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 | 353 +++++++++++++++++++++++++++++++++++++++++++++-------------
1 files changed, 271 insertions(+), 82 deletions(-)
diff --git a/src/network.c b/src/network.c
index 292bba0..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"
@@ -8,7 +9,9 @@
#include "connected_layer.h"
#include "convolutional_layer.h"
#include "maxpool_layer.h"
+#include "cost_layer.h"
#include "normalization_layer.h"
+#include "freeweight_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
@@ -22,62 +25,156 @@
net.outputs = 0;
net.output = 0;
#ifdef GPU
- net.input_cl = 0;
+ net.input_cl = calloc(1, sizeof(cl_mem));
+ net.truth_cl = calloc(1, sizeof(cl_mem));
#endif
return net;
}
#ifdef GPU
-void forward_network(network net, float *input, int train)
+
+void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
{
- cl_setup();
- size_t size = get_network_input_size(net);
- if(!net.input_cl){
- net.input_cl = clCreateBuffer(cl.context,
- CL_MEM_READ_WRITE, size*sizeof(float), 0, &cl.error);
- check_error(cl);
- }
- cl_write_array(net.input_cl, input, size);
- cl_mem input_cl = net.input_cl;
+ //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_cl);
- input_cl = layer.output_cl;
- input = layer.output;
+ forward_convolutional_layer_gpu(layer, input);
+ input = layer.output_cl;
+ }
+ else if(net.types[i] == COST){
+ cost_layer layer = *(cost_layer *)net.layers[i];
+ forward_cost_layer_gpu(layer, input, truth);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- forward_connected_layer(layer, input, train);
- input = layer.output;
- }
- 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;
+ forward_connected_layer_gpu(layer, input);
+ input = layer.output_cl;
}
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;
+ }
+ */
+ }
+}
+
+void backward_network_gpu(network net, cl_mem input)
+{
+ int i;
+ cl_mem prev_input;
+ cl_mem prev_delta;
+ for(i = net.n-1; i >= 0; --i){
+ if(i == 0){
+ prev_input = input;
+ prev_delta = 0;
+ }else{
+ prev_input = get_network_output_cl_layer(net, i-1);
+ prev_delta = get_network_delta_cl_layer(net, i-1);
+ }
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ backward_convolutional_layer_gpu(layer, prev_delta);
+ }
+ else if(net.types[i] == COST){
+ cost_layer layer = *(cost_layer *)net.layers[i];
+ backward_cost_layer_gpu(layer, prev_input, prev_delta);
+ }
+ else if(net.types[i] == CONNECTED){
+ 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);
}
}
}
-#else
+void update_network_gpu(network net)
+{
+ int i;
+ for(i = 0; i < net.n; ++i){
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ update_convolutional_layer_gpu(layer);
+ }
+ else if(net.types[i] == CONNECTED){
+ connected_layer layer = *(connected_layer *)net.layers[i];
+ update_connected_layer_gpu(layer);
+ }
+ }
+}
-void forward_network(network net, float *input, int train)
+cl_mem get_network_output_cl_layer(network net, int i)
+{
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ return layer.output_cl;
+ }
+ else if(net.types[i] == CONNECTED){
+ 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;
+}
+
+cl_mem get_network_delta_cl_layer(network net, int i)
+{
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ return layer.delta_cl;
+ }
+ else if(net.types[i] == CONNECTED){
+ 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;
+}
+
+#endif
+
+void forward_network(network net, float *input, float *truth, int train)
{
int i;
for(i = 0; i < net.n; ++i){
@@ -96,6 +193,10 @@
forward_crop_layer(layer, input);
input = layer.output;
}
+ else if(net.types[i] == COST){
+ cost_layer layer = *(cost_layer *)net.layers[i];
+ forward_cost_layer(layer, input, truth);
+ }
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
forward_softmax_layer(layer, input);
@@ -116,9 +217,13 @@
dropout_layer layer = *(dropout_layer *)net.layers[i];
forward_dropout_layer(layer, input);
}
+ else if(net.types[i] == FREEWEIGHT){
+ if(!train) continue;
+ freeweight_layer layer = *(freeweight_layer *)net.layers[i];
+ forward_freeweight_layer(layer, input);
+ }
}
}
-#endif
void update_network(network net)
{
@@ -157,6 +262,8 @@
return layer.output;
} else if(net.types[i] == DROPOUT){
return get_network_output_layer(net, i-1);
+ } else if(net.types[i] == FREEWEIGHT){
+ return get_network_output_layer(net, i-1);
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.output;
@@ -168,7 +275,9 @@
}
float *get_network_output(network net)
{
- return get_network_output_layer(net, net.n-1);
+ int i;
+ for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
+ return get_network_output_layer(net, i);
}
