public ReaderKerasModel(string fname)
{
JObject model = JObject.Parse(File.ReadAllText(fname));
String modelType = (String)model.SelectToken("model_type");
if (!modelType.Equals("Sequential"))
{
throw new Exception("This reader only supports Sequential type models!");
}
SequentialModel seq = new SequentialModel();
List <IKernelDescriptor> descriptors = ReadDescriptors(model);
foreach (var d in descriptors)
{
seq.Add(d);
}
seq.Compile(new DefaultExecutor());
List <IData> weights = ReadWeights(model, descriptors);
seq.SetWeights(weights);
sequential = seq;
}
private Brain(double[] weights)
{
_model = TrainManager.Instance.DefaultModel;
_layers = _model.GetLayers();
_weights = weights;
_model.SetWeights(_layers, _weights);
}
public Brain(SequentialModel model = null)
{
if (!TrainManager.Instance) // when not training, just load the passed model
{
if (model == null)
{
throw new NullReferenceException(
"When not training, there should always a model be passed! But model is null!");
}
Debug.Log("Running passed model.");
_model = model;
_weights = _model.GetWeights();
}
else // when training, load default model
{
_model = TrainManager.Instance.DefaultModel;
if (model != null) // if model not null, try to get the weights from pretrained model
{
_weights = model.GetWeights();
var defaultWeights = _model.GetWeights();
if (defaultWeights.Length != _weights.Length)
{
throw new ArgumentOutOfRangeException(nameof(ArgumentOutOfRangeException),
$"The length of the default model {defaultWeights.Length} and the length of " +
$"the pretrained model {_weights.Length} should be identical.");
}
//Debug.Log("Training with pretrained models.");
}
else // if model null, get weights from default model and train from scratch
{
_weights = _model.GetWeights();
for (var i = 0; i < _weights.Length; i++)
{
_weights[i] = Random.value * 2 - 1f;
}
//Debug.Log("Training from scratch models.");
}
}
_layers = _model.GetLayers();
_model.SetWeights(_layers, _weights);
}
public static void SetWeights(this SequentialModel sequentialModel, int[] layers, double[] weights)
{
var dim = sequentialModel.GetInputDimension();
if (dim.c != layers[0])
{
throw new ArgumentOutOfRangeException(nameof(ArgumentOutOfRangeException),
"Size of input layer should be" +
$"{sequentialModel.GetInputDimension().c} but is {layers[0]}.");
}
var data = new List <IData>();
var count = 0;
for (var i = 0; i < layers.Length - 1; i++)
{
// create weight list
data.Add(new Data2D(1, 1, layers[i], layers[i + 1]));
data.Add(new DataArray(layers[i + 1]));
data.Add(new DataArray(0));
// assign weights to weight list
for (var j = 0; j < layers[i]; j++)
{
for (var k = 0; k < layers[i + 1]; k++)
{
((Data2D)data[i * 3])[0, 0, j, k] = weights[count++];
}
}
// assign biases to weight list
for (var j = 0; j < layers[i + 1]; j++)
{
((DataArray)data[i * 3 + 1])[j] = weights[count++];
}
}
sequentialModel.SetWeights(data);
}