public OnlineLinTrainer(int imm_buf_size, double[] gen_triggers, int[] gen_buf_sizes,
Action <Vector> input_method, Func <double> output_method, params LinApproxTask[] tasks)
{
foreach (var t in tasks)
{
this.tasks.Add(t);
}
input_count = tasks[0].linmodel.input_count;
linearity_check_task = tasks[0];
// Immediate and gradient buffer initialization
imm_buf_inputs = new CircularBufferAA <Vector>(imm_buf_size, true);
imm_buf_vectors = new VectorArray(input_count, imm_buf_size);
imm_training_inputs = new CircularBufferAA <Vector>(imm_buf_size, true);
imm_training_vectors = new VectorArray(input_count, imm_buf_size);
grad_training = new CircularBufferAA <GenStruct>(imm_buf_size / 2, true);
grad_training_vectors = new VectorArray(input_count, imm_buf_size / 2);
imm_buf_outputs = new CircularBufferAA <double>(imm_buf_size, true);
imm_training_outputs = new CircularBufferAA <double>(imm_buf_size, true);
// bind vectors in circular buffers to vector arrays
for (int i = 0; i < imm_buf_size; i++)
{
imm_buf_inputs[i] = imm_buf_vectors[i];
imm_training_inputs[i] = imm_training_vectors[i];
if (i < imm_buf_size / 2)
{
grad_training[i] = new GenStruct(0.0, 0, grad_training_vectors[i]);
}
}
// Generalization space initialization
this.gen_triggers = gen_triggers;
gen_space_size = 0;
gen_buffers = new CircularBufferAA <GenStruct> [input_count];
for (int i = 0; i < input_count; i++)
{
gen_buffers[i] = new CircularBufferAA <GenStruct>(gen_buf_sizes[i], true);
gen_space_size += gen_buf_sizes[i];
VectorArray gen_array = new VectorArray(input_count, gen_buf_sizes[i]);
for (int j = 0; j < gen_buf_sizes[i]; j++)
{
gen_buffers[i][j] = new GenStruct(0.0, 0, gen_array[j]);
}
}
// Delegates assignment
input_update_dlg = input_method;
output_update_dlg = output_method;
// Preallocate buffers for matrix operations in linear approximator
foreach (var t in tasks)
{
t.linmodel.preallocate((int)(imm_buf_size * 1.5) + gen_buf_sizes.Sum());
}
// Misc
inputs_changed = new bool[input_count];
nothing_changed = new bool[input_count];
reassigned = new bool[input_count];
}
public OnlineAnnTrainer(SimpleAnn ann, int imm_buf_size, int[] gen_cells, double[] l_gen_bound, double[] u_gen_bound,
Action <Vector> input_method, Func <double> output_method)
{
this.ann = ann;
// Immediate buffer initialization
imm_buf_inputs = new CircularBufferAA <Vector>(imm_buf_size, true);
imm_buf_vectors = new VectorArray(ann.input_count, imm_buf_size);
imm_training_inputs = new CircularBufferAA <Vector>(imm_buf_size, true);
imm_training_vectors = new VectorArray(ann.input_count, imm_buf_size);
imm_buf_outputs = new CircularBufferAA <double>(imm_buf_size, true);
imm_training_outputs = new CircularBufferAA <double>(imm_buf_size, true);
// bind vectors in circular buffers to vector arrays
for (int i = 0; i < imm_buf_size; i++)
{
imm_buf_inputs[i] = imm_buf_vectors[i];
imm_training_inputs[i] = imm_training_vectors[i];
}
// Generalization space initialization
gen_space = new GridSpace <GenStruct>(ann.input_count, gen_cells, l_gen_bound, u_gen_bound);
linear_gen_buff = gen_space.Linearized;
gen_space.put_method = GenBufPutCriteria;
// Delegates assignment
input_update_dlg = input_method;
output_update_dlg = output_method;
// Preallocate buffers for ann
int supercell_size = gen_cells[0];
for (int i = 1; i < ann.input_count; i++)
{
supercell_size *= gen_cells[i];
}
ann.preallocate(imm_buf_size + supercell_size);
// Misc
batch_size = imm_buf_size;
coord_vector = new Vector(ann.input_count);
}
public Vector(VectorArray binding, int index)
{
source = binding;
this.index = index;
}
public Vector(int size)
{
source = new VectorArray(size, 1);
index = 0;
}
