From ace5aeb0f59fdceb99e607af9780added20da37c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 24 Jan 2014 22:51:17 +0000
Subject: [PATCH] MNIST connected network showing off matrices
---
src/network.c | 201 +++++++++++++++++++++++++++++++++++++++++++------
1 files changed, 175 insertions(+), 26 deletions(-)
diff --git a/src/network.c b/src/network.c
index a77d607..07ac621 100644
--- a/src/network.c
+++ b/src/network.c
@@ -2,10 +2,12 @@
#include "network.h"
#include "image.h"
#include "data.h"
+#include "utils.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
#include "maxpool_layer.h"
+#include "softmax_layer.h"
network make_network(int n)
{
@@ -13,6 +15,8 @@
net.n = n;
net.layers = calloc(net.n, sizeof(void *));
net.types = calloc(net.n, sizeof(LAYER_TYPE));
+ net.outputs = 0;
+ net.output = 0;
return net;
}
@@ -30,6 +34,11 @@
forward_connected_layer(layer, input);
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] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
forward_maxpool_layer(layer, input);
@@ -38,20 +47,23 @@
}
}
-void update_network(network net, double step)
+void update_network(network net, double step, double momentum, double decay)
{
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(layer, step);
+ update_convolutional_layer(layer, step, momentum, decay);
}
else if(net.types[i] == MAXPOOL){
//maxpool_layer layer = *(maxpool_layer *)net.layers[i];
}
+ else if(net.types[i] == SOFTMAX){
+ //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, .3, 0);
+ update_connected_layer(layer, step, momentum, 0);
}
}
}
@@ -64,6 +76,9 @@
} else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.output;
+ } else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ return layer.output;
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.output;
@@ -83,6 +98,9 @@
} else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.delta;
+ } else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ return layer.delta;
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.delta;
@@ -95,8 +113,26 @@
return get_network_delta_layer(net, net.n-1);
}
-void learn_network(network net, double *input)
+void calculate_error_network(network net, double *truth)
{
+ double *delta = get_network_delta(net);
+ double *out = get_network_output(net);
+ int i, k = get_network_output_size(net);
+ for(i = 0; i < k; ++i){
+ delta[i] = truth[i] - out[i];
+ }
+}
+
+int get_predicted_class_network(network net)
+{
+ double *out = get_network_output(net);
+ int k = get_network_output_size(net);
+ return max_index(out, k);
+}
+
+void backward_network(network net, double *input, double *truth)
+{
+ calculate_error_network(net, truth);
int i;
double *prev_input;
double *prev_delta;
@@ -114,7 +150,12 @@
if(i != 0) backward_convolutional_layer(layer, prev_input, prev_delta);
}
else if(net.types[i] == MAXPOOL){
- //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ if(i != 0) backward_maxpool_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];
@@ -124,26 +165,61 @@
}
}
-void train_network_batch(network net, batch b)
+int train_network_datum(network net, double *x, double *y, double step, double momentum, double decay)
{
- int i,j;
- int k = get_network_output_size(net);
+ forward_network(net, x);
+ int class = get_predicted_class_network(net);
+ backward_network(net, x, y);
+ update_network(net, step, momentum, decay);
+ return (y[class]?1:0);
+}
+
+double train_network_sgd(network net, data d, int n, double step, double momentum,double decay)
+{
+ int i;
int correct = 0;
- for(i = 0; i < b.n; ++i){
- forward_network(net, b.images[i].data);
- image o = get_network_image(net);
- double *output = get_network_output(net);
- double *delta = get_network_delta(net);
- for(j = 0; j < k; ++j){
- //printf("%f %f\n", b.truth[i][j], output[j]);
