C++ (Cpp) backward_connected_layer_gpu Examples

Example #1

File: rnn_layer.cpp Project: abggcv/darknet-1

void backward_rnn_layer_gpu(layer l, network_state state)
{
    network_state s = {0};
    s.train = state.train;
    int i;
    layer input_layer = *(l.input_layer);
    layer self_layer = *(l.self_layer);
    layer output_layer = *(l.output_layer);
    increment_layer(&input_layer,  l.steps - 1);
    increment_layer(&self_layer,   l.steps - 1);
    increment_layer(&output_layer, l.steps - 1);
    l.state_gpu += l.hidden*l.batch*l.steps;
    for (i = l.steps-1; i >= 0; --i) {

        s.input = l.state_gpu;
        s.delta = self_layer.delta_gpu;
        backward_connected_layer_gpu(output_layer, s);

        l.state_gpu -= l.hidden*l.batch;

        copy_ongpu(l.hidden*l.batch, self_layer.delta_gpu, 1, input_layer.delta_gpu, 1);

        s.input = l.state_gpu;
        s.delta = self_layer.delta_gpu - l.hidden*l.batch;
        if (i == 0) s.delta = 0;
        backward_connected_layer_gpu(self_layer, s);

        //copy_ongpu(l.hidden*l.batch, self_layer.delta_gpu, 1, input_layer.delta_gpu, 1);
        if (i > 0 && l.shortcut) axpy_ongpu(l.hidden*l.batch, 1, self_layer.delta_gpu, 1, self_layer.delta_gpu - l.hidden*l.batch, 1);
        s.input = state.input + i*l.inputs*l.batch;
        if(state.delta) s.delta = state.delta + i*l.inputs*l.batch;
        else s.delta = 0;
        backward_connected_layer_gpu(input_layer, s);

        increment_layer(&input_layer,  -1);
        increment_layer(&self_layer,   -1);
        increment_layer(&output_layer, -1);
    }
}

Example #2

File: lstm_layer.c Project: dagoliveira/radiation-benchmarks

void backward_lstm_layer_gpu(layer l, network state) {
	network s = { 0 };
	s.train = state.train;
	int i;
	layer wf = *(l.wf);
	layer wi = *(l.wi);
	layer wg = *(l.wg);
	layer wo = *(l.wo);

	layer uf = *(l.uf);
	layer ui = *(l.ui);
	layer ug = *(l.ug);
	layer uo = *(l.uo);

	increment_layer(&wf, l.steps - 1);
	increment_layer(&wi, l.steps - 1);
	increment_layer(&wg, l.steps - 1);
	increment_layer(&wo, l.steps - 1);

	increment_layer(&uf, l.steps - 1);
	increment_layer(&ui, l.steps - 1);
	increment_layer(&ug, l.steps - 1);
	increment_layer(&uo, l.steps - 1);

	state.input_gpu += l.inputs * l.batch * (l.steps - 1);
	if (state.delta_gpu)
		state.delta_gpu += l.inputs * l.batch * (l.steps - 1);

	l.output_gpu += l.outputs * l.batch * (l.steps - 1);
	l.cell_gpu += l.outputs * l.batch * (l.steps - 1);
	l.delta_gpu += l.outputs * l.batch * (l.steps - 1);

	for (i = l.steps - 1; i >= 0; --i) {
		if (i != 0)
			copy_gpu(l.outputs * l.batch, l.cell_gpu - l.outputs * l.batch, 1,
					l.prev_cell_gpu, 1, state.st);
		copy_gpu(l.outputs * l.batch, l.cell_gpu, 1, l.c_gpu, 1, state.st);
		if (i != 0)
			copy_gpu(l.outputs * l.batch, l.output_gpu - l.outputs * l.batch, 1,
					l.prev_state_gpu, 1, state.st);
		copy_gpu(l.outputs * l.batch, l.output_gpu, 1, l.h_gpu, 1, state.st);

		l.dh_gpu = (i == 0) ? 0 : l.delta_gpu - l.outputs * l.batch;

		copy_gpu(l.outputs * l.batch, wf.output_gpu, 1, l.f_gpu, 1, state.st);
		axpy_gpu(l.outputs * l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1, state.st);

		copy_gpu(l.outputs * l.batch, wi.output_gpu, 1, l.i_gpu, 1, state.st);
		axpy_gpu(l.outputs * l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1, state.st);

		copy_gpu(l.outputs * l.batch, wg.output_gpu, 1, l.g_gpu, 1, state.st);
		axpy_gpu(l.outputs * l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1, state.st);

		copy_gpu(l.outputs * l.batch, wo.output_gpu, 1, l.o_gpu, 1, state.st);
		axpy_gpu(l.outputs * l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1, state.st);

