mirror/darknet
1
0
Fork 0
mirror of https://github.com/pjreddie/darknet.git synced 2023-08-10 21:13:14 +03:00
Code Issues Projects Releases Wiki Activity

💀 :space:

Browse Source
This commit is contained in:
Joseph Redmon 2017-06-08 14:10:11 -07:00
parent 9963d70481
commit c3e0d90e9f
2 changed files with 124 additions and 124 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

102
src/gru_layer.c
View File

@ -156,33 +156,33 @@ void forward_gru_layer_gpu(layer l, network state)
for (i = 0; i < l.steps; ++i) { for (i = 0; i < l.steps; ++i) {
s.input = l.h_gpu; s.input = l.h_gpu;
forward_connected_layer_gpu(uz, s); forward_connected_layer_gpu(uz, s);
forward_connected_layer_gpu(ur, s); forward_connected_layer_gpu(ur, s);
s.input = state.input; s.input = state.input;
forward_connected_layer_gpu(wz, s); forward_connected_layer_gpu(wz, s);
forward_connected_layer_gpu(wr, s); forward_connected_layer_gpu(wr, s);
forward_connected_layer_gpu(wh, s); forward_connected_layer_gpu(wh, s);
copy_ongpu(l.outputs*l.batch, wz.output_gpu, 1, l.z_gpu, 1); copy_ongpu(l.outputs*l.batch, wz.output_gpu, 1, l.z_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uz.output_gpu, 1, l.z_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uz.output_gpu, 1, l.z_gpu, 1);
copy_ongpu(l.outputs*l.batch, wr.output_gpu, 1, l.r_gpu, 1); copy_ongpu(l.outputs*l.batch, wr.output_gpu, 1, l.r_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, ur.output_gpu, 1, l.r_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, ur.output_gpu, 1, l.r_gpu, 1);
activate_array_ongpu(l.z_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.z_gpu, l.outputs*l.batch, LOGISTIC);
activate_array_ongpu(l.r_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.r_gpu, l.outputs*l.batch, LOGISTIC);
copy_ongpu(l.outputs*l.batch, l.h_gpu, 1, l.hh_gpu, 1); copy_ongpu(l.outputs*l.batch, l.h_gpu, 1, l.hh_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.r_gpu, 1, l.hh_gpu, 1); mul_ongpu(l.outputs*l.batch, l.r_gpu, 1, l.hh_gpu, 1);
s.input = l.hh_gpu; s.input = l.hh_gpu;
forward_connected_layer_gpu(uh, s); forward_connected_layer_gpu(uh, s);
copy_ongpu(l.outputs*l.batch, wh.output_gpu, 1, l.hh_gpu, 1); copy_ongpu(l.outputs*l.batch, wh.output_gpu, 1, l.hh_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uh.output_gpu, 1, l.hh_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uh.output_gpu, 1, l.hh_gpu, 1);
activate_array_ongpu(l.hh_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.hh_gpu, l.outputs*l.batch, TANH);
weighted_sum_gpu(l.h_gpu, l.hh_gpu, l.z_gpu, l.outputs*l.batch, l.output_gpu); weighted_sum_gpu(l.h_gpu, l.hh_gpu, l.z_gpu, l.outputs*l.batch, l.output_gpu);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.h_gpu, 1); copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.h_gpu, 1);
@ -231,44 +231,44 @@ void backward_gru_layer_gpu(layer l, network state)
if (i>0) copy_ongpu(l.outputs*l.batch, l.output_gpu - l.outputs*l.batch, 1, l.prev_state_gpu, 1); if (i>0) copy_ongpu(l.outputs*l.batch, l.output_gpu - l.outputs*l.batch, 1, l.prev_state_gpu, 1);
l.dh_gpu = (i == 0) ? 0 : l.delta_gpu - l.outputs*l.batch; l.dh_gpu = (i == 0) ? 0 : l.delta_gpu - l.outputs*l.batch;
copy_ongpu(l.outputs*l.batch, wz.output_gpu, 1, l.z_gpu, 1); copy_ongpu(l.outputs*l.batch, wz.output_gpu, 1, l.z_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uz.output_gpu, 1, l.z_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uz.output_gpu, 1, l.z_gpu, 1);
