/// <summary>
/// Setup the layer.
/// </summary>
/// <param name="colBottom">Specifies the collection of bottom (input) Blobs.</param>
/// <param name="colTop">Specifies the collection of top (output) Blobs.</param>
public override void LayerSetUp(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
m_log.CHECK(colTop[0] != colBottom[0], type.ToString() + " Layer does not allow in-place computation.");
m_nInferredAxis = -1;
m_nConstantCount = 1;
m_rgCopyAxes.Clear();
BlobShape top_blob_shape = m_param.reshape_param.shape;
int top_num_axes = top_blob_shape.dim.Count();
for (int i = 0; i < top_num_axes; i++)
{
int top_dim = top_blob_shape.dim[i];
if (top_dim == 0)
{
m_rgCopyAxes.Add(i);
}
else if (top_dim == -1)
{
m_log.CHECK_EQ(m_nInferredAxis, -1, "new shape contains multiple -1 dims; at most a single (1) value of -1 may be specified.");
m_nInferredAxis = i;
}
else
{
m_nConstantCount *= top_dim;
}
}
}
/// <summary>
/// Parses the parameter from a RawProto.
/// </summary>
/// <param name="rp">Specifies the RawProto to parse.</param>
/// <returns>A new instance of the parameter is returned.</returns>
public static TransposeParameter FromProto(RawProto rp)
{
TransposeParameter p = new TransposeParameter();
p.m_shape = BlobShape.FromProto(rp);
return(p);
}
/// <summary>
/// Parses the parameter from a RawProto.
/// </summary>
/// <param name="rp">Specifies the RawProto to parse.</param>
/// <returns>A new instance of the parameter is returned.</returns>
public static SqueezeParameter FromProto(RawProto rp)
{
SqueezeParameter p = new SqueezeParameter();
p.m_shape = BlobShape.FromProto(rp);
return(p);
}
/// <summary>
/// Setup the layer.
/// </summary>
/// <param name="colBottom">Specifies the input blobs, which are not used.</param>
/// <param name="colTop">Specifies the output blobs.</param>
public override void LayerSetUp(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
m_shape = m_param.constant_param.output_shape;
m_rgF = m_param.constant_param.values_f;
if (!string.IsNullOrEmpty(m_param.constant_param.binary_data_file))
{
m_log.CHECK(File.Exists(m_param.constant_param.binary_data_file), "The 'binary_data_file' specified ('" + m_param.constant_param.binary_data_file + "') could not be found!");
using (FileStream fs = new FileStream(m_param.constant_param.binary_data_file, FileMode.Open, FileAccess.Read))
using (BinaryReader br = new BinaryReader(fs))
{
BlobProto proto = BlobProto.Load(br);
m_rgF = proto.data;
}
}
}
/// <summary>
/// Initialize the gym with the specified properties.
/// </summary>
/// <param name="log">Specifies the output log to use.</param>
/// <param name="properties">Specifies the properties containing Gym specific initialization parameters.</param>
/// <remarks>
/// The ModelGym uses the following initialization properties.
///
/// 'GpuID' - the GPU to run on.
/// 'ModelDescription' - the model description of the model to use.
/// 'Dataset' - the name of the dataset to use.
/// 'Weights' - the model trained weights.
/// 'CudaPath' - the path of the CudaDnnDLL to use.
/// 'BatchSize' - the batch size used when running images through the model (default = 16).
/// 'RecreateData' - when 'True' the data is re-run through the model, otherwise if already run the data is loaded from file (faster).
