void backward_crnn_layer_gpu(layer_t l, network_state state)
{
NETWORK_STATE(s);
s.train = state.train;
int i;
layer_t input_layer = *(l.input_layer);
layer_t self_layer = *(l.self_layer);
layer_t 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) {
copy_ongpu(l.hidden * l.batch, input_layer.output_gpu, 1, l.state_gpu, 1);
axpy_ongpu(l.hidden * l.batch, 1, self_layer.output_gpu, 1, l.state_gpu, 1);
s.input = l.state_gpu;
s.delta = self_layer.delta_gpu;
backward_convolutional_layer_gpu(output_layer, s);
l.state_gpu -= l.hidden*l.batch;
s.input = l.state_gpu;
s.delta = self_layer.delta_gpu - l.hidden*l.batch;
if (i == 0) s.delta = NULL;
backward_convolutional_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 = NULL;
backward_convolutional_layer_gpu(input_layer, s);
increment_layer(&input_layer, -1);
increment_layer(&self_layer, -1);
increment_layer(&output_layer, -1);
}
}
void backward_crnn_layer_gpu(layer l, network net)
{
network s = net;
s.train = net.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) {
copy_gpu(l.hidden * l.batch, input_layer.output_gpu, 1, l.state_gpu, 1);
axpy_gpu(l.hidden * l.batch, 1, self_layer.output_gpu, 1, l.state_gpu, 1);
s.input_gpu = l.state_gpu;
s.delta_gpu = self_layer.delta_gpu;
backward_convolutional_layer_gpu(output_layer, s);
l.state_gpu -= l.hidden*l.batch;
s.input_gpu = l.state_gpu;
s.delta_gpu = self_layer.delta_gpu - l.hidden*l.batch;
if (i == 0) s.delta_gpu = 0;
backward_convolutional_layer_gpu(self_layer, s);
copy_gpu(l.hidden*l.batch, self_layer.delta_gpu, 1, input_layer.delta_gpu, 1);
if (i > 0 && l.shortcut) axpy_gpu(l.hidden*l.batch, 1, self_layer.delta_gpu, 1, self_layer.delta_gpu - l.hidden*l.batch, 1);
s.input_gpu = net.input_gpu + i*l.inputs*l.batch;
if(net.delta_gpu) s.delta_gpu = net.delta_gpu + i*l.inputs*l.batch;
else s.delta_gpu = 0;
backward_convolutional_layer_gpu(input_layer, s);
increment_layer(&input_layer, -1);
increment_layer(&self_layer, -1);
increment_layer(&output_layer, -1);
}
}