string l = OptionList.option_find(options, "steps");
string p = OptionList.option_find(options, "scales");
if (string.IsNullOrEmpty(l) || string.IsNullOrEmpty(p))
{
Utils.Error("STEPS policy must have steps and scales in cfg file");
}
var lines = l.Split(',');
int[] steps = new int[lines.Length];
float[] scales = new float[lines.Length];
for (var i = 0; i < lines.Length; ++i)
{
steps[i] = int.Parse(lines[i]);
scales[i] = float.Parse(lines[i]);
}
net.Scales = scales;
net.Steps = steps;
net.NumSteps = lines.Length;
}
else if (net.Policy == LearningRatePolicy.Exp)
{
net.Gamma = OptionList.option_find_float(options, "gamma", 1);
}
else if (net.Policy == LearningRatePolicy.Sig)
{
net.Gamma = OptionList.option_find_float(options, "gamma", 1);
net.Step = OptionList.option_find_int(options, "step", 1);
}
else if (net.Policy == LearningRatePolicy.Poly || net.Policy == LearningRatePolicy.Random)
{
net.Power = OptionList.option_find_float(options, "power", 1);
}
net.MaxBatches = OptionList.option_find_int(options, "max_batches", 0);
}