/// </remarks>
public void Initialize(Log log, PropertySet properties)
{
m_nGpuID = properties.GetPropertyAsInt("GpuID");
m_strModelDesc = properties.GetProperty("ModelDescription");
m_strDataset = properties.GetProperty("Dataset");
m_rgWeights = properties.GetPropertyBlob("Weights");
m_nBatchSize = properties.GetPropertyAsInt("BatchSize", 16);
m_bRecreateData = properties.GetPropertyAsBool("RecreateData", false);
m_strProject = properties.GetProperty("ProjectName");
if (string.IsNullOrEmpty(m_strProject))
{
m_strProject = "default";
}
string strCudaPath = properties.GetProperty("CudaPath");
SettingsCaffe s = new SettingsCaffe();
s.GpuIds = m_nGpuID.ToString();
s.ImageDbLoadMethod = IMAGEDB_LOAD_METHOD.LOAD_ON_DEMAND_BACKGROUND;
m_imgdb = new MyCaffeImageDatabase2(log);
m_imgdb.InitializeWithDsName1(s, m_strDataset);
m_ds = m_imgdb.GetDatasetByName(m_strDataset);
SimpleDatum sd = m_imgdb.QueryImage(m_ds.TrainingSource.ID, 0, IMGDB_LABEL_SELECTION_METHOD.NONE, IMGDB_IMAGE_SELECTION_METHOD.NONE);
BlobShape shape = new BlobShape(1, sd.Channels, sd.Height, sd.Width);
if (m_evtCancel == null)
{
m_evtCancel = new CancelEvent();
}
m_mycaffe = new MyCaffeControl <float>(s, log, m_evtCancel, null, null, null, null, strCudaPath);
m_mycaffe.LoadToRun(m_strModelDesc, m_rgWeights, shape);
m_log = log;
}
/// <summary>
/// Reshape the bottom (input) and top (output) blobs.
/// </summary>
/// <param name="colBottom">Specifies the collection of bottom (input) Blobs.</param>
/// <param name="colTop">Specifies the collection of top (output) Blobs.</param>
public override void Reshape(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
int input_start_axis = m_param.reshape_param.axis;
int start_axis = (input_start_axis >= 0) ? input_start_axis : colBottom[0].num_axes + input_start_axis + 1;
m_log.CHECK_GE(start_axis, 0, "axis " + input_start_axis.ToString() + " out of range");
m_log.CHECK_LE(start_axis, colBottom[0].num_axes, "axis " + input_start_axis.ToString() + " out of range for " + colBottom[0].num_axes.ToString() + "-D input blob");
int num_axes = m_param.reshape_param.num_axes;
m_log.CHECK_GE(num_axes, -1, "num_axes must be >= 0, or -1 for all");
int end_axis = (num_axes == -1) ? colBottom[0].num_axes : (start_axis + num_axes);
m_log.CHECK_LE(end_axis, colBottom[0].num_axes, "end_axis = axis + num_axes is out of range");
int num_axes_replaced = end_axis - start_axis;
int num_axes_retained = colBottom[0].num_axes - num_axes_replaced;
BlobShape top_blob_shape = m_param.reshape_param.shape;
int num_new_axes = top_blob_shape.dim.Count;
List <int> rgTopShape = new List <int>();
int top_shape_index = 0;
for (int i = 0; i < start_axis; i++)
{
rgTopShape.Add(colBottom[0].shape(i));
top_shape_index++;
}
for (int i = 0; i < num_new_axes; i++)
{
rgTopShape.Add(top_blob_shape.dim[i]);
top_shape_index++;
}
for (int i = end_axis; i < colBottom[0].num_axes; i++)
{
rgTopShape.Add(colBottom[0].shape(i));
top_shape_index++;
}
m_log.CHECK_EQ(top_shape_index, rgTopShape.Count, "The top shape count should equal the top_shape_index.");
for (int i = 0; i < m_rgCopyAxes.Count; i++)
{
int copy_axis_index = m_rgCopyAxes[i];
m_log.CHECK_GT(colBottom[0].num_axes, start_axis + copy_axis_index, "new shape contains a 0, but there was no corresponding bottom axis to copy");
rgTopShape[start_axis + copy_axis_index] = colBottom[0].shape(start_axis + copy_axis_index);
}
if (m_nInferredAxis >= 0)
{
// A -1 dim was specified; infer the correct dimension by computing the
// product of the other dimensions.
int explicit_count = m_nConstantCount;
explicit_count *= colBottom[0].count(0, start_axis);
explicit_count *= colBottom[0].count(end_axis);
for (int i = 0; i < m_rgCopyAxes.Count; i++)
{
int copy_axis_index = m_rgCopyAxes[i];
explicit_count *= rgTopShape[start_axis + copy_axis_index];
}
m_log.CHECK_EQ(0, colBottom[0].count() % explicit_count, "bottom count (" + colBottom[0].count().ToString() + ") must be divisible by the product of the specified dimensions( " + explicit_count.ToString() + ")");
int inferred_dim = colBottom[0].count() / explicit_count;
rgTopShape[start_axis + m_nInferredAxis] = inferred_dim;
}
colTop[0].Reshape(rgTopShape);
m_log.CHECK_EQ(colTop[0].count(), colBottom[0].count(), "output count must match input count");
colTop[0].ShareData(colBottom[0]);
colTop[0].ShareDiff(colBottom[0]);
}
/// <summary>
/// Fills the NetParameter with the LSTM network architecture.
