From cd8d53df21f3ad2810add2a8cff766c745f55a17 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 09 May 2014 22:14:52 +0000
Subject: [PATCH] So there WAS this huge bug. Gone now
---
src/network.c | 212 ++++++++++++++++++++++++++++++++++++++--------------
1 files changed, 154 insertions(+), 58 deletions(-)
diff --git a/src/network.c b/src/network.c
index f5fea60..b75eddf 100644
--- a/src/network.c
+++ b/src/network.c
@@ -6,33 +6,38 @@
#include "connected_layer.h"
#include "convolutional_layer.h"
-//#include "old_conv.h"
#include "maxpool_layer.h"
+#include "normalization_layer.h"
#include "softmax_layer.h"
-network make_network(int n)
+network make_network(int n, int batch)
{
network net;
net.n = n;
+ net.batch = batch;
net.layers = calloc(net.n, sizeof(void *));
net.types = calloc(net.n, sizeof(LAYER_TYPE));
net.outputs = 0;
net.output = 0;
+ #ifdef GPU
+ net.input_cl = 0;
+ #endif
return net;
}
-void print_convolutional_cfg(FILE *fp, convolutional_layer *l)
+void print_convolutional_cfg(FILE *fp, convolutional_layer *l, int first)
{
int i;
- fprintf(fp, "[convolutional]\n"
- "height=%d\n"
- "width=%d\n"
- "channels=%d\n"
- "filters=%d\n"
+ fprintf(fp, "[convolutional]\n");
+ if(first) fprintf(fp, "batch=%d\n"
+ "height=%d\n"
+ "width=%d\n"
+ "channels=%d\n",
+ l->batch,l->h, l->w, l->c);
+ fprintf(fp, "filters=%d\n"
"size=%d\n"
"stride=%d\n"
"activation=%s\n",
- l->h, l->w, l->c,
l->n, l->size, l->stride,
get_activation_string(l->activation));
fprintf(fp, "data=");
@@ -40,37 +45,51 @@
for(i = 0; i < l->n*l->c*l->size*l->size; ++i) fprintf(fp, "%g,", l->filters[i]);
fprintf(fp, "\n\n");
}
-void print_connected_cfg(FILE *fp, connected_layer *l)
+void print_connected_cfg(FILE *fp, connected_layer *l, int first)
{
int i;
- fprintf(fp, "[connected]\n"
- "input=%d\n"
- "output=%d\n"
- "activation=%s\n",
- l->inputs, l->outputs,
- get_activation_string(l->activation));
+ fprintf(fp, "[connected]\n");
+ if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs);
+ fprintf(fp, "output=%d\n"
+ "activation=%s\n",
+ l->outputs,
+ get_activation_string(l->activation));
fprintf(fp, "data=");
for(i = 0; i < l->outputs; ++i) fprintf(fp, "%g,", l->biases[i]);
for(i = 0; i < l->inputs*l->outputs; ++i) fprintf(fp, "%g,", l->weights[i]);
fprintf(fp, "\n\n");
}
-void print_maxpool_cfg(FILE *fp, maxpool_layer *l)
+void print_maxpool_cfg(FILE *fp, maxpool_layer *l, int first)
{
- fprintf(fp, "[maxpool]\n"
- "height=%d\n"
- "width=%d\n"
- "channels=%d\n"
- "stride=%d\n\n",
- l->h, l->w, l->c,
- l->stride);
+ fprintf(fp, "[maxpool]\n");
+ if(first) fprintf(fp, "batch=%d\n"
+ "height=%d\n"
+ "width=%d\n"
+ "channels=%d\n",
+ l->batch,l->h, l->w, l->c);
+ fprintf(fp, "stride=%d\n\n", l->stride);
}
-void print_softmax_cfg(FILE *fp, softmax_layer *l)
+void print_normalization_cfg(FILE *fp, normalization_layer *l, int first)
{
- fprintf(fp, "[softmax]\n"
- "input=%d\n\n",
- l->inputs);
+ fprintf(fp, "[localresponsenormalization]\n");
+ if(first) fprintf(fp, "batch=%d\n"
+ "height=%d\n"
+ "width=%d\n"
+ "channels=%d\n",
+ l->batch,l->h, l->w, l->c);
+ fprintf(fp, "size=%d\n"
+ "alpha=%g\n"
+ "beta=%g\n"
+ "kappa=%g\n\n", l->size, l->alpha, l->beta, l->kappa);
+}
+
+void print_softmax_cfg(FILE *fp, softmax_layer *l, int first)
+{
+ fprintf(fp, "[softmax]\n");
+ if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs);
+ fprintf(fp, "\n");
}
void save_network(network net, char *filename)
@@ -81,29 +100,47 @@
for(i = 0; i < net.n; ++i)
{
if(net.types[i] == CONVOLUTIONAL)
- print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i]);