- delta[j] = b.truth[i][j]-output[j];
- if(fabs(delta[j]) < .5) ++correct;
- //printf("%f\n", output[j]);
- }
- learn_network(net, b.images[i].data);
- update_network(net, .00001);
+ for(i = 0; i < n; ++i){
+ int index = rand()%d.X.rows;
+ correct += train_network_datum(net, d.X.vals[index], d.y.vals[index], step, momentum, decay);
+ //if((i+1)%10 == 0){
+ // printf("%d: %f\n", (i+1), (double)correct/(i+1));
+ //}
}
- printf("Accuracy: %f\n", (double)correct/b.n);
+ return (double)correct/n;
+}
+double train_network_batch(network net, data d, int n, double step, double momentum,double decay)
+{
+ int i;
+ int correct = 0;
+ for(i = 0; i < n; ++i){
+ int index = rand()%d.X.rows;
+ double *x = d.X.vals[index];
+ double *y = d.y.vals[index];
+ forward_network(net, x);
+ int class = get_predicted_class_network(net);
+ backward_network(net, x, y);
+ correct += (y[class]?1:0);
+ }
+ update_network(net, step, momentum, decay);
+ return (double)correct/n;
+
+}
+
+
+void train_network(network net, data d, double step, double momentum, double decay)
+{
+ int i;
+ int correct = 0;
+ for(i = 0; i < d.X.rows; ++i){
+ correct += train_network_datum(net, d.X.vals[i], d.y.vals[i], step, momentum, decay);
+ if(i%100 == 0){
+ visualize_network(net);
+ cvWaitKey(10);
+ }
+ }
+ visualize_network(net);
+ cvWaitKey(100);
+ printf("Accuracy: %f\n", (double)correct/d.X.rows);
}
int get_network_output_size_layer(network net, int i)
@@ -162,6 +238,10 @@
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
}
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ return layer.inputs;
+ }
return 0;
}
@@ -181,7 +261,7 @@
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return get_maxpool_image(layer);
}
- return make_image(0,0,0);
+ return make_empty_image(0,0,0);
}
image get_network_image(network net)
@@ -191,17 +271,86 @@
image m = get_network_image_layer(net, i);
if(m.h != 0) return m;
}
- return make_image(1,1,1);
+ return make_empty_image(0,0,0);
}
void visualize_network(network net)
{
int i;
- for(i = 0; i < 1; ++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);
+ visualize_convolutional_filters(layer, buff);
}
}
}
+double *network_predict(network net, double *input)
+{
+ forward_network(net, input);
+ double *out = get_network_output(net);
+ return out;
+}
+
+matrix network_predict_data(network net, data test)
+{
+ int i,j;
+ int k = get_network_output_size(net);
+ matrix pred = make_matrix(test.X.rows, k);
+ for(i = 0; i < test.X.rows; ++i){
+ double *out = network_predict(net, test.X.vals[i]);
+ for(j = 0; j < k; ++j){
+ pred.vals[i][j] = out[j];
+ }
+ }
+ return pred;
+}
+
+void print_network(network net)
+{
+ int i,j;
+ for(i = 0; i < net.n; ++i){
+ double *output = 0;
+ int n = 0;
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ output = layer.output;
+ image m = get_convolutional_image(layer);
+ n = m.h*m.w*m.c;
+ }
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ output = layer.output;
+ image m = get_maxpool_image(layer);
+ n = m.h*m.w*m.c;
+ }
+ else if(net.types[i] == CONNECTED){
+ connected_layer layer = *(connected_layer *)net.layers[i];
+ output = layer.output;
+ n = layer.outputs;
+ }
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ output = layer.output;
+ n = layer.inputs;
+ }
+ double mean = mean_array(output, n);
+ double vari = variance_array(output, n);
+ fprintf(stderr, "Layer %d - Mean: %f, Variance: %f\n",i,mean, vari);
+ if(n > 100) n = 100;
+ for(j = 0; j < n; ++j) fprintf(stderr, "%f, ", output[j]);
+ if(n == 100)fprintf(stderr,".....\n");
+ fprintf(stderr, "\n");
+ }
+}
+
+double network_accuracy(network net, data d)
+{
+ matrix guess = network_predict_data(net, d);
+ double acc = matrix_accuracy(d.y, guess);
+ free_matrix(guess);
+ return acc;
+}
+
--
Gitblit v1.10.0