/// </summary>
/// <param name="net_param"></param>
protected override void FillUnrolledNet(NetParameter net_param)
{
uint nNumOutput = m_param.recurrent_param.num_output;
m_log.CHECK_GT(nNumOutput, 0, "num_output must be positive.");
FillerParameter weight_filler = m_param.recurrent_param.weight_filler;
FillerParameter bias_filler = m_param.recurrent_param.bias_filler;
// Add generic LayerParameter's (without bottoms/tops) of layer types we'll
// use to save redundant code.
LayerParameter hidden_param = new param.LayerParameter(LayerParameter.LayerType.INNERPRODUCT);
hidden_param.inner_product_param.num_output = nNumOutput * 4;
hidden_param.inner_product_param.bias_term = false;
hidden_param.inner_product_param.axis = 2;
hidden_param.inner_product_param.weight_filler = weight_filler.Clone();
LayerParameter biased_hidden_param = hidden_param.Clone(false);
biased_hidden_param.inner_product_param.bias_term = true;
biased_hidden_param.inner_product_param.bias_filler = bias_filler.Clone();
LayerParameter sum_param = new param.LayerParameter(LayerParameter.LayerType.ELTWISE);
sum_param.eltwise_param.operation = EltwiseParameter.EltwiseOp.SUM;
LayerParameter scale_param = new LayerParameter(LayerParameter.LayerType.SCALE);
scale_param.scale_param.axis = 0;
LayerParameter slice_param = new LayerParameter(LayerParameter.LayerType.SLICE);
slice_param.slice_param.axis = 0;
LayerParameter split_param = new LayerParameter(LayerParameter.LayerType.SPLIT);
List <BlobShape> rgInputShapes = new List <BlobShape>();
RecurrentInputShapes(rgInputShapes);
m_log.CHECK_EQ(2, rgInputShapes.Count, "There should be 2 input shapes.");
//--- Add the layers ---
LayerParameter input_layer_param = new LayerParameter(LayerParameter.LayerType.INPUT);
input_layer_param.top.Add("c_0");
input_layer_param.input_param.shape.Add(rgInputShapes[0].Clone());
input_layer_param.top.Add("h_0");
input_layer_param.input_param.shape.Add(rgInputShapes[1].Clone());
net_param.layer.Add(input_layer_param);
LayerParameter cont_slice_param = slice_param.Clone(false);
cont_slice_param.name = "cont_slice";
cont_slice_param.bottom.Add("cont");
cont_slice_param.slice_param.axis = 0;
net_param.layer.Add(cont_slice_param);
// Add layer to transform all timesteps of x to the hidden state dimension.
// W_xc_x = W_xc * x + b_c
{
LayerParameter x_transform_param = biased_hidden_param.Clone(false);
x_transform_param.name = "x_transform";
x_transform_param.parameters.Add(new ParamSpec("W_xc"));
x_transform_param.parameters.Add(new ParamSpec("b_c"));
x_transform_param.bottom.Add("x");
x_transform_param.top.Add("W_xc_x");
x_transform_param.propagate_down.Add(true);
net_param.layer.Add(x_transform_param);
}
if (m_bStaticInput)
{
// Add layer to transform x_static to the hidden state dimension.
