From 8e7a51a4925977acde34daffe9ea1fcffd8aae47 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Wed, 18 Oct 2017 23:56:38 +0000
Subject: [PATCH] circleci: opencv
---
src/rnn_layer.c | 16 ++++++++++++----
1 files changed, 12 insertions(+), 4 deletions(-)
diff --git a/src/rnn_layer.c b/src/rnn_layer.c
index a6b025d..83fda13 100644
--- a/src/rnn_layer.c
+++ b/src/rnn_layer.c
@@ -10,7 +10,7 @@
#include <stdlib.h>
#include <string.h>
-void increment_layer(layer *l, int steps)
+static void increment_layer(layer *l, int steps)
{
int num = l->outputs*l->batch*steps;
l->output += num;
@@ -18,10 +18,12 @@
l->x += num;
l->x_norm += num;
+#ifdef GPU
l->output_gpu += num;
l->delta_gpu += num;
l->x_gpu += num;
l->x_norm_gpu += num;
+#endif
}
layer make_rnn_layer(int batch, int inputs, int hidden, int outputs, int steps, ACTIVATION activation, int batch_normalize, int log)
@@ -56,7 +58,13 @@
l.output = l.output_layer->output;
l.delta = l.output_layer->delta;
+ l.forward = forward_rnn_layer;
+ l.backward = backward_rnn_layer;
+ l.update = update_rnn_layer;
#ifdef GPU
+ l.forward_gpu = forward_rnn_layer_gpu;
+ l.backward_gpu = backward_rnn_layer_gpu;
+ l.update_gpu = update_rnn_layer_gpu;
l.state_gpu = cuda_make_array(l.state, batch*hidden*(steps+1));
l.output_gpu = l.output_layer->output_gpu;
l.delta_gpu = l.output_layer->delta_gpu;
@@ -240,8 +248,6 @@
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;
@@ -249,12 +255,14 @@
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);
+ //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;
--
Gitblit v1.10.0