copy_ongpu(l.outputs*l.batch, wr.output_gpu, 1, l.r_gpu, 1); copy_ongpu(l.outputs*l.batch, wr.output_gpu, 1, l.r_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, ur.output_gpu, 1, l.r_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, ur.output_gpu, 1, l.r_gpu, 1);
activate_array_ongpu(l.z_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.z_gpu, l.outputs*l.batch, LOGISTIC);
activate_array_ongpu(l.r_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.r_gpu, l.outputs*l.batch, LOGISTIC);
copy_ongpu(l.outputs*l.batch, wh.output_gpu, 1, l.hh_gpu, 1); copy_ongpu(l.outputs*l.batch, wh.output_gpu, 1, l.hh_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uh.output_gpu, 1, l.hh_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uh.output_gpu, 1, l.hh_gpu, 1);
activate_array_ongpu(l.hh_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.hh_gpu, l.outputs*l.batch, TANH);
copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, l.temp3_gpu, 1); copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, l.temp3_gpu, 1);
fill_ongpu(l.outputs*l.batch, 1, l.temp_gpu, 1); fill_ongpu(l.outputs*l.batch, 1, l.temp_gpu, 1);
axpy_ongpu(l.outputs*l.batch, -1, l.z_gpu, 1, l.temp_gpu, 1); axpy_ongpu(l.outputs*l.batch, -1, l.z_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp_gpu, 1);
gradient_array_ongpu(l.hh_gpu, l.outputs*l.batch, TANH, l.temp_gpu); gradient_array_ongpu(l.hh_gpu, l.outputs*l.batch, TANH, l.temp_gpu);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wh.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wh.delta_gpu, 1);
s.input = state.input; s.input = state.input;
s.delta = state.delta; s.delta = state.delta;
backward_connected_layer_gpu(wh, s); backward_connected_layer_gpu(wh, s);
copy_ongpu(l.outputs*l.batch, l.prev_state_gpu, 1, l.temp2_gpu, 1); copy_ongpu(l.outputs*l.batch, l.prev_state_gpu, 1, l.temp2_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.r_gpu, 1, l.temp2_gpu, 1); mul_ongpu(l.outputs*l.batch, l.r_gpu, 1, l.temp2_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, uh.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, uh.delta_gpu, 1);
fill_ongpu(l.outputs*l.batch, 0, l.temp_gpu, 1); fill_ongpu(l.outputs*l.batch, 0, l.temp_gpu, 1);
s.input = l.temp2_gpu; s.input = l.temp2_gpu;
s.delta = l.temp_gpu; s.delta = l.temp_gpu;
backward_connected_layer_gpu(uh, s); backward_connected_layer_gpu(uh, s);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, l.temp2_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, l.temp2_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.prev_state_gpu, 1, l.temp2_gpu, 1); mul_ongpu(l.outputs*l.batch, l.prev_state_gpu, 1, l.temp2_gpu, 1);
gradient_array_ongpu(l.r_gpu, l.outputs*l.batch, LOGISTIC, l.temp2_gpu); gradient_array_ongpu(l.r_gpu, l.outputs*l.batch, LOGISTIC, l.temp2_gpu);
copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, wr.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, wr.delta_gpu, 1);
s.input = state.input; s.input = state.input;
@ -278,21 +278,21 @@ void backward_gru_layer_gpu(layer l, network state)
copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, ur.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, ur.delta_gpu, 1);
s.input = l.prev_state_gpu; s.input = l.prev_state_gpu;
s.delta = l.dh_gpu; s.delta = l.dh_gpu;
backward_connected_layer_gpu(ur, s); backward_connected_layer_gpu(ur, s);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, l.temp2_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, l.temp2_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.r_gpu, 1, l.temp2_gpu, 1); mul_ongpu(l.outputs*l.batch, l.r_gpu, 1, l.temp2_gpu, 1);