// W_xc_x_static = W_xc_static * x_static
LayerParameter x_static_transform_param = hidden_param.Clone(false);
x_static_transform_param.inner_product_param.axis = 1;
x_static_transform_param.name = "W_xc_x_static";
x_static_transform_param.parameters.Add(new ParamSpec("W_xc_static"));
x_static_transform_param.bottom.Add("x_static");
x_static_transform_param.top.Add("W_xc_x_static_preshape");
x_static_transform_param.propagate_down.Add(true);
net_param.layer.Add(x_static_transform_param);
LayerParameter reshape_param = new LayerParameter(LayerParameter.LayerType.RESHAPE);
BlobShape new_shape = reshape_param.reshape_param.shape;
new_shape.dim.Add(1); // One timestep.
new_shape.dim.Add(-1); // Should infer m_nN as the dimension so we can reshape on batch size.
new_shape.dim.Add((int)x_static_transform_param.inner_product_param.num_output);
reshape_param.name = "W_xc_x_static_reshape";
reshape_param.bottom.Add("W_xc_x_static_preshape");
reshape_param.top.Add("W_xc_x_static");
net_param.layer.Add(reshape_param);
}
LayerParameter x_slice_param = slice_param.Clone(false);
x_slice_param.name = "W_xc_x_slice";
x_slice_param.bottom.Add("W_xc_x");
net_param.layer.Add(x_slice_param);
LayerParameter output_concat_layer = new LayerParameter(LayerParameter.LayerType.CONCAT);
output_concat_layer.name = "h_concat";
output_concat_layer.top.Add("h");
output_concat_layer.concat_param.axis = 0;
for (int t = 1; t <= m_nT; t++)
{
string tm1s = (t - 1).ToString();
string ts = t.ToString();
cont_slice_param.top.Add("cont_" + ts);
x_slice_param.top.Add("W_xc_x_" + ts);
// Add layer to flush the hidden state when beginning a new sequence,
// as indicated by cont_t.
// h_conted_{t-1} := cont_t * h_{t-1}
//
// Normally, cont_t is binary (i.e., 0 or 1), so:
// h_conted_{t-1} := h_{t-1} if cont_t == 1
// 0 otherwise.
{
LayerParameter cont_h_param = scale_param.Clone(false);
cont_h_param.group_start = true;
cont_h_param.name = "h_conted_" + tm1s;
cont_h_param.bottom.Add("h_" + tm1s);
cont_h_param.bottom.Add("cont_" + ts);
cont_h_param.top.Add("h_conted_" + tm1s);
net_param.layer.Add(cont_h_param);
}
// Add layer to compute
// W_hc_h_{t-1} := W_hc * h_conted_{t-1}
{
LayerParameter w_param = hidden_param.Clone(false);
w_param.name = "transform_" + ts;
w_param.parameters.Add(new ParamSpec("W_hc"));
w_param.bottom.Add("h_conted_" + tm1s);
w_param.top.Add("W_hc_h_" + tm1s);
w_param.inner_product_param.axis = 2;
net_param.layer.Add(w_param);
}
// Add the outputs of the linear transformations to compute the gate input.
// get_input_t := W_hc * h_conted_{t-1} + W_xc * x_t + b_c
// = W_hc_h_{t-1} + W_xc_x_t + b_c
{
LayerParameter input_sum_layer = sum_param.Clone(false);
input_sum_layer.name = "gate_input_" + ts;
input_sum_layer.bottom.Add("W_hc_h_" + tm1s);
input_sum_layer.bottom.Add("W_xc_x_" + ts);
if (m_bStaticInput)
{
input_sum_layer.bottom.Add("W_xc_x_static");
}
input_sum_layer.top.Add("gate_input_" + ts);
net_param.layer.Add(input_sum_layer);
}
// Add LSTMUnit layer to compute the cell & hidden vectors c_t and h_t.
// Inputs: c_{t-1}, gate_input_t = (i_t, f_t, o_t, g_t), cont_t
// Outputs: c_t, h_t
// [ i_t' ]
// [ f_t' ] := gate_input_t
// [ o_t' ]
// [ g_t' ]
// i_t := \sigmoid[i_t']
// f_t := \sigmoid[f_t']
// o_t := \sigmoid[o_t']
// g_t := \tanh[g_t']
// c_t := cont_t * (f_t .* c_{t-1}) + (i_t .* g_t)
// h_t := o_t .* \tanh[c_t]
{
LayerParameter lstm_unit_param = new LayerParameter(LayerParameter.LayerType.LSTM_UNIT);
lstm_unit_param.bottom.Add("c_" + tm1s);
lstm_unit_param.bottom.Add("gate_input_" + ts);
lstm_unit_param.bottom.Add("cont_" + ts);
lstm_unit_param.top.Add("c_" + ts);
lstm_unit_param.top.Add("h_" + ts);
lstm_unit_param.name = "unit_" + ts;
net_param.layer.Add(lstm_unit_param);
}
output_concat_layer.bottom.Add("h_" + ts);
}
{
LayerParameter c_T_copy_param = split_param.Clone(false);
c_T_copy_param.bottom.Add("c_" + m_nT.ToString());
c_T_copy_param.top.Add("c_T");
net_param.layer.Add(c_T_copy_param);
}
net_param.layer.Add(output_concat_layer.Clone(false));
}
/// <summary>
/// Fills the NetParameter with the RNN network architecture.
