From af4e4f92dc9e5da160eb6c6870a7b38b863f1c6c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 28 Oct 2014 02:45:06 +0000
Subject: [PATCH] getting rid of sub_arrays, nvidia driver memory leak
---
src/network.c | 127 ++++++++++++++++++++++++++++++++++--------
1 files changed, 102 insertions(+), 25 deletions(-)
diff --git a/src/network.c b/src/network.c
index f9b4667..69942e8 100644
--- a/src/network.c
+++ b/src/network.c
@@ -1,4 +1,5 @@
#include <stdio.h>
+#include <time.h>
#include "network.h"
#include "image.h"
#include "data.h"
@@ -31,10 +32,13 @@
}
#ifdef GPU
+
void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
{
+ //printf("start\n");
int i;
for(i = 0; i < net.n; ++i){
+ //clock_t time = clock();
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
forward_convolutional_layer_gpu(layer, input);
@@ -49,28 +53,29 @@
forward_connected_layer_gpu(layer, input);
input = layer.output_cl;
}
- /*
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- forward_softmax_layer(layer, input);
- input = layer.output;
- }
- else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_layer(layer, input);
- input = layer.output;
- }
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- forward_maxpool_layer(layer, input);
- input = layer.output;
+ forward_maxpool_layer_gpu(layer, input);
+ input = layer.output_cl;
}
- else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- forward_normalization_layer(layer, input);
- input = layer.output;
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ forward_softmax_layer_gpu(layer, input);
+ input = layer.output_cl;
}
- */
+ //printf("%d %f\n", i, sec(clock()-time));
+ /*
+ else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *)net.layers[i];
+ forward_crop_layer(layer, input);
+ input = layer.output;
+ }
+ else if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ forward_normalization_layer(layer, input);
+ input = layer.output;
+ }
+ */
}
}
@@ -99,6 +104,14 @@
connected_layer layer = *(connected_layer *)net.layers[i];
backward_connected_layer_gpu(layer, prev_input, prev_delta);
}
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ backward_maxpool_layer_gpu(layer, prev_delta);
+ }
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ backward_softmax_layer_gpu(layer, prev_delta);
+ }
}
}
@@ -127,6 +140,14 @@
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.output_cl;
}
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ return layer.output_cl;
+ }
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ return layer.output_cl;
+ }
return 0;
}
@@ -140,6 +161,14 @@
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.delta_cl;
}
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ return layer.delta_cl;
+ }
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ return layer.delta_cl;
+ }
return 0;
}
@@ -330,7 +359,7 @@
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- if(i != 0) backward_maxpool_layer(layer, prev_input, prev_delta);
+ if(i != 0) backward_maxpool_layer(layer, prev_delta);
}
else if(net.types[i] == NORMALIZATION){
normalization_layer layer = *(normalization_layer *)net.layers[i];
@@ -338,7 +367,7 @@
}
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
- if(i != 0) backward_softmax_layer(layer, prev_input, prev_delta);
+ if(i != 0) backward_softmax_layer(layer, prev_delta);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
@@ -351,11 +380,13 @@
}
}
+
#ifdef GPU
float train_network_datum_gpu(network net, float *x, float *y)
{
int x_size = get_network_input_size(net)*net.batch;
int y_size = get_network_output_size(net)*net.batch;
+ clock_t time = clock();
if(!*net.input_cl){
*net.input_cl = cl_make_array(x, x_size);
*net.truth_cl = cl_make_array(y, y_size);
@@ -363,14 +394,21 @@
cl_write_array(*net.input_cl, x, x_size);
cl_write_array(*net.truth_cl, y, y_size);
}
+ //printf("trans %f\n", sec(clock()-time));
+ time = clock();
forward_network_gpu(net, *net.input_cl, *net.truth_cl, 1);
- //int class = get_predicted_class_network(net);
+ //printf("forw %f\n", sec(clock()-time));
+ time = clock();
backward_network_gpu(net, *net.input_cl);
+ //printf("back %f\n", sec(clock()-time));
+ time = clock();
float error = get_network_cost(net);
update_network_gpu(net);
- //return (y[class]?1:0);
+ //printf("updt %f\n", sec(clock()-time));
+ time = clock();
return error;
}
+
float train_network_sgd_gpu(network net, data d, int n)
{
int batch = net.batch;
@@ -380,7 +418,25 @@
int i;
float sum = 0;
for(i = 0; i < n; ++i){
- get_batch(d, batch, X, y);
+ get_random_batch(d, batch, X, y);
+ float err = train_network_datum_gpu(net, X, y);
+ sum += err;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+
+float train_network_data_gpu(network net, data d, int n)
+{
+ int batch = net.batch;
+ float *X = calloc(batch*d.X.cols, sizeof(float));
+ float *y = calloc(batch*d.y.cols, sizeof(float));
+
+ int i;
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ get_next_batch(d, batch, i*batch, X, y);
float err = train_network_datum_gpu(net, X, y);
sum += err;
}
@@ -411,7 +467,7 @@
int i;
float sum = 0;
for(i = 0; i < n; ++i){
- get_batch(d, batch, X, y);
+ get_random_batch(d, batch, X, y);
float err = train_network_datum(net, X, y);
sum += err;
}
@@ -594,7 +650,7 @@
image *prev = 0;
int i;
char buff[256];
- show_image(get_network_image_layer(net, 0), "Crop");
+ //show_image(get_network_image_layer(net, 0), "Crop");
for(i = 0; i < net.n; ++i){
sprintf(buff, "Layer %d", i);
if(net.types[i] == CONVOLUTIONAL){
@@ -608,6 +664,27 @@
}
}
+void top_predictions(network net, int n, int *index)
+{
+ int i,j;
+ int k = get_network_output_size(net);
+ float *out = get_network_output(net);
+ float thresh = FLT_MAX;
+ for(i = 0; i < n; ++i){
+ float max = -FLT_MAX;
+ int max_i = -1;
+ for(j = 0; j < k; ++j){
+ float val = out[j];
+ if(val > max && val < thresh){
+ max = val;
+ max_i = j;
+ }
+ }
+ index[i] = max_i;
+ thresh = max;
+ }
+}
+
float *network_predict(network net, float *input)
{
forward_network(net, input, 0, 0);
--
Gitblit v1.10.0