From 2ea63c0e99a5358eaf38785ea83b9c5923fcc9cd Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 13 Mar 2014 04:57:34 +0000
Subject: [PATCH] Better VOC handling and resizing
---
src/network.c | 238 ++++++++++++++++++++++++++++++++++++++++++++++++----------
1 files changed, 195 insertions(+), 43 deletions(-)
diff --git a/src/network.c b/src/network.c
index 07ac621..e2c44b0 100644
--- a/src/network.c
+++ b/src/network.c
@@ -6,13 +6,15 @@
#include "connected_layer.h"
#include "convolutional_layer.h"
+//#include "old_conv.h"
#include "maxpool_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;
@@ -20,7 +22,79 @@
return net;
}
-void forward_network(network net, double *input)
+void print_convolutional_cfg(FILE *fp, convolutional_layer *l, int first)
+{
+ int i;
+ 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->n, l->size, l->stride,
+ get_activation_string(l->activation));
+ fprintf(fp, "data=");
+ for(i = 0; i < l->n; ++i) fprintf(fp, "%g,", l->biases[i]);
+ 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, int first)
+{
+ int i;
+ 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, int first)
+{
+ 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, 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)
+{
+ FILE *fp = fopen(filename, "w");
+ if(!fp) file_error(filename);
+ int i;
+ for(i = 0; i < net.n; ++i)
+ {
+ if(net.types[i] == CONVOLUTIONAL)
+ 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], i==0);
+ else if(net.types[i] == MAXPOOL)
+ print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i], i==0);
+ else if(net.types[i] == SOFTMAX)
+ print_softmax_cfg(fp, (softmax_layer *)net.layers[i], i==0);
+ }
+ fclose(fp);
+}
+
+void forward_network(network net, float *input)
{
int i;
for(i = 0; i < net.n; ++i){
@@ -47,7 +121,7 @@
}
}
-void update_network(network net, double step, double momentum, double decay)
+void update_network(network net, float step, float momentum, float decay)
{
int i;
for(i = 0; i < net.n; ++i){
@@ -63,12 +137,12 @@
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer(layer, step, momentum, 0);
+ update_connected_layer(layer, step, momentum, decay);
}
}
}
-double *get_network_output_layer(network net, int i)
+float *get_network_output_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
@@ -85,12 +159,12 @@
}
return 0;
}
-double *get_network_output(network net)
+float *get_network_output(network net)
{
return get_network_output_layer(net, net.n-1);
}
-double *get_network_delta_layer(network net, int i)
+float *get_network_delta_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
@@ -108,34 +182,39 @@
return 0;
}
-double *get_network_delta(network net)
+float *get_network_delta(network net)
{
return get_network_delta_layer(net, net.n-1);
}
-void calculate_error_network(network net, double *truth)
+float calculate_error_network(network net, float *truth)
{
- double *delta = get_network_delta(net);
- double *out = get_network_output(net);
+ float sum = 0;
+ float *delta = get_network_delta(net);
+ float *out = get_network_output(net);
int i, k = get_network_output_size(net);
for(i = 0; i < k; ++i){
+ //printf("%f, ", out[i]);
delta[i] = truth[i] - out[i];
+ sum += delta[i]*delta[i];
}
+ //printf("\n");
+ return sum;
}
int get_predicted_class_network(network net)
{
- double *out = get_network_output(net);
+ float *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)
+float backward_network(network net, float *input, float *truth)
{
- calculate_error_network(net, truth);
+ float error = calculate_error_network(net, truth);
int i;
- double *prev_input;
- double *prev_delta;
+ float *prev_input;
+ float *prev_delta;
for(i = net.n-1; i >= 0; --i){
if(i == 0){
prev_input = input;
@@ -146,8 +225,9 @@
}
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);
+ learn_convolutional_layer(layer);
+ //learn_convolutional_layer(layer);
+ if(i != 0) backward_convolutional_layer(layer, prev_delta);
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
@@ -163,50 +243,65 @@
if(i != 0) backward_connected_layer(layer, prev_input, prev_delta);
}
}
+ return error;
}
-int train_network_datum(network net, double *x, double *y, double step, double momentum, double decay)
+float train_network_datum(network net, float *x, float *y, float step, float momentum, float decay)
{
- 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);
+ forward_network(net, x);
+ //int class = get_predicted_class_network(net);
+ float error = backward_network(net, x, y);
+ update_network(net, step, momentum, decay);
+ //return (y[class]?1:0);
+ return error;
}
-double train_network_sgd(network net, data d, int n, double step, double momentum,double decay)
+float train_network_sgd(network net, data d, int n, float step, float momentum,float decay)
{
int i;
+ float error = 0;
int correct = 0;
+ int pos = 0;
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);
+ 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), (double)correct/(i+1));
+ // printf("%d: %f\n", (i+1), (float)correct/(i+1));
//}
}
- return (double)correct/n;
+ //printf("Accuracy: %f\n",(float) correct/n);
+ return error/pos;
}
-double train_network_batch(network net, data d, int n, double step, double momentum,double decay)
+float train_network_batch(network net, data d, int n, float step, float momentum,float 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];
+ float *x = d.X.vals[index];
+ float *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;
+ return (float)correct/n;
}
-void train_network(network net, data d, double step, double momentum, double decay)
+void train_network(network net, data d, float step, float momentum, float decay)
{
int i;
int correct = 0;
@@ -219,7 +314,7 @@
}
visualize_network(net);
cvWaitKey(100);
- printf("Accuracy: %f\n", (double)correct/d.X.rows);
+ fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}
int get_network_output_size_layer(network net, int i)
@@ -245,6 +340,63 @@
return 0;
}
+/*
+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;
+ image output = get_convolutional_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
+ }
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer *layer = (maxpool_layer *)net.layers[i];
+ layer->h = h;
+ layer->w = w;
+ layer->c = c;
+ image output = get_maxpool_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
+ }
+ }
+ return 0;
+}
+*/
+
+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];
+ 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){
+ maxpool_layer *layer = (maxpool_layer *)net.layers[i];
+ resize_maxpool_layer(layer, h, w, c);
+ image output = get_maxpool_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
+ }
+ else{
+ error("Cannot resize this type of layer");
+ }
+ }
+ return 0;
+}
+
int get_network_output_size(network net)
{
int i = net.n-1;
@@ -282,15 +434,15 @@
sprintf(buff, "Layer %d", i);
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- visualize_convolutional_filters(layer, buff);
+ visualize_convolutional_layer(layer, buff);
}
}
}
-double *network_predict(network net, double *input)
+float *network_predict(network net, float *input)
{
forward_network(net, input);
- double *out = get_network_output(net);
+ float *out = get_network_output(net);
return out;
}
@@ -300,7 +452,7 @@
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]);
+ float *out = network_predict(net, test.X.vals[i]);
for(j = 0; j < k; ++j){
pred.vals[i][j] = out[j];
}
@@ -312,7 +464,7 @@
{
int i,j;
for(i = 0; i < net.n; ++i){
- double *output = 0;
+ float *output = 0;
int n = 0;
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
@@ -336,8 +488,8 @@
output = layer.output;
n = layer.inputs;
}
- double mean = mean_array(output, n);
- double vari = variance_array(output, n);
+ float mean = mean_array(output, n);
+ float 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]);
@@ -346,10 +498,10 @@
}
}
-double network_accuracy(network net, data d)
+float network_accuracy(network net, data d)
{
matrix guess = network_predict_data(net, d);
- double acc = matrix_accuracy(d.y, guess);
+ float acc = matrix_accuracy(d.y, guess);
free_matrix(guess);
return acc;
}
--
Gitblit v1.10.0