/// </summary>
/// <param name="net_param"></param>
protected override void FillUnrolledNet(NetParameter net_param)
{
uint nNumOutput = m_param.recurrent_param.num_output;
m_log.CHECK_GT(nNumOutput, 0, "num_output must be positive.");
FillerParameter weight_filler = m_param.recurrent_param.weight_filler;
FillerParameter bias_filler = m_param.recurrent_param.bias_filler;
// Add generic LayerParameter's (without bottoms/tops) of layer types we'll
// use to save redundant code.
LayerParameter hidden_param = new param.LayerParameter(LayerParameter.LayerType.INNERPRODUCT);
hidden_param.inner_product_param.num_output = nNumOutput;
hidden_param.inner_product_param.bias_term = false;
hidden_param.inner_product_param.axis = 2;
hidden_param.inner_product_param.weight_filler = weight_filler.Clone();
LayerParameter biased_hidden_param = hidden_param.Clone(false);
biased_hidden_param.inner_product_param.bias_term = true;
biased_hidden_param.inner_product_param.bias_filler = bias_filler.Clone();
LayerParameter sum_param = new param.LayerParameter(LayerParameter.LayerType.ELTWISE);
sum_param.eltwise_param.operation = EltwiseParameter.EltwiseOp.SUM;
LayerParameter tanh_param = new LayerParameter(LayerParameter.LayerType.TANH);
LayerParameter scale_param = new LayerParameter(LayerParameter.LayerType.SCALE);
scale_param.scale_param.axis = 0;
LayerParameter slice_param = new LayerParameter(LayerParameter.LayerType.SLICE);
slice_param.slice_param.axis = 0;
List <BlobShape> rgInputShapes = new List <BlobShape>();
RecurrentInputShapes(rgInputShapes);
m_log.CHECK_EQ(1, rgInputShapes.Count, "There should only be one input shape.");
//--- Add the layers ---
LayerParameter input_layer_param = new LayerParameter(LayerParameter.LayerType.INPUT);
input_layer_param.top.Add("h_0");
input_layer_param.input_param.shape.Add(rgInputShapes[0]);
net_param.layer.Add(input_layer_param);
LayerParameter cont_slice_param = slice_param.Clone(false);
cont_slice_param.name = "cont_slice";
cont_slice_param.bottom.Add("cont");
cont_slice_param.slice_param.axis = 0;
net_param.layer.Add(cont_slice_param);
// Add layer to transform all timesteps of x to the hidden state dimension.
// W_xh_x = W_xh * x + b_h
{
LayerParameter x_transform_param = biased_hidden_param.Clone(false);
x_transform_param.name = "x_transform";
x_transform_param.parameters.Add(new ParamSpec("W_xh"));
x_transform_param.parameters.Add(new ParamSpec("b_h"));
x_transform_param.bottom.Add("x");
x_transform_param.top.Add("W_xh_x");
x_transform_param.propagate_down.Add(true);
net_param.layer.Add(x_transform_param);
}
if (m_bStaticInput)
{
// Add layer to transform x_static to the hidden state dimension.
