void forward_network(network net, network_state state)
{
int i;
for(i = 0; i < net.n; ++i){
layer l = net.layers[i];
if(l.type == CONVOLUTIONAL){
forward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
forward_deconvolutional_layer(l, state);
} else if(l.type == DETECTION){
forward_detection_layer(l, state);
} else if(l.type == CONNECTED){
forward_connected_layer(l, state);
} else if(l.type == CROP){
forward_crop_layer(l, state);
} else if(l.type == COST){
forward_cost_layer(l, state);
} else if(l.type == SOFTMAX){
forward_softmax_layer(l, state);
} else if(l.type == MAXPOOL){
forward_maxpool_layer(l, state);
} else if(l.type == DROPOUT){
forward_dropout_layer(l, state);
} else if(l.type == ROUTE){
forward_route_layer(l, net);
}
state.input = l.output;
}
}
void forward_rnn_layer(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);
fill_cpu(l.outputs * l.batch * l.steps, 0, output_layer.delta, 1);
fill_cpu(l.hidden * l.batch * l.steps, 0, self_layer.delta, 1);
fill_cpu(l.hidden * l.batch * l.steps, 0, input_layer.delta, 1);
if (state.train)
fill_cpu(l.hidden * l.batch, 0, l.state, 1);
for (i = 0; i < l.steps; ++i) {
s.input = state.input;
forward_connected_layer(input_layer, s);
s.input = l.state;
forward_connected_layer(self_layer, s);
float *old_state = l.state;
if (state.train)
l.state += l.hidden * l.batch;
if (l.shortcut) {
copy_cpu(l.hidden * l.batch, old_state, 1, l.state, 1);
} else {
fill_cpu(l.hidden * l.batch, 0, l.state, 1);
}
axpy_cpu(l.hidden * l.batch, 1, input_layer.output, 1, l.state, 1);
axpy_cpu(l.hidden * l.batch, 1, self_layer.output, 1, l.state, 1);
s.input = l.state;
forward_connected_layer(output_layer, s);
state.input += l.inputs * l.batch;
increment_layer(&input_layer, 1);
increment_layer(&self_layer, 1);
increment_layer(&output_layer, 1);
}
}
void forward_network(network net, network_state state)
{
state.workspace = net.workspace;
int i;
for(i = 0; i < net.n; ++i){
state.index = i;
layer l = net.layers[i];
if(l.delta){
scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
}
if(l.type == CONVOLUTIONAL){
forward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
forward_deconvolutional_layer(l, state);
} else if(l.type == ACTIVE){
forward_activation_layer(l, state);
} else if(l.type == LOCAL){
forward_local_layer(l, state);
} else if(l.type == NORMALIZATION){
forward_normalization_layer(l, state);
} else if(l.type == BATCHNORM){
forward_batchnorm_layer(l, state);
} else if(l.type == DETECTION){
forward_detection_layer(l, state);
} else if(l.type == CONNECTED){
forward_connected_layer(l, state);
} else if(l.type == RNN){
forward_rnn_layer(l, state);
} else if(l.type == GRU){
forward_gru_layer(l, state);
} else if(l.type == CRNN){
forward_crnn_layer(l, state);
} else if(l.type == CROP){
forward_crop_layer(l, state);
} else if(l.type == COST){
forward_cost_layer(l, state);
} else if(l.type == SOFTMAX){
forward_softmax_layer(l, state);
} else if(l.type == MAXPOOL){
forward_maxpool_layer(l, state);
} else if(l.type == AVGPOOL){
forward_avgpool_layer(l, state);
} else if(l.type == DROPOUT){
forward_dropout_layer(l, state);
} else if(l.type == ROUTE){
forward_route_layer(l, net);
} else if(l.type == SHORTCUT){
forward_shortcut_layer(l, state);
}
state.input = l.output;
}
}
