From 11c72b1132feca7c1252ea01d02da4cb497e723f Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 11 Jun 2015 22:38:58 +0000
Subject: [PATCH] testing on one image
---
src/network.c | 570 ++++++++++++++++++++++++++++++++++++++------------------
1 files changed, 388 insertions(+), 182 deletions(-)
diff --git a/src/network.c b/src/network.c
index faedb8c..68790e5 100644
--- a/src/network.c
+++ b/src/network.c
@@ -1,226 +1,325 @@
#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 "deconvolutional_layer.h"
+#include "detection_layer.h"
#include "maxpool_layer.h"
+#include "cost_layer.h"
#include "softmax_layer.h"
+#include "dropout_layer.h"
+#include "route_layer.h"
+
+char *get_layer_string(LAYER_TYPE a)
+{
+ switch(a){
+ case CONVOLUTIONAL:
+ return "convolutional";
+ case DECONVOLUTIONAL:
+ return "deconvolutional";
+ case CONNECTED:
+ return "connected";
+ case MAXPOOL:
+ return "maxpool";
+ case SOFTMAX:
+ return "softmax";
+ case DETECTION:
+ return "detection";
+ case DROPOUT:
+ return "dropout";
+ case CROP:
+ return "crop";
+ case COST:
+ return "cost";
+ case ROUTE:
+ return "route";
+ default:
+ break;
+ }
+ return "none";
+}
network make_network(int n)
{
- network net;
+ network net = {0};
net.n = n;
- net.layers = calloc(net.n, sizeof(void *));
- net.types = calloc(net.n, sizeof(LAYER_TYPE));
+ net.layers = calloc(net.n, sizeof(layer));
+ #ifdef GPU
+ net.input_gpu = calloc(1, sizeof(float *));
+ net.truth_gpu = calloc(1, sizeof(float *));
+ #endif
return net;
}
-void forward_network(network net, double *input)
+void forward_network(network net, network_state state)
{
int i;
for(i = 0; i < net.n; ++i){
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- forward_convolutional_layer(layer, input);
- input = layer.output;
+ layer l = net.layers[i];
+ if(l.type == CONVOLUTIONAL){
+ forward_convolutional_layer(l, state);
+ } else if(l.type == DECONVOLUTIONAL){
+ forward_deconvolutional_layer(l, state);
+ } else if(l.type == DETECTION){
+ forward_detection_layer(l, state);
+ } else if(l.type == CONNECTED){
+ forward_connected_layer(l, state);
+ } else if(l.type == CROP){
+ forward_crop_layer(l, state);
+ } else if(l.type == COST){
+ forward_cost_layer(l, state);
+ } else if(l.type == SOFTMAX){
+ forward_softmax_layer(l, state);
+ } else if(l.type == MAXPOOL){
+ forward_maxpool_layer(l, state);
+ } else if(l.type == DROPOUT){
+ forward_dropout_layer(l, state);
+ } else if(l.type == ROUTE){
+ forward_route_layer(l, net);
}
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- 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);
- input = layer.output;
- }
+ state.input = l.output;
}
}
-void update_network(network net, double step)
+void update_network(network net)
{
int i;
+ int update_batch = net.batch*net.subdivisions;
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, 0.9, .01);
- }
- 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, .9, 0);
+ layer l = net.layers[i];
+ if(l.type == CONVOLUTIONAL){
+ update_convolutional_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
+ } else if(l.type == DECONVOLUTIONAL){
+ update_deconvolutional_layer(l, net.learning_rate, net.momentum, net.decay);
+ } else if(l.type == CONNECTED){
+ update_connected_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
}
}
}
-double *get_network_output_layer(network net, int i)
+float *get_network_output(network net)
{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output;
- } 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;
- }
- return 0;
-}
-double *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.layers[i].type != COST) break;
+ return net.layers[i].output;
}
-double *get_network_delta_layer(network net, int i)
+float get_network_cost(network net)
{
- 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;
+ if(net.layers[net.n-1].type == COST){
+ return net.layers[net.n-1].output[0];
+ }
+ if(net.layers[net.n-1].type == DETECTION){
+ return net.layers[net.n-1].cost[0];
}
return 0;