// W_xh_x_static = W_xh_static * x_static
LayerParameter x_static_transform_param = hidden_param.Clone(false);
x_static_transform_param.inner_product_param.axis = 1;
x_static_transform_param.name = "W_xh_x_static";
x_static_transform_param.parameters.Add(new ParamSpec("W_xh_static"));
x_static_transform_param.bottom.Add("x_static");
x_static_transform_param.top.Add("W_xh_x_static_preshape");
x_static_transform_param.propagate_down.Add(true);
net_param.layer.Add(x_static_transform_param);
LayerParameter reshape_param = new LayerParameter(LayerParameter.LayerType.RESHAPE);
BlobShape new_shape = reshape_param.reshape_param.shape;
new_shape.dim.Add(1); // One timestep.
new_shape.dim.Add(-1); // Should infer m_nN as the dimension so we can reshape on batch size.
new_shape.dim.Add((int)x_static_transform_param.inner_product_param.num_output);
reshape_param.name = "W_xh_x_static_reshape";
reshape_param.bottom.Add("W_xh_x_static_preshape");
reshape_param.top.Add("W_xh_x_static");
net_param.layer.Add(reshape_param);
}
LayerParameter x_slice_param = slice_param.Clone(false);
x_slice_param.name = "W_xh_x_slice";
x_slice_param.bottom.Add("W_xh_x");
net_param.layer.Add(x_slice_param);
LayerParameter output_concat_layer = new LayerParameter(LayerParameter.LayerType.CONCAT);
output_concat_layer.name = "o_concat";
output_concat_layer.top.Add("o");
output_concat_layer.concat_param.axis = 0;
for (int t = 1; t <= m_nT; t++)
{
string tm1s = (t - 1).ToString();
string ts = t.ToString();
cont_slice_param.top.Add("cont_" + ts);
x_slice_param.top.Add("W_xh_x_" + ts);
// Add layer to flush the hidden state when beginning a new sequence,
// as indicated by cont_t.
// h_conted_{t-1} := cont_t * h_{t-1}
//
// Normally, cont_t is binary (i.e., 0 or 1), so:
// h_conted_{t-1} := h_{t-1} if cont_t == 1
// 0 otherwise.
{
LayerParameter cont_h_param = scale_param.Clone(false);
cont_h_param.name = "h_conted_" + tm1s;
cont_h_param.bottom.Add("h_" + tm1s);
cont_h_param.bottom.Add("cont_" + ts);
cont_h_param.top.Add("h_conted_" + tm1s);
net_param.layer.Add(cont_h_param);
}
// Add layer to compute
// W_hh_h_{t-1} := W_hh * h_conted_{t-1}
{
LayerParameter w_param = hidden_param.Clone(false);
w_param.name = "W_hh_h_" + tm1s;
w_param.parameters.Add(new ParamSpec("W_hh"));
w_param.bottom.Add("h_conted_" + tm1s);
w_param.top.Add("W_hh_h_" + tm1s);
w_param.inner_product_param.axis = 2;
net_param.layer.Add(w_param);
}
// Add layers to compute
// h_t := \tanh( W_hh * h_conted_t{t-1} + W_xh * x_t + b_h )
// = \tanh( W_hh_h_{t-1} + W_xh_t )
{
LayerParameter h_input_sum_param = sum_param.Clone(false);
h_input_sum_param.name = "h_input_sum_" + ts;
h_input_sum_param.bottom.Add("W_hh_h_" + tm1s);
h_input_sum_param.bottom.Add("W_xh_x_" + ts);
if (m_bStaticInput)
{
h_input_sum_param.bottom.Add("W_xh_x_static");
}
h_input_sum_param.top.Add("h_neuron_input_" + ts);
net_param.layer.Add(h_input_sum_param);
}
{
LayerParameter h_neuron_param = tanh_param.Clone(false);
h_neuron_param.name = "h_neuron_input_" + ts;
h_neuron_param.bottom.Add("h_neuron_input_" + ts);
h_neuron_param.top.Add("h_" + ts);
net_param.layer.Add(h_neuron_param);
}
// Add layer to compute
// W_ho_h_t := W_ho * h_t + b_o
{
LayerParameter w_param = biased_hidden_param.Clone(false);
w_param.name = "W_ho_h_" + ts;
w_param.parameters.Add(new ParamSpec("W_ho"));
w_param.parameters.Add(new ParamSpec("b_o"));
w_param.bottom.Add("h_" + ts);
w_param.top.Add("W_ho_h_" + ts);
w_param.inner_product_param.axis = 2;
net_param.layer.Add(w_param);
}
// Add layer to compute
// o_t := \tanh( W_ho * h_t + b_o
// = \tanh( W_ho_h_t )
{
LayerParameter o_neuron_param = tanh_param.Clone(false);
o_neuron_param.name = "o_neuron_" + ts;
o_neuron_param.bottom.Add("W_ho_h_" + ts);
o_neuron_param.top.Add("o_" + ts);
net_param.layer.Add(o_neuron_param);
}
output_concat_layer.bottom.Add("o_" + ts);
}
net_param.layer.Add(output_concat_layer.Clone(false));
}
private void layerSetUpCaffe(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
// Get (recurrent) input/output names.