if (l.dh_gpu) axpy_ongpu(l.outputs*l.batch, 1, l.temp2_gpu, 1, l.dh_gpu, 1); if (l.dh_gpu) axpy_ongpu(l.outputs*l.batch, 1, l.temp2_gpu, 1, l.dh_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.z_gpu, 1, l.temp2_gpu, 1); mul_ongpu(l.outputs*l.batch, l.z_gpu, 1, l.temp2_gpu, 1);
if (l.dh_gpu) axpy_ongpu(l.outputs*l.batch, 1, l.temp2_gpu, 1, l.dh_gpu, 1); if (l.dh_gpu) axpy_ongpu(l.outputs*l.batch, 1, l.temp2_gpu, 1, l.dh_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.prev_state_gpu, 1, l.temp3_gpu, 1); mul_ongpu(l.outputs*l.batch, l.prev_state_gpu, 1, l.temp3_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.hh_gpu, 1, l.temp2_gpu, 1); mul_ongpu(l.outputs*l.batch, l.hh_gpu, 1, l.temp2_gpu, 1);
axpy_ongpu(l.outputs*l.batch, -1, l.temp2_gpu, 1, l.temp3_gpu, 1); axpy_ongpu(l.outputs*l.batch, -1, l.temp2_gpu, 1, l.temp3_gpu, 1);
gradient_array_ongpu(l.z_gpu, l.outputs*l.batch, LOGISTIC, l.temp3_gpu); gradient_array_ongpu(l.z_gpu, l.outputs*l.batch, LOGISTIC, l.temp3_gpu);
copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, wz.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, wz.delta_gpu, 1);
s.input = state.input; s.input = state.input;
@ -302,7 +302,7 @@ void backward_gru_layer_gpu(layer l, network state)
copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, uz.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, uz.delta_gpu, 1);
s.input = l.prev_state_gpu; s.input = l.prev_state_gpu;
s.delta = l.dh_gpu; s.delta = l.dh_gpu;
backward_connected_layer_gpu(uz, s); backward_connected_layer_gpu(uz, s);
state.input -= l.inputs*l.batch; state.input -= l.inputs*l.batch;
if (state.delta) state.delta -= l.inputs*l.batch; if (state.delta) state.delta -= l.inputs*l.batch;

146
src/lstm_layer.c
View File

@ -80,7 +80,7 @@ layer make_lstm_layer(int batch, int inputs, int outputs, int steps, int batch_n
l.outputs = outputs; l.outputs = outputs;
l.output = calloc(outputs*batch*steps, sizeof(float)); l.output = calloc(outputs*batch*steps, sizeof(float));
l.state = calloc(outputs*batch, sizeof(float)); l.state = calloc(outputs*batch, sizeof(float));
l.forward = forward_lstm_layer; l.forward = forward_lstm_layer;
l.update = update_lstm_layer; l.update = update_lstm_layer;
@ -95,7 +95,7 @@ layer make_lstm_layer(int batch, int inputs, int outputs, int steps, int batch_n
l.output_gpu = cuda_make_array(0, batch*outputs*steps); l.output_gpu = cuda_make_array(0, batch*outputs*steps);
l.cell_gpu = cuda_make_array(0, batch*outputs*steps); l.cell_gpu = cuda_make_array(0, batch*outputs*steps);
l.delta_gpu = cuda_make_array(0, batch*l.outputs*steps); l.delta_gpu = cuda_make_array(0, batch*l.outputs*steps);
l.f_gpu = cuda_make_array(l.output, batch*outputs); l.f_gpu = cuda_make_array(l.output, batch*outputs);
l.i_gpu = cuda_make_array(l.output, batch*outputs); l.i_gpu = cuda_make_array(l.output, batch*outputs);
@ -121,7 +121,7 @@ void forward_lstm_layer(layer l, network state)
{ {
} }
#ifdef GPU #ifdef GPU
void update_lstm_layer_gpu(layer l, int batch, float learning_rate, float momentum, float decay) void update_lstm_layer_gpu(layer l, int batch, float learning_rate, float momentum, float decay)
{ {
update_connected_layer_gpu(*(l.wf), batch, learning_rate, momentum, decay); update_connected_layer_gpu(*(l.wf), batch, learning_rate, momentum, decay);
@ -135,7 +135,7 @@ void update_lstm_layer_gpu(layer l, int batch, float learning_rate, float moment
} }
void forward_lstm_layer_gpu(layer l, network state) void forward_lstm_layer_gpu(layer l, network state)
{ {
network s = { 0 }; network s = { 0 };
