From 1edcf73a73d2007afc61289245763f5cf0c29e10 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 04 Dec 2014 07:20:29 +0000
Subject: [PATCH] Detection good, split up col images
---
src/network.c | 388 +++++++++++++++++++++++++++++++++++++++++++++----------
1 files changed, 317 insertions(+), 71 deletions(-)
diff --git a/src/network.c b/src/network.c
index 07ac621..3a6a184 100644
--- a/src/network.c
+++ b/src/network.c
@@ -1,26 +1,38 @@
#include <stdio.h>
+#include <time.h>
#include "network.h"
#include "image.h"
#include "data.h"
#include "utils.h"
+#include "crop_layer.h"
#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"
-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 = calloc(1, sizeof(cl_mem));
+ net.truth_cl = calloc(1, sizeof(cl_mem));
+ #endif
return net;
}
-void forward_network(network net, double *input)
+
+void forward_network(network net, float *input, float *truth, int train)
{
int i;
for(i = 0; i < net.n; ++i){
@@ -34,6 +46,15 @@
forward_connected_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;
+ }
+ 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);
@@ -44,16 +65,31 @@
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;
+ }
+ else if(net.types[i] == DROPOUT){
+ if(!train) continue;
+ 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);
+ }
}
}
-void update_network(network net, double step, double momentum, double decay)
+void update_network(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(layer, step, momentum, decay);
+ update_convolutional_layer(layer);
}
else if(net.types[i] == MAXPOOL){
//maxpool_layer layer = *(maxpool_layer *)net.layers[i];
@@ -61,14 +97,17 @@
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, 0);
+ update_connected_layer(layer);
}
}
}
-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];
@@ -79,18 +118,27 @@
} else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
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;
+ } else if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ return layer.output;
}
return 0;
}
-double *get_network_output(network net)
+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);
}
-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];
@@ -101,6 +149,10 @@
} else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
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;
@@ -108,34 +160,49 @@
return 0;
}
-double *get_network_delta(network net)
+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);
}
-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);
- int i, k = get_network_output_size(net);
- for(i = 0; i < k; ++i){
+ float sum = 0;
+ float *delta = get_network_delta(net);
+ float *out = get_network_output(net);
+ int i;
+ for(i = 0; i < get_network_output_size(net)*net.batch; ++i){
+ //if(i %get_network_output_size(net) == 0) printf("\n");
+ //printf("%5.2f %5.2f, ", out[i], truth[i]);
+ //if(i == get_network_output_size(net)) printf("\n");
delta[i] = truth[i] - out[i];
+ //printf("%.10f, ", 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)
+void backward_network(network net, float *input)
{
- 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;
@@ -146,72 +213,106 @@
}
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);
+ 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);
+ if(i != 0) backward_maxpool_layer(layer, 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);
+ if(i != 0) backward_softmax_layer(layer, 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);
+ }
+ else if(net.types[i] == COST){
+ cost_layer layer = *(cost_layer *)net.layers[i];
+ backward_cost_layer(layer, prev_input, prev_delta);
}
}
}
-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)
{
- 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, y, 1);
+ //int class = get_predicted_class_network(net);
+ backward_network(net, x);
+ float error = get_network_cost(net);
+ update_network(net);
+ //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)
{
+ int batch = net.batch;
+ float *X = calloc(batch*d.X.cols, sizeof(float));
+ float *y = calloc(batch*d.y.cols, sizeof(float));
+
int i;
- int correct = 0;
+ float sum = 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);
- //if((i+1)%10 == 0){
- // printf("%d: %f\n", (i+1), (double)correct/(i+1));
- //}
+ get_random_batch(d, batch, X, y);
+ float err = train_network_datum(net, X, y);
+ sum += err;
}
- return (double)correct/n;
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
}
-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)
{
- int i;
- int correct = 0;
+ int i,j;
+ float sum = 0;
+ int batch = 2;
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);