List <string> rgOutputNames = new List <string>();
OutputBlobNames(rgOutputNames);
List <string> rgRecurInputNames = new List <string>();
RecurrentInputBlobNames(rgRecurInputNames);
List <string> rgRecurOutputNames = new List <string>();
RecurrentOutputBlobNames(rgRecurOutputNames);
int nNumRecurBlobs = rgRecurInputNames.Count;
m_log.CHECK_EQ(nNumRecurBlobs, rgRecurOutputNames.Count, "The number of recurrent input names must equal the number of recurrent output names.");
// If provided, bottom[2] is a static input to the recurrent net.
int nNumHiddenExposed = (m_bExposeHidden) ? nNumRecurBlobs : 0;
m_bStaticInput = (colBottom.Count > 2 + nNumHiddenExposed) ? true : false;
if (m_bStaticInput)
{
m_log.CHECK_GE(colBottom[2].num_axes, 1, "When static input is present, the bottom[2].num_axes must be >= 1");
m_log.CHECK_EQ(m_nN, colBottom[2].shape(0), "When static input is present, the bottom[2].shape(0) must = N which is " + m_nN.ToString());
}
// Create a NetParameter; setup the inputs that aren't unique to particular
// recurrent architectures.
NetParameter net_param = new NetParameter();
LayerParameter input_layer = new LayerParameter(LayerParameter.LayerType.INPUT);
input_layer.top.Add("x");
BlobShape input_shape1 = new param.BlobShape();
for (int i = 0; i < colBottom[0].num_axes; i++)
{
input_shape1.dim.Add(colBottom[0].shape(i));
}
input_layer.input_param.shape.Add(input_shape1);
input_layer.top.Add("cont");
BlobShape input_shape2 = new param.BlobShape();
for (int i = 0; i < colBottom[1].num_axes; i++)
{
input_shape2.dim.Add(colBottom[1].shape(i));
}
input_layer.input_param.shape.Add(input_shape2);
if (m_bStaticInput)
{
input_layer.top.Add("x_static");
BlobShape input_shape3 = new BlobShape();
for (int i = 0; i < colBottom[2].num_axes; i++)
{
input_shape3.dim.Add(colBottom[2].shape(i));
}
input_layer.input_param.shape.Add(input_shape3);
}
net_param.layer.Add(input_layer);
// Call the child's FillUnrolledNet implementation to specify the unrolled
// recurrent architecture.
FillUnrolledNet(net_param);
// Prepend this layer's name to the names of each layer in the unrolled net.
string strLayerName = m_param.name;
if (strLayerName.Length > 0)
{
for (int i = 0; i < net_param.layer.Count; i++)
{
LayerParameter layer = net_param.layer[i];
layer.name = strLayerName + "_" + layer.name;
}
}
// Add 'pseudo-losses' to all outputs to force backpropagation.
// (Setting force_backward is too agressive as we may not need to backprop to
// all inputs, e.g., the sequence continuation indicators.)
List <string> rgPseudoLosses = new List <string>();
for (int i = 0; i < rgOutputNames.Count; i++)
{
rgPseudoLosses.Add(rgOutputNames[i] + "_pseudoloss");
LayerParameter layer = new LayerParameter(LayerParameter.LayerType.REDUCTION, rgPseudoLosses[i]);
layer.bottom.Add(rgOutputNames[i]);
layer.top.Add(rgPseudoLosses[i]);
layer.loss_weight.Add(1.0);
net_param.layer.Add(layer);
}
// Create the unrolled net.