}
-double *get_network_delta(network net)
+int get_predicted_class_network(network net)
{
- return get_network_delta_layer(net, net.n-1);
+ float *out = get_network_output(net);
+ int k = get_network_output_size(net);
+ return max_index(out, k);
}
-void learn_network(network net, double *input)
+void backward_network(network net, network_state state)
{
int i;
- double *prev_input;
- double *prev_delta;
+ float *original_input = state.input;
for(i = net.n-1; i >= 0; --i){
if(i == 0){
- prev_input = input;
- prev_delta = 0;
+ state.input = original_input;
+ state.delta = 0;
}else{
- prev_input = get_network_output_layer(net, i-1);
- prev_delta = get_network_delta_layer(net, i-1);
+ layer prev = net.layers[i-1];
+ state.input = prev.output;
+ state.delta = prev.delta;
}
- 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);
+ layer l = net.layers[i];
+ if(l.type == CONVOLUTIONAL){
+ backward_convolutional_layer(l, state);
+ } else if(l.type == DECONVOLUTIONAL){
+ backward_deconvolutional_layer(l, state);
+ } else if(l.type == MAXPOOL){
+ if(i != 0) backward_maxpool_layer(l, state);
+ } else if(l.type == DROPOUT){
+ backward_dropout_layer(l, state);
+ } else if(l.type == DETECTION){
+ backward_detection_layer(l, state);
+ } else if(l.type == SOFTMAX){
+ if(i != 0) backward_softmax_layer(l, state);
+ } else if(l.type == CONNECTED){
+ backward_connected_layer(l, state);
+ } else if(l.type == COST){
+ backward_cost_layer(l, state);
+ } else if(l.type == ROUTE){
+ backward_route_layer(l, net);
}
}
}
-void train_network_batch(network net, batch b)
+float train_network_datum(network net, float *x, float *y)
+{
+ #ifdef GPU
+ if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
+ #endif
+ network_state state;
+ state.input = x;
+ state.truth = y;
+ state.train = 1;
+ forward_network(net, state);
+ backward_network(net, state);
+ float error = get_network_cost(net);
+ if((net.seen/net.batch)%net.subdivisions == 0) update_network(net);
+ return error;
+}
+
+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;
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ net.seen += batch;
+ get_random_batch(d, batch, X, y);
+ float err = train_network_datum(net, X, y);
+ sum += err;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+
+float train_network(network net, data d)
+{
+ int batch = net.batch;
+ int n = d.X.rows / batch;
+ float *X = calloc(batch*d.X.cols, sizeof(float));
+ float *y = calloc(batch*d.y.cols, sizeof(float));
+
+ int i;
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ get_next_batch(d, batch, i*batch, X, y);
+ net.seen += batch;
+ float err = train_network_datum(net, X, y);
+ sum += err;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+
+float train_network_batch(network net, data d, int n)
{
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;
- }
+ network_state state;
+ state.train = 1;
+ float sum = 0;
+ int batch = 2;
+ for(i = 0; i < n; ++i){
+ for(j = 0; j < batch; ++j){
+ int index = rand()%d.X.rows;
+ state.input = d.X.vals[index];
+ state.truth = d.y.vals[index];
+ forward_network(net, state);
+ backward_network(net, state);
+ sum += get_network_cost(net);
}
- 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);
- }
+ update_network(net);
}
- visualize_network(net);
- print_network(net);
- cvWaitKey(100);
- printf("Accuracy: %f\n", (double)correct/b.n);
+ return (float)sum/(n*batch);
}
-int get_network_output_size_layer(network net, int i)
+void set_batch_network(network *net, int b)
{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- image output = get_convolutional_image(layer);
- return output.h*output.w*output.c;
+ net->batch = b;
+ int i;
+ for(i = 0; i < net->n; ++i){
+ net->layers[i].batch = b;
}
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- 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;
+}
+
+/*
+int resize_network(network net, int h, int w, int c)
+{
+ fprintf(stderr, "Might be broken, careful!!");
+ 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);