+ print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i], i==0);
else if(net.types[i] == CONNECTED)
- print_connected_cfg(fp, (connected_layer *)net.layers[i]);
+ print_connected_cfg(fp, (connected_layer *)net.layers[i], i==0);
else if(net.types[i] == MAXPOOL)
- print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i]);
+ print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i], i==0);
+ else if(net.types[i] == NORMALIZATION)
+ print_normalization_cfg(fp, (normalization_layer *)net.layers[i], i==0);
else if(net.types[i] == SOFTMAX)
- print_softmax_cfg(fp, (softmax_layer *)net.layers[i]);
+ print_softmax_cfg(fp, (softmax_layer *)net.layers[i], i==0);
}
fclose(fp);
}
-void forward_network(network net, float *input)
+void forward_network(network net, float *input, int train)
{
int i;
+ #ifdef GPU
+ 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;
+ #endif
for(i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ #ifdef GPU
+ forward_convolutional_layer_gpu(layer, input_cl);
+ input_cl = layer.output_cl;
+ #else
forward_convolutional_layer(layer, input);
+ #endif
input = layer.output;
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- forward_connected_layer(layer, input);
+ forward_connected_layer(layer, input, train);
input = layer.output;
}
else if(net.types[i] == SOFTMAX){
@@ -116,6 +153,11 @@
forward_maxpool_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;
+ }
}
}
@@ -133,6 +175,9 @@
else if(net.types[i] == SOFTMAX){
//maxpool_layer layer = *(maxpool_layer *)net.layers[i];
}
+ else if(net.types[i] == NORMALIZATION){
+ //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ }
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
update_connected_layer(layer, step, momentum, decay);
@@ -154,6 +199,9 @@
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.output;
+ } else if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ return layer.output;
}
return 0;
}
@@ -192,11 +240,11 @@
float *out = get_network_output(net);
int i, k = get_network_output_size(net);
for(i = 0; i < k; ++i){
- printf("%f, ", out[i]);
+ //printf("%f, ", out[i]);
delta[i] = truth[i] - out[i];
sum += delta[i]*delta[i];
}
- printf("\n");
+ //printf("\n");
return sum;
}
@@ -223,22 +271,23 @@
}
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- learn_convolutional_layer(layer);
- //learn_convolutional_layer(layer);
- if(i != 0) backward_convolutional_layer(layer, prev_delta);
+ backward_convolutional_layer(layer, prev_delta);
}
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);
}
+ else if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ if(i != 0) backward_normalization_layer(layer, prev_input, prev_delta);
+ }
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);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- learn_connected_layer(layer, prev_input);
- if(i != 0) backward_connected_layer(layer, prev_input, prev_delta);
+ backward_connected_layer(layer, prev_input, prev_delta);
}
}
return error;
@@ -246,7 +295,7 @@
float train_network_datum(network net, float *x, float *y, float step, float momentum, float decay)
{
- forward_network(net, x);
+ forward_network(net, x, 1);
//int class = get_predicted_class_network(net);
float error = backward_network(net, x, y);
update_network(net, step, momentum, decay);
@@ -259,19 +308,26 @@
int i;
float error = 0;
int correct = 0;
+ int pos = 0;
for(i = 0; i < n; ++i){
int index = rand()%d.X.rows;
- error += train_network_datum(net, d.X.vals[index], d.y.vals[index], step, momentum, decay);
+ float err = train_network_datum(net, d.X.vals[index], d.y.vals[index], step, momentum, decay);