+ for(j = 0; j < batch; ++j){
+ int index = rand()%d.X.rows;
+ float *x = d.X.vals[index];
+ float *y = d.y.vals[index];
+ forward_network(net, x, y, 1);
+ backward_network(net, x);
+ sum += get_network_cost(net);
+ }
+ update_network(net);
}
- update_network(net, step, momentum, decay);
- return (double)correct/n;
-
+ return (float)sum/(n*batch);
}
+float train_network_data_cpu(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));
-void train_network(network net, data d, double step, double momentum, double decay)
+ 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(net, X, y);
+ sum += err;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+
+void train_network(network net, data d)
{
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);
+ correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]);
if(i%100 == 0){
visualize_network(net);
cvWaitKey(10);
@@ -219,7 +320,35 @@
}
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_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] == 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;
+ }
+ return 0;
}
int get_network_output_size_layer(network net, int i)
@@ -238,6 +367,14 @@
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
}
+ 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;
@@ -245,12 +382,50 @@
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 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;
+}
+
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);
}
+int get_network_input_size(network net)
+{
+ return get_network_input_size_layer(net, 0);
+}
+
image get_network_image_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
@@ -261,6 +436,14 @@
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);
+ }
+ else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *)net.layers[i];
+ return get_crop_image(layer);
+ }
return make_empty_image(0,0,0);
}
@@ -276,35 +459,83 @@
void visualize_network(network net)
{
+ image *prev = 0;
int i;
char buff[256];
+ //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){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- visualize_convolutional_filters(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);
}
}
}
-double *network_predict(network net, double *input)
+void top_predictions(network net, int k, int *index)
{
- forward_network(net, input);
- double *out = get_network_output(net);
+ int size = get_network_output_size(net);
+ float *out = get_network_output(net);
+ top_k(out, size, k, index);
+}
+
+
+float *network_predict(network net, float *input)
+{
+ forward_network(net, input, 0, 0);
+ float *out = get_network_output(net);
return out;
}
+matrix network_predict_data_multi(network net, data test, int n)
+{
+ int i,j,b,m;
+ int k = get_network_output_size(net);
+ matrix pred = make_matrix(test.X.rows, k);
+ float *X = calloc(net.batch*test.X.rows, sizeof(float));
+ for(i = 0; i < test.X.rows; i += net.batch){
+ for(b = 0; b < net.batch; ++b){
+ if(i+b == test.X.rows) break;
+ memcpy(X+b*test.X.cols, test.X.vals[i+b], test.X.cols*sizeof(float));
+ }
+ for(m = 0; m < n; ++m){
+ float *out = network_predict(net, X);
+ for(b = 0; b < net.batch; ++b){
+ if(i+b == test.X.rows) break;
+ for(j = 0; j < k; ++j){
+ pred.vals[i+b][j] += out[j+b*k]/n;
+ }
+ }
+ }
+ }
+ free(X);
+ return pred;
+}
+
matrix network_predict_data(network net, data test)
{
- int i,j;
+ int i,j,b;
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];
+ float *X = calloc(net.batch*test.X.cols, sizeof(float));
+ for(i = 0; i < test.X.rows; i += net.batch){
+ for(b = 0; b < net.batch; ++b){
+ if(i+b == test.X.rows) break;
+ memcpy(X+b*test.X.cols, test.X.vals[i+b], test.X.cols*sizeof(float));
+ }
+ float *out = network_predict(net, X);
+ for(b = 0; b < net.batch; ++b){
+ if(i+b == test.X.rows) break;
+ for(j = 0; j < k; ++j){
+ pred.vals[i+b][j] = out[j+b*k];
+ }
}
}
+ free(X);
return pred;
}
@@ -312,7 +543,7 @@
{
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];
@@ -326,6 +557,12 @@
image m = get_maxpool_image(layer);
n = m.h*m.w*m.c;
}
+ else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *)net.layers[i];
+ output = layer.output;
+ image m = get_crop_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;
@@ -336,8 +573,8 @@
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]);
@@ -346,11 +583,20 @@
}
}
-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;
}
+float network_accuracy_multi(network net, data d, int n)
+{
+ matrix guess = network_predict_data_multi(net, d, n);
+ float acc = matrix_accuracy(d.y, guess);
+ free_matrix(guess);
+ return acc;
+}
+
+
--
Gitblit v1.10.0