s.train = state.train; s.train = state.train;
int i; int i;
@ -159,50 +159,50 @@ void forward_lstm_layer_gpu(layer l, network state)
fill_ongpu(l.outputs * l.batch * l.steps, 0, ug.delta_gpu, 1); fill_ongpu(l.outputs * l.batch * l.steps, 0, ug.delta_gpu, 1);
fill_ongpu(l.outputs * l.batch * l.steps, 0, uo.delta_gpu, 1); fill_ongpu(l.outputs * l.batch * l.steps, 0, uo.delta_gpu, 1);
if (state.train) { if (state.train) {
fill_ongpu(l.outputs * l.batch * l.steps, 0, l.delta_gpu, 1); fill_ongpu(l.outputs * l.batch * l.steps, 0, l.delta_gpu, 1);
} }
for (i = 0; i < l.steps; ++i) { for (i = 0; i < l.steps; ++i) {
s.input = l.h_gpu; s.input = l.h_gpu;
forward_connected_layer_gpu(wf, s); forward_connected_layer_gpu(wf, s);
forward_connected_layer_gpu(wi, s); forward_connected_layer_gpu(wi, s);
forward_connected_layer_gpu(wg, s); forward_connected_layer_gpu(wg, s);
forward_connected_layer_gpu(wo, s); forward_connected_layer_gpu(wo, s);
s.input = state.input; s.input = state.input;
forward_connected_layer_gpu(uf, s); forward_connected_layer_gpu(uf, s);
forward_connected_layer_gpu(ui, s); forward_connected_layer_gpu(ui, s);
forward_connected_layer_gpu(ug, s); forward_connected_layer_gpu(ug, s);
forward_connected_layer_gpu(uo, s); forward_connected_layer_gpu(uo, s);
copy_ongpu(l.outputs*l.batch, wf.output_gpu, 1, l.f_gpu, 1); copy_ongpu(l.outputs*l.batch, wf.output_gpu, 1, l.f_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1);
copy_ongpu(l.outputs*l.batch, wi.output_gpu, 1, l.i_gpu, 1); copy_ongpu(l.outputs*l.batch, wi.output_gpu, 1, l.i_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1);
copy_ongpu(l.outputs*l.batch, wg.output_gpu, 1, l.g_gpu, 1); copy_ongpu(l.outputs*l.batch, wg.output_gpu, 1, l.g_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1);
copy_ongpu(l.outputs*l.batch, wo.output_gpu, 1, l.o_gpu, 1); copy_ongpu(l.outputs*l.batch, wo.output_gpu, 1, l.o_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1);
activate_array_ongpu(l.f_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.f_gpu, l.outputs*l.batch, LOGISTIC);
activate_array_ongpu(l.i_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.i_gpu, l.outputs*l.batch, LOGISTIC);
activate_array_ongpu(l.g_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.g_gpu, l.outputs*l.batch, TANH);
activate_array_ongpu(l.o_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.o_gpu, l.outputs*l.batch, LOGISTIC);
copy_ongpu(l.outputs*l.batch, l.i_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.i_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.g_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.g_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.f_gpu, 1, l.c_gpu, 1); mul_ongpu(l.outputs*l.batch, l.f_gpu, 1, l.c_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, l.temp_gpu, 1, l.c_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, l.temp_gpu, 1, l.c_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.h_gpu, 1); copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.h_gpu, 1);
activate_array_ongpu(l.h_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.h_gpu, l.outputs*l.batch, TANH);
mul_ongpu(l.outputs*l.batch, l.o_gpu, 1, l.h_gpu, 1); mul_ongpu(l.outputs*l.batch, l.o_gpu, 1, l.h_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.cell_gpu, 1); copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.cell_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.h_gpu, 1, l.output_gpu, 1); copy_ongpu(l.outputs*l.batch, l.h_gpu, 1, l.output_gpu, 1);
state.input += l.inputs*l.batch; state.input += l.inputs*l.batch;