Net <T> sharedNet = null;
if (m_param is LayerParameterEx <T> )
{
RecurrentLayer <T> sharedLayer = ((LayerParameterEx <T>)m_param).SharedLayer as RecurrentLayer <T>;
if (sharedLayer != null)
{
sharedNet = sharedLayer.m_unrolledNet;
}
}
m_unrolledNet = new Net <T>(m_cuda, m_log, net_param, m_evtCancel, null, m_phase, null, sharedNet);
m_unrolledNet.set_debug_info(m_param.recurrent_param.debug_info);
// Setup pointers to the inputs.
m_blobXInputBlob = m_unrolledNet.blob_by_name("x");
m_blobContInputBlob = m_unrolledNet.blob_by_name("cont");
if (m_bStaticInput)
{
m_blobXStaticInputBlob = m_unrolledNet.blob_by_name("x_static");
}
// Setup pointers to paired recurrent inputs/outputs.
m_colRecurInputBlobs = new common.BlobCollection <T>();
m_colRecurOutputBlobs = new common.BlobCollection <T>();
for (int i = 0; i < nNumRecurBlobs; i++)
{
m_colRecurInputBlobs.Add(m_unrolledNet.blob_by_name(rgRecurInputNames[i]));
m_colRecurOutputBlobs.Add(m_unrolledNet.blob_by_name(rgRecurOutputNames[i]));
}
// Setup pointers to outputs.
m_log.CHECK_EQ(colTop.Count() - nNumHiddenExposed, rgOutputNames.Count, "OutputBlobNames must provide output blob name for each top.");
m_colOutputBlobs = new common.BlobCollection <T>();
for (int i = 0; i < rgOutputNames.Count; i++)
{
m_colOutputBlobs.Add(m_unrolledNet.blob_by_name(rgOutputNames[i]));
}
// We should have 2 inputs (x and cont), plus a number of recurrent inputs,
// plus maybe a static input.
int nStaticInput = (m_bStaticInput) ? 1 : 0;
m_log.CHECK_EQ(2 + nNumRecurBlobs + nStaticInput, m_unrolledNet.input_blobs.Count, "The unrolled net input count should equal 2 + number of recurrent blobs (" + nNumRecurBlobs.ToString() + ") + static inputs (" + nStaticInput.ToString() + ")");
// This layer's parameters are any parameters in the layers of the unrolled
// net. We only want one copy of each parameter, so check that the parameter
// is 'owned' by the layer, rather than shared with another.
blobs.Clear();
for (int i = 0; i < m_unrolledNet.parameters.Count; i++)
{
if (m_unrolledNet.param_owners[i] == -1)
{
m_log.WriteLine("Adding parameter " + i.ToString() + ": " + m_unrolledNet.param_display_names[i]);
blobs.Add(m_unrolledNet.parameters[i]);
}
}
// Check that param_propagate_down is set for all of the parameters in the
// unrolled net; set param_propagate_down to true in this layer.
for (int i = 0; i < m_unrolledNet.layers.Count; i++)
{
for (int j = 0; j < m_unrolledNet.layers[i].blobs.Count; j++)
{
m_log.CHECK(m_unrolledNet.layers[i].param_propagate_down(j), "param_propagate_down not set for layer " + i.ToString() + ", param " + j.ToString());
}
}
m_rgbParamPropagateDown = new DictionaryMap <bool>(blobs.Count, true);
// Set the diffs of recurrent outputs to 0 -- we can't backpropagate across
// batches.
for (int i = 0; i < m_colRecurOutputBlobs.Count; i++)
{
m_colRecurOutputBlobs[i].SetDiff(0);
}
// Check that the last output_names.count layers are the pseudo-losses;
// set last_layer_index so that we don't actually run these layers.
List <string> rgLayerNames = m_unrolledNet.layer_names;
m_nLastLayerIndex = rgLayerNames.Count - 1 - rgPseudoLosses.Count;
for (int i = m_nLastLayerIndex + 1, j = 0; i < rgLayerNames.Count; i++, j++)
{
m_log.CHECK(rgLayerNames[i] == rgPseudoLosses[j], "The last layer at idx " + i.ToString() + " should be the pseudo layer named " + rgPseudoLosses[j]);
}
}
/** @copydoc LayerParameterBase::Copy */
public override void Copy(LayerParameterBase src)
{
TransposeParameter p = (TransposeParameter)src;
m_shape = p.m_shape.Clone();
}
/** @copydoc LayerParameterBase::Copy */
public override void Copy(LayerParameterBase src)
{
SqueezeParameter p = (SqueezeParameter)src;
m_shape = p.m_shape.Clone();
}