void forward_network(network net, network_state state)
{
int i;
for(i = 0; i < net.n; ++i){
layer l = net.layers[i];
if(l.delta){
scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
}
if(l.type == CONVOLUTIONAL){
forward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
forward_deconvolutional_layer(l, state);
} else if(l.type == NORMALIZATION){
forward_normalization_layer(l, state);
} else if(l.type == DETECTION){
forward_detection_layer(l, state);
} else if(l.type == CONNECTED){
forward_connected_layer(l, state);
} else if(l.type == CROP){
forward_crop_layer(l, state);
} else if(l.type == COST){
forward_cost_layer(l, state);
} else if(l.type == SOFTMAX){
forward_softmax_layer(l, state);
} else if(l.type == MAXPOOL){
forward_maxpool_layer(l, state);
} else if(l.type == AVGPOOL){
forward_avgpool_layer(l, state);
} else if(l.type == DROPOUT){
forward_dropout_layer(l, state);
} else if(l.type == ROUTE){
forward_route_layer(l, net);
}
state.input = l.output;
}
}
void forward_lstm_layer(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);
fill_cpu(l.outputs * l.batch * l.steps, 0, wf.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, wi.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, wg.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, wo.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, uf.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, ui.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, ug.delta, 1);
fill_cpu(l.outputs * l.batch * l.steps, 0, uo.delta, 1);
if (state.train) {
fill_cpu(l.outputs * l.batch * l.steps, 0, l.delta, 1);
}
for (i = 0; i < l.steps; ++i) {
s.input = l.h_cpu;
forward_connected_layer(wf, s);
forward_connected_layer(wi, s);
forward_connected_layer(wg, s);
forward_connected_layer(wo, s);
s.input = state.input;
forward_connected_layer(uf, s);
forward_connected_layer(ui, s);
forward_connected_layer(ug, s);
forward_connected_layer(uo, s);
copy_cpu(l.outputs * l.batch, wf.output, 1, l.f_cpu, 1);
axpy_cpu(l.outputs * l.batch, 1, uf.output, 1, l.f_cpu, 1);
copy_cpu(l.outputs * l.batch, wi.output, 1, l.i_cpu, 1);
axpy_cpu(l.outputs * l.batch, 1, ui.output, 1, l.i_cpu, 1);
copy_cpu(l.outputs * l.batch, wg.output, 1, l.g_cpu, 1);
axpy_cpu(l.outputs * l.batch, 1, ug.output, 1, l.g_cpu, 1);
copy_cpu(l.outputs * l.batch, wo.output, 1, l.o_cpu, 1);
axpy_cpu(l.outputs * l.batch, 1, uo.output, 1, l.o_cpu, 1);
activate_array(l.f_cpu, l.outputs * l.batch, LOGISTIC);
activate_array(l.i_cpu, l.outputs * l.batch, LOGISTIC);
activate_array(l.g_cpu, l.outputs * l.batch, TANH);
activate_array(l.o_cpu, l.outputs * l.batch, LOGISTIC);
copy_cpu(l.outputs * l.batch, l.i_cpu, 1, l.temp_cpu, 1);
mul_cpu(l.outputs * l.batch, l.g_cpu, 1, l.temp_cpu, 1);
mul_cpu(l.outputs * l.batch, l.f_cpu, 1, l.c_cpu, 1);
axpy_cpu(l.outputs * l.batch, 1, l.temp_cpu, 1, l.c_cpu, 1);
copy_cpu(l.outputs * l.batch, l.c_cpu, 1, l.h_cpu, 1);
activate_array(l.h_cpu, l.outputs * l.batch, TANH);
mul_cpu(l.outputs * l.batch, l.o_cpu, 1, l.h_cpu, 1);
copy_cpu(l.outputs * l.batch, l.c_cpu, 1, l.cell_cpu, 1);
copy_cpu(l.outputs * l.batch, l.h_cpu, 1, l.output, 1);
state.input += l.inputs * l.batch;
l.output += l.outputs * l.batch;
l.cell_cpu += 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);
}
}