l.output_gpu += l.outputs*l.batch; l.output_gpu += l.outputs*l.batch;
@ -217,11 +217,11 @@ void forward_lstm_layer_gpu(layer l, network state)
increment_layer(&ui, 1); increment_layer(&ui, 1);
increment_layer(&ug, 1); increment_layer(&ug, 1);
increment_layer(&uo, 1); increment_layer(&uo, 1);
} }
} }
void backward_lstm_layer_gpu(layer l, network state) void backward_lstm_layer_gpu(layer l, network state)
{ {
network s = { 0 }; network s = { 0 };
s.train = state.train; s.train = state.train;
int i; int i;
@ -250,9 +250,9 @@ void backward_lstm_layer_gpu(layer l, network state)
l.output_gpu += l.outputs*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.cell_gpu += l.outputs*l.batch*(l.steps - 1);
l.delta_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) { for (i = l.steps - 1; i >= 0; --i) {
if (i != 0) copy_ongpu(l.outputs*l.batch, l.cell_gpu - l.outputs*l.batch, 1, l.prev_cell_gpu, 1); if (i != 0) copy_ongpu(l.outputs*l.batch, l.cell_gpu - l.outputs*l.batch, 1, l.prev_cell_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.cell_gpu, 1, l.c_gpu, 1); copy_ongpu(l.outputs*l.batch, l.cell_gpu, 1, l.c_gpu, 1);
if (i != 0) copy_ongpu(l.outputs*l.batch, l.output_gpu - l.outputs*l.batch, 1, l.prev_state_gpu, 1); if (i != 0) copy_ongpu(l.outputs*l.batch, l.output_gpu - l.outputs*l.batch, 1, l.prev_state_gpu, 1);
@ -260,38 +260,38 @@ void backward_lstm_layer_gpu(layer l, network state)
l.dh_gpu = (i == 0) ? 0 : l.delta_gpu - l.outputs*l.batch; l.dh_gpu = (i == 0) ? 0 : l.delta_gpu - l.outputs*l.batch;
copy_ongpu(l.outputs*l.batch, wf.output_gpu, 1, l.f_gpu, 1); copy_ongpu(l.outputs*l.batch, wf.output_gpu, 1, l.f_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1);
copy_ongpu(l.outputs*l.batch, wi.output_gpu, 1, l.i_gpu, 1); copy_ongpu(l.outputs*l.batch, wi.output_gpu, 1, l.i_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1);
copy_ongpu(l.outputs*l.batch, wg.output_gpu, 1, l.g_gpu, 1); copy_ongpu(l.outputs*l.batch, wg.output_gpu, 1, l.g_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1);
copy_ongpu(l.outputs*l.batch, wo.output_gpu, 1, l.o_gpu, 1); copy_ongpu(l.outputs*l.batch, wo.output_gpu, 1, l.o_gpu, 1);
axpy_ongpu(l.outputs*l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1);
activate_array_ongpu(l.f_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.f_gpu, l.outputs*l.batch, LOGISTIC);
activate_array_ongpu(l.i_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.i_gpu, l.outputs*l.batch, LOGISTIC);
activate_array_ongpu(l.g_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.g_gpu, l.outputs*l.batch, TANH);
activate_array_ongpu(l.o_gpu, l.outputs*l.batch, LOGISTIC); activate_array_ongpu(l.o_gpu, l.outputs*l.batch, LOGISTIC);
copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, l.temp3_gpu, 1); copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, l.temp3_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.temp_gpu, 1);
activate_array_ongpu(l.temp_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.temp_gpu, l.outputs*l.batch, TANH);
copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.o_gpu, 1, l.temp2_gpu, 1); mul_ongpu(l.outputs*l.batch, l.o_gpu, 1, l.temp2_gpu, 1);
gradient_array_ongpu(l.temp_gpu, l.outputs*l.batch, TANH, l.temp2_gpu); gradient_array_ongpu(l.temp_gpu, l.outputs*l.batch, TANH, l.temp2_gpu);
axpy_ongpu(l.outputs*l.batch, 1, l.dc_gpu, 1, l.temp2_gpu, 1); axpy_ongpu(l.outputs*l.batch, 1, l.dc_gpu, 1, l.temp2_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.temp_gpu, 1);
