From b715671988a4f3e476586df52fa3bf052cce7f80 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 05 Dec 2013 21:17:16 +0000
Subject: [PATCH] Works well on MNIST
---
src/network.c | 244 ++++++++++++++++++++++++++++++++++++------------
1 files changed, 182 insertions(+), 62 deletions(-)
diff --git a/src/network.c b/src/network.c
index 53184d9..faedb8c 100644
--- a/src/network.c
+++ b/src/network.c
@@ -1,9 +1,13 @@
+#include <stdio.h>
#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)
{
@@ -14,27 +18,29 @@
return net;
}
-void run_network(image input, network net)
+void forward_network(network net, double *input)
{
int i;
- double *input_d = input.data;
for(i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- run_convolutional_layer(input, layer);
+ forward_convolutional_layer(layer, input);
input = layer.output;
- input_d = layer.output.data;
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- run_connected_layer(input_d, layer);
- input_d = layer.output;
+ 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];
- run_maxpool_layer(input, layer);
+ forward_maxpool_layer(layer, input);
input = layer.output;
- input_d = layer.output.data;
}
}
}
@@ -45,121 +51,235 @@
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, 0.9, .01);
}
else if(net.types[i] == MAXPOOL){
//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);
- }
- }
-}
-
-void learn_network(image input, network net)
-{
- int i;
- image prev;
- double *prev_p;
- for(i = net.n-1; i >= 0; --i){
- if(i == 0){
- prev = input;
- prev_p = prev.data;
- } else if(net.types[i-1] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i-1];
- prev = layer.output;
- prev_p = prev.data;
- } else if(net.types[i-1] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i-1];
- prev = layer.output;
- prev_p = prev.data;
- } else if(net.types[i-1] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i-1];
- prev_p = layer.output;
- }
-
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- learn_convolutional_layer(prev, layer);
- }
- else if(net.types[i] == MAXPOOL){
+ 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];
- learn_connected_layer(prev_p, layer);
+ update_connected_layer(layer, step, .9, 0);
}
}
}
-
double *get_network_output_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output.data;
- }
- else if(net.types[i] == MAXPOOL){
+ return layer.output;
+ } else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output.data;
- }
- else if(net.types[i] == CONNECTED){
+ 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;
}
return 0;
}
+double *get_network_output(network net)
+{
+ return get_network_output_layer(net, net.n-1);
+}
+
+double *get_network_delta_layer(network net, int i)
+{
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ return layer.delta;
+ } 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;
+ }
+ return 0;
+}
+
+double *get_network_delta(network net)
+{
+ return get_network_delta_layer(net, net.n-1);
+}
+
+void learn_network(network net, double *input)
+{
+ int i;
+ double *prev_input;
+ double *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_layer(net, i-1);
+ prev_delta = get_network_delta_layer(net, i-1);
+ }
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ learn_convolutional_layer(layer, prev_input);
+ 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];
+ 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];
+ learn_connected_layer(layer, prev_input);
+ if(i != 0) backward_connected_layer(layer, prev_input, prev_delta);
+ }
+ }
+}
+
+void train_network_batch(network net, batch b)
+{
+ int i,j;
+ int k = get_network_output_size(net);
+ int correct = 0;
+ for(i = 0; i < b.n; ++i){
+ show_image(b.images[i], "Input");
+ forward_network(net, b.images[i].data);
+ image o = get_network_image(net);
+ if(o.h) show_image_collapsed(o, "Output");
+ double *output = get_network_output(net);
+ double *delta = get_network_delta(net);
+ int max_k = 0;
+ double max = 0;
+ for(j = 0; j < k; ++j){
+ delta[j] = b.truth[i][j]-output[j];
+ if(output[j] > max) {
+ max = output[j];
+ max_k = j;
+ }
+ }
+ if(b.truth[i][max_k]) ++correct;
+ printf("%f\n", (double)correct/(i+1));
+ learn_network(net, b.images[i].data);
+ update_network(net, .001);
+ if(i%100 == 0){
+ visualize_network(net);
+ cvWaitKey(100);
+ }
+ }
+ visualize_network(net);
+ print_network(net);
+ cvWaitKey(100);
+ printf("Accuracy: %f\n", (double)correct/b.n);
+}
int get_network_output_size_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output.h*layer.output.w*layer.output.c;
+ image output = get_convolutional_image(layer);
+ return output.h*output.w*output.c;
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output.h*layer.output.w*layer.output.c;
+ image output = get_maxpool_image(layer);
+ return output.h*output.w*output.c;
}
else if(net.types[i] == CONNECTED){
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;
}
-double *get_network_output(network net)
+int get_network_output_size(network net)
{
int i = net.n-1;
- return get_network_output_layer(net, i);
+ return get_network_output_size_layer(net, i);
}
image get_network_image_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output;
+ return get_convolutional_image(layer);
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output;
+ return get_maxpool_image(layer);
}
- return make_image(0,0,0);
+ return make_empty_image(0,0,0);
}
image get_network_image(network net)
{
int i;
for(i = net.n-1; i >= 0; --i){
+ image m = get_network_image_layer(net, i);
+ if(m.h != 0) return m;
+ }
+ return make_empty_image(0,0,0);
+}
+
+void visualize_network(network net)
+{
+ 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];
- return layer.output;
+ visualize_convolutional_filters(layer, buff);
+ }
+ }
+}
+
+void print_network(network net)
+{
+ int i,j;
+ for(i = 0; i < net.n; ++i){
+ double *output;
+ 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];
- return layer.output;
+ 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");
}
- return make_image(1,1,1);
}
-
--
Gitblit v1.10.0