+ image output = get_convolutional_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer *layer = (deconvolutional_layer *)net.layers[i];
+ resize_deconvolutional_layer(layer, h, w);
+ image output = get_deconvolutional_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);
+ image output = get_maxpool_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
+ }else if(net.types[i] == DROPOUT){
+ dropout_layer *layer = (dropout_layer *)net.layers[i];
+ resize_dropout_layer(layer, h*w*c);
+ }else{
+ error("Cannot resize this type of layer");
+ }
}
return 0;
}
+*/
int get_network_output_size(network net)
{
- int i = net.n-1;
- return get_network_output_size_layer(net, i);
+ int i;
+ for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
+ return net.layers[i].outputs;
+}
+
+int get_network_input_size(network net)
+{
+ return net.layers[0].inputs;
+}
+
+detection_layer get_network_detection_layer(network net)
+{
+ int i;
+ for(i = 0; i < net.n; ++i){
+ if(net.layers[i].type == DETECTION){
+ return net.layers[i];
+ }
+ }
+ fprintf(stderr, "Detection layer not found!!\n");
+ detection_layer l = {0};
+ return l;
}
image get_network_image_layer(network net, int i)
{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return get_convolutional_image(layer);
+ layer l = net.layers[i];
+ if (l.out_w && l.out_h && l.out_c){
+ return float_to_image(l.out_w, l.out_h, l.out_c, l.output);
}
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return get_maxpool_image(layer);
- }
- return make_empty_image(0,0,0);
+ image def = {0};
+ return def;
}
image get_network_image(network net)
@@ -230,52 +329,105 @@
image m = get_network_image_layer(net, i);
if(m.h != 0) return m;
}
- return make_empty_image(0,0,0);
+ image def = {0};
+ return def;
}
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_filters(layer, buff);
+ layer l = net.layers[i];
+ if(l.type == CONVOLUTIONAL){
+ prev = visualize_convolutional_layer(l, buff, prev);
}
}
}
+void top_predictions(network net, int k, int *index)
+{
+ 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)
+{
+#ifdef GPU
+ if(gpu_index >= 0) return network_predict_gpu(net, input);
+#endif
+
+ network_state state;
+ state.input = input;
+ state.truth = 0;
+ state.train = 0;
+ state.delta = 0;
+ forward_network(net, state);
+ 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,b;
+ int k = get_network_output_size(net);
+ matrix pred = make_matrix(test.X.rows, k);
+ 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;
+}
+
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];
- 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);
+ layer l = net.layers[i];
+ float *output = l.output;
+ int n = l.outputs;
+ 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]);
@@ -283,3 +435,57 @@
fprintf(stderr, "\n");
}
}
+
+void compare_networks(network n1, network n2, data test)
+{
+ matrix g1 = network_predict_data(n1, test);
+ matrix g2 = network_predict_data(n2, test);
+ int i;
+ int a,b,c,d;
+ a = b = c = d = 0;
+ for(i = 0; i < g1.rows; ++i){
+ int truth = max_index(test.y.vals[i], test.y.cols);
+ int p1 = max_index(g1.vals[i], g1.cols);
+ int p2 = max_index(g2.vals[i], g2.cols);
+ if(p1 == truth){
+ if(p2 == truth) ++d;
+ else ++c;
+ }else{
+ if(p2 == truth) ++b;
+ else ++a;
+ }
+ }
+ printf("%5d %5d\n%5d %5d\n", a, b, c, d);
+ float num = pow((abs(b - c) - 1.), 2.);
+ float den = b + c;
+ printf("%f\n", num/den);
+}
+
+float network_accuracy(network net, data d)
+{
+ matrix guess = network_predict_data(net, d);
+ float acc = matrix_topk_accuracy(d.y, guess,1);
+ free_matrix(guess);
+ return acc;
+}
+
+float *network_accuracies(network net, data d)
+{
+ static float acc[2];
+ matrix guess = network_predict_data(net, d);
+ acc[0] = matrix_topk_accuracy(d.y, guess,1);
+ acc[1] = matrix_topk_accuracy(d.y, guess,5);
+ 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_topk_accuracy(d.y, guess,1);
+ free_matrix(guess);
+ return acc;
+}
+
+
--
Gitblit v1.10.0