		activate_array_gpu(l.f_gpu, l.outputs * l.batch, LOGISTIC, state.st);
		activate_array_gpu(l.i_gpu, l.outputs * l.batch, LOGISTIC, state.st);
		activate_array_gpu(l.g_gpu, l.outputs * l.batch, TANH, state.st);
		activate_array_gpu(l.o_gpu, l.outputs * l.batch, LOGISTIC, state.st);

		copy_gpu(l.outputs * l.batch, l.delta_gpu, 1, l.temp3_gpu, 1, state.st);

		copy_gpu(l.outputs * l.batch, l.c_gpu, 1, l.temp_gpu, 1, state.st);
		activate_array_gpu(l.temp_gpu, l.outputs * l.batch, TANH, state.st);

		copy_gpu(l.outputs * l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1, state.st);
		mul_gpu(l.outputs * l.batch, l.o_gpu, 1, l.temp2_gpu, 1, state.st);

		gradient_array_gpu(l.temp_gpu, l.outputs * l.batch, TANH, l.temp2_gpu, state.st);
		axpy_gpu(l.outputs * l.batch, 1, l.dc_gpu, 1, l.temp2_gpu, 1, state.st);

		copy_gpu(l.outputs * l.batch, l.c_gpu, 1, l.temp_gpu, 1, state.st);
		activate_array_gpu(l.temp_gpu, l.outputs * l.batch, TANH, state.st);
		mul_gpu(l.outputs * l.batch, l.temp3_gpu, 1, l.temp_gpu, 1, state.st);
		gradient_array_gpu(l.o_gpu, l.outputs * l.batch, LOGISTIC, l.temp_gpu, state.st);
		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wo.delta_gpu, 1, state.st);
		s.input_gpu = l.prev_state_gpu;
		s.delta_gpu = l.dh_gpu;
		backward_connected_layer_gpu(wo, s);

		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, uo.delta_gpu, 1, state.st);
		s.input_gpu = state.input_gpu;
		s.delta_gpu = state.delta_gpu;
		backward_connected_layer_gpu(uo, s);

		copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
		mul_gpu(l.outputs * l.batch, l.i_gpu, 1, l.temp_gpu, 1, state.st);
		gradient_array_gpu(l.g_gpu, l.outputs * l.batch, TANH, l.temp_gpu, state.st);
		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wg.delta_gpu, 1, state.st);
		s.input_gpu = l.prev_state_gpu;
		s.delta_gpu = l.dh_gpu;
		backward_connected_layer_gpu(wg, s);

		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, ug.delta_gpu, 1, state.st);
		s.input_gpu = state.input_gpu;
		s.delta_gpu = state.delta_gpu;
		backward_connected_layer_gpu(ug, s);

		copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
		mul_gpu(l.outputs * l.batch, l.g_gpu, 1, l.temp_gpu, 1, state.st);
		gradient_array_gpu(l.i_gpu, l.outputs * l.batch, LOGISTIC, l.temp_gpu, state.st);
		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wi.delta_gpu, 1, state.st);
		s.input_gpu = l.prev_state_gpu;
		s.delta_gpu = l.dh_gpu;
		backward_connected_layer_gpu(wi, s);

		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, ui.delta_gpu, 1, state.st);
		s.input_gpu = state.input_gpu;
		s.delta_gpu = state.delta_gpu;
		backward_connected_layer_gpu(ui, s);

		copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
		mul_gpu(l.outputs * l.batch, l.prev_cell_gpu, 1, l.temp_gpu, 1, state.st);
		gradient_array_gpu(l.f_gpu, l.outputs * l.batch, LOGISTIC, l.temp_gpu, state.st);
		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wf.delta_gpu, 1, state.st);
		s.input_gpu = l.prev_state_gpu;
		s.delta_gpu = l.dh_gpu;
		backward_connected_layer_gpu(wf, s);

		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, uf.delta_gpu, 1, state.st);
		s.input_gpu = state.input_gpu;
		s.delta_gpu = state.delta_gpu;
		backward_connected_layer_gpu(uf, s);

		copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
		mul_gpu(l.outputs * l.batch, l.f_gpu, 1, l.temp_gpu, 1, state.st);
		copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, l.dc_gpu, 1, state.st);

		state.input_gpu -= l.inputs * l.batch;
		if (state.delta_gpu)
			state.delta_gpu -= l.inputs * l.batch;
		l.output_gpu -= l.outputs * l.batch;
		l.cell_gpu -= l.outputs * l.batch;
		l.delta_gpu -= l.outputs * l.batch;

		increment_layer(&wf, -1);
		increment_layer(&wi, -1);
		increment_layer(&wg, -1);
		increment_layer(&wo, -1);

		increment_layer(&uf, -1);
		increment_layer(&ui, -1);
		increment_layer(&ug, -1);
		increment_layer(&uo, -1);
	}
}