float *y = d.y.vals[index];
int class = get_predicted_class_network(net);
correct += (y[class]?1:0);
+ if(y[1]){
+ error += err;
+ ++pos;
+ }
+
+
//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);
- return error/n;
+ //printf("Accuracy: %f\n",(float) correct/n);
+ return error/pos;
}
float train_network_batch(network net, data d, int n, float step, float momentum,float decay)
{
@@ -281,7 +337,7 @@
int index = rand()%d.X.rows;
float *x = d.X.vals[index];
float *y = d.y.vals[index];
- forward_network(net, x);
+ forward_network(net, x, 1);
int class = get_predicted_class_network(net);
backward_network(net, x, y);
correct += (y[class]?1:0);
@@ -305,7 +361,28 @@
}
visualize_network(net);
cvWaitKey(100);
- printf("Accuracy: %f\n", (float)correct/d.X.rows);
+ fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
+}
+
+int get_network_input_size_layer(network net, int i)
+{
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ return layer.h*layer.w*layer.c;
+ }
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ return layer.h*layer.w*layer.c;
+ }
+ else if(net.types[i] == CONNECTED){
+ connected_layer layer = *(connected_layer *)net.layers[i];
+ return layer.inputs;
+ }
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ return layer.inputs;
+ }
+ return 0;
}
int get_network_output_size_layer(network net, int i)
@@ -331,29 +408,33 @@
return 0;
}
-int reset_network_size(network net, int h, int w, int c)
+int resize_network(network net, int h, int w, int c)
{
int i;
for (i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer *layer = (convolutional_layer *)net.layers[i];
- layer->h = h;
- layer->w = w;
- layer->c = c;
+ resize_convolutional_layer(layer, h, w, c);
image output = get_convolutional_image(*layer);
h = output.h;
w = output.w;
c = output.c;
- }
- else if(net.types[i] == MAXPOOL){
+ }else if(net.types[i] == MAXPOOL){
maxpool_layer *layer = (maxpool_layer *)net.layers[i];
- layer->h = h;
- layer->w = w;
- layer->c = c;
+ resize_maxpool_layer(layer, h, w, c);
image output = get_maxpool_image(*layer);
h = output.h;
w = output.w;
c = output.c;
+ }else if(net.types[i] == NORMALIZATION){
+ normalization_layer *layer = (normalization_layer *)net.layers[i];
+ resize_normalization_layer(layer, h, w, c);
+ image output = get_normalization_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
+ }else{
+ error("Cannot resize this type of layer");
}
}
return 0;
@@ -365,6 +446,11 @@
return get_network_output_size_layer(net, i);
}
+int get_network_input_size(network net)
+{
+ return get_network_output_size_layer(net, 0);
+}
+
image get_network_image_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
@@ -375,6 +461,10 @@
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return get_maxpool_image(layer);
}
+ else if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ return get_normalization_image(layer);
+ }
return make_empty_image(0,0,0);
}
@@ -390,20 +480,25 @@
void visualize_network(network net)
{
+ image *prev = 0;
int i;
char buff[256];
for(i = 0; i < net.n; ++i){
sprintf(buff, "Layer %d", i);
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- visualize_convolutional_layer(layer, buff);
+ prev = visualize_convolutional_layer(layer, buff, prev);
+ }
+ if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ visualize_normalization_layer(layer, buff);
}
}
}
float *network_predict(network net, float *input)
{
- forward_network(net, input);
+ forward_network(net, input, 0);
float *out = get_network_output(net);
return out;
}
@@ -468,3 +563,4 @@
return acc;
}
+
--
Gitblit v1.10.0