float *get_network_delta_layer(network net, int i)
@@ -184,6 +293,8 @@
return layer.delta;
} else if(net.types[i] == DROPOUT){
return get_network_delta_layer(net, i-1);
+ } else if(net.types[i] == FREEWEIGHT){
+ return get_network_delta_layer(net, i-1);
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.delta;
@@ -191,6 +302,14 @@
return 0;
}
+float get_network_cost(network net)
+{
+ if(net.types[net.n-1] == COST){
+ return ((cost_layer *)net.layers[net.n-1])->output[0];
+ }
+ return 0;
+}
+
float *get_network_delta(network net)
{
return get_network_delta_layer(net, net.n-1);
@@ -221,9 +340,8 @@
return max_index(out, k);
}
-float backward_network(network net, float *input, float *truth)
+void backward_network(network net, float *input)
{
- float error = calculate_error_network(net, truth);
int i;
float *prev_input;
float *prev_delta;
@@ -241,7 +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];
@@ -249,21 +367,92 @@
}
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];
backward_connected_layer(layer, prev_input, prev_delta);
}
+ else if(net.types[i] == COST){
+ cost_layer layer = *(cost_layer *)net.layers[i];
+ backward_cost_layer(layer, prev_input, prev_delta);
+ }
}
+}
+
+
+#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);
+ }else{
+ 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);
+ //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);
+ //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;
+ 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_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;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+#endif
+
+
float train_network_datum(network net, float *x, float *y)
{
- forward_network(net, x, 1);
+ forward_network(net, x, y, 1);
//int class = get_predicted_class_network(net);
- float error = backward_network(net, x, y);
+ backward_network(net, x);
+ float error = get_network_cost(net);
update_network(net);
//return (y[class]?1:0);
return error;
@@ -275,45 +464,13 @@
float *X = calloc(batch*d.X.cols, sizeof(float));
float *y = calloc(batch*d.y.cols, sizeof(float));
- int i,j;
+ int i;
float sum = 0;
- int index = 0;
for(i = 0; i < n; ++i){
- for(j = 0; j < batch; ++j){
- index = rand()%d.X.rows;
- memcpy(X+j*d.X.cols, d.X.vals[index], d.X.cols*sizeof(float));
- memcpy(y+j*d.y.cols, d.y.vals[index], d.y.cols*sizeof(float));
- }
-
+ get_random_batch(d, batch, X, y);
float err = train_network_datum(net, X, y);
sum += err;
- //train_network_datum(net, X, y);
- /*
- float *y = d.y.vals[index];
- int class = get_predicted_class_network(net);
- correct += (y[class]?1:0);
- */
-
-/*
- for(j = 0; j < d.y.cols*batch; ++j){
- printf("%6.3f ", y[j]);
- }
- printf("\n");
- for(j = 0; j < d.y.cols*batch; ++j){
- printf("%6.3f ", get_network_output(net)[j]);
- }
- printf("\n");
- printf("\n");
- */
-
-
- //printf("%d %f %f\n", i,net.output[0], d.y.vals[index][0]);
- //if((i+1)%10 == 0){
- // printf("%d: %f\n", (i+1), (float)correct/(i+1));
- //}
}
- //printf("Accuracy: %f\n",(float) correct/n);
- //show_image(float_to_image(32,32,3,X), "Orig");
free(X);
free(y);
return (float)sum/(n*batch);
@@ -328,8 +485,9 @@
int index = rand()%d.X.rows;
float *x = d.X.vals[index];
float *y = d.y.vals[index];
- forward_network(net, x, 1);
- sum += backward_network(net, x, y);
+ forward_network(net, x, y, 1);
+ backward_network(net, x);
+ sum += get_network_cost(net);
}
update_network(net);
}
@@ -370,6 +528,10 @@
dropout_layer layer = *(dropout_layer *) net.layers[i];
return layer.inputs;
}
+ else if(net.types[i] == FREEWEIGHT){
+ freeweight_layer layer = *(freeweight_layer *) net.layers[i];
+ return layer.inputs;
+ }
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.inputs;
@@ -392,10 +554,15 @@
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
- } else if(net.types[i] == DROPOUT){
+ }
+ else if(net.types[i] == DROPOUT){
dropout_layer layer = *(dropout_layer *) net.layers[i];
return layer.inputs;
}
+ else if(net.types[i] == FREEWEIGHT){
+ freeweight_layer layer = *(freeweight_layer *) net.layers[i];
+ return layer.inputs;
+ }
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.inputs;
@@ -437,7 +604,8 @@
int get_network_output_size(network net)
{
- int i = net.n-1;
+ int i;
+ for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
return get_network_output_size_layer(net, i);
}
@@ -482,7 +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){
@@ -496,9 +664,30 @@
}
}
+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);
+ forward_network(net, input, 0, 0);
float *out = get_network_output(net);
return out;
}
--
Gitblit v1.10.0