activate_array_ongpu(l.temp_gpu, l.outputs*l.batch, TANH); activate_array_ongpu(l.temp_gpu, l.outputs*l.batch, TANH);
mul_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.temp3_gpu, 1, l.temp_gpu, 1);
gradient_array_ongpu(l.o_gpu, l.outputs*l.batch, LOGISTIC, l.temp_gpu); gradient_array_ongpu(l.o_gpu, l.outputs*l.batch, LOGISTIC, l.temp_gpu);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wo.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wo.delta_gpu, 1);
s.input = l.prev_state_gpu; s.input = l.prev_state_gpu;
s.delta = l.dh_gpu; s.delta = l.dh_gpu;
@ -302,9 +302,9 @@ void backward_lstm_layer_gpu(layer l, network state)
s.delta = state.delta; s.delta = state.delta;
backward_connected_layer_gpu(uo, s); backward_connected_layer_gpu(uo, s);
copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.i_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.i_gpu, 1, l.temp_gpu, 1);
gradient_array_ongpu(l.g_gpu, l.outputs*l.batch, TANH, l.temp_gpu); gradient_array_ongpu(l.g_gpu, l.outputs*l.batch, TANH, l.temp_gpu);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wg.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wg.delta_gpu, 1);
s.input = l.prev_state_gpu; s.input = l.prev_state_gpu;
s.delta = l.dh_gpu; s.delta = l.dh_gpu;
@ -315,9 +315,9 @@ void backward_lstm_layer_gpu(layer l, network state)
s.delta = state.delta; s.delta = state.delta;
backward_connected_layer_gpu(ug, s); backward_connected_layer_gpu(ug, s);
copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.g_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.g_gpu, 1, l.temp_gpu, 1);
gradient_array_ongpu(l.i_gpu, l.outputs*l.batch, LOGISTIC, l.temp_gpu); gradient_array_ongpu(l.i_gpu, l.outputs*l.batch, LOGISTIC, l.temp_gpu);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wi.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wi.delta_gpu, 1);
s.input = l.prev_state_gpu; s.input = l.prev_state_gpu;
s.delta = l.dh_gpu; s.delta = l.dh_gpu;
@ -328,9 +328,9 @@ void backward_lstm_layer_gpu(layer l, network state)
s.delta = state.delta; s.delta = state.delta;
backward_connected_layer_gpu(ui, s); backward_connected_layer_gpu(ui, s);
copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.prev_cell_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.prev_cell_gpu, 1, l.temp_gpu, 1);
gradient_array_ongpu(l.f_gpu, l.outputs*l.batch, LOGISTIC, l.temp_gpu); gradient_array_ongpu(l.f_gpu, l.outputs*l.batch, LOGISTIC, l.temp_gpu);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wf.delta_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, wf.delta_gpu, 1);
s.input = l.prev_state_gpu; s.input = l.prev_state_gpu;
s.delta = l.dh_gpu; s.delta = l.dh_gpu;
@ -341,9 +341,9 @@ void backward_lstm_layer_gpu(layer l, network state)
s.delta = state.delta; s.delta = state.delta;
backward_connected_layer_gpu(uf, s); backward_connected_layer_gpu(uf, s);
copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp2_gpu, 1, l.temp_gpu, 1);
mul_ongpu(l.outputs*l.batch, l.f_gpu, 1, l.temp_gpu, 1); mul_ongpu(l.outputs*l.batch, l.f_gpu, 1, l.temp_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, l.dc_gpu, 1); copy_ongpu(l.outputs*l.batch, l.temp_gpu, 1, l.dc_gpu, 1);
state.input -= l.inputs*l.batch; state.input -= l.inputs*l.batch;
if (state.delta) state.delta -= l.inputs*l.batch; if (state.delta) state.delta -= l.inputs*l.batch;
@ -360,6 +360,6 @@ void backward_lstm_layer_gpu(layer l, network state)
increment_layer(&ui, -1); increment_layer(&ui, -1);
increment_layer(&ug, -1); increment_layer(&ug, -1);
increment_layer(&uo, -1); increment_layer(&uo, -1);
} }
} }
#endif #endif