From 0f645836f193e75c4c3b718369e6fab15b5d19c5 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 11 Feb 2015 03:41:03 +0000
Subject: [PATCH] Detection is back, baby\!
---
src/network.c | 382 ++++++++++++++++++++++++++++++++++++++++++------------
1 files changed, 294 insertions(+), 88 deletions(-)
diff --git a/src/network.c b/src/network.c
index 3761bf9..bf0d63f 100644
--- a/src/network.c
+++ b/src/network.c
@@ -1,4 +1,5 @@
#include <stdio.h>
+#include <time.h>
#include "network.h"
#include "image.h"
#include "data.h"
@@ -7,11 +8,43 @@
#include "crop_layer.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
+#include "deconvolutional_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"
+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 NORMALIZATION:
+ return "normalization";
+ case DROPOUT:
+ return "dropout";
+ case FREEWEIGHT:
+ return "freeweight";
+ case CROP:
+ return "crop";
+ case COST:
+ return "cost";
+ default:
+ break;
+ }
+ return "none";
+}
+
network make_network(int n, int batch)
{
network net;
@@ -21,55 +54,15 @@
net.types = calloc(net.n, sizeof(LAYER_TYPE));
net.outputs = 0;
net.output = 0;
+ net.seen = 0;
#ifdef GPU
- net.input_cl = 0;
+ net.input_gpu = calloc(1, sizeof(float *));
+ net.truth_gpu = calloc(1, sizeof(float *));
#endif
return net;
}
-#ifdef GPU
-void forward_network_gpu(network net, cl_mem input_cl, int train)
-{
- 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_gpu(layer, input_cl);
- input_cl = layer.output_cl;
- }
- /*
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- forward_connected_layer(layer, input, train);
- 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] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_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;
- }
- else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- forward_normalization_layer(layer, input);
- input = layer.output;
- }
- */
- }
-}
-
-#endif
-
-void forward_network(network net, float *input, int train)
+void forward_network(network net, float *input, float *truth, int train)
{
int i;
for(i = 0; i < net.n; ++i){
@@ -78,6 +71,11 @@
forward_convolutional_layer(layer, input);
input = layer.output;
}
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ forward_deconvolutional_layer(layer, input);
+ input = layer.output;
+ }
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
forward_connected_layer(layer, input);
@@ -85,9 +83,13 @@
}
else if(net.types[i] == CROP){
crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_layer(layer, input);
+ forward_crop_layer(layer, train, 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);
@@ -107,7 +109,16 @@
if(!train) continue;
dropout_layer layer = *(dropout_layer *)net.layers[i];
forward_dropout_layer(layer, input);
+ input = layer.output;
}
+ else if(net.types[i] == FREEWEIGHT){
+ if(!train) continue;
+ //freeweight_layer layer = *(freeweight_layer *)net.layers[i];
+ //forward_freeweight_layer(layer, input);
+ }
+ //char buff[256];
+ //sprintf(buff, "layer %d", i);
+ //cuda_compare(get_network_output_gpu_layer(net, i), input, get_network_output_size_layer(net, i)*net.batch, buff);
}
}
@@ -119,14 +130,9 @@
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
update_convolutional_layer(layer);
}
- 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] == NORMALIZATION){
- //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ update_deconvolutional_layer(layer);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
@@ -140,6 +146,9 @@
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
return layer.output;
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ return layer.output;
} else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.output;
@@ -147,10 +156,16 @@
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.output;
} else if(net.types[i] == DROPOUT){
+ dropout_layer layer = *(dropout_layer *)net.layers[i];
+ return layer.output;
+ } 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] == CROP){
+ crop_layer layer = *(crop_layer *)net.layers[i];
+ return layer.output;
} else if(net.types[i] == NORMALIZATION){
normalization_layer layer = *(normalization_layer *)net.layers[i];
return layer.output;
@@ -159,7 +174,9 @@
}
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);
}
float *get_network_delta_layer(network net, int i)
@@ -167,6 +184,9 @@
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
return layer.delta;
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ return layer.delta;
} else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.delta;
@@ -174,6 +194,9 @@
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.delta;
} else if(net.types[i] == DROPOUT){
+ if(i == 0) return 0;
+ 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];
@@ -182,6 +205,14 @@
return 0;
}
+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);
@@ -212,9 +243,8 @@
return max_index(out, k);
}
-float backward_network(network net, float *input, float *truth)
+void backward_network(network net, float *input)
{
- float error = calculate_error_network(net, truth);
int i;
float *prev_input;
float *prev_delta;
@@ -226,13 +256,21 @@
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];
- backward_convolutional_layer(layer, prev_delta);
+ backward_convolutional_layer(layer, prev_input, prev_delta);
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ backward_deconvolutional_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] == DROPOUT){
+ dropout_layer layer = *(dropout_layer *)net.layers[i];
+ backward_dropout_layer(layer, prev_delta);
}
else if(net.types[i] == NORMALIZATION){
normalization_layer layer = *(normalization_layer *)net.layers[i];
@@ -240,23 +278,28 @@
}
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];
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);
+ }
}
- return error;
}
float train_network_datum(network net, float *x, float *y)
{
- forward_network(net, x, 1);
- //int class = get_predicted_class_network(net);
- float error = backward_network(net, x, y);
+ #ifdef GPU
+ if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
+ #endif
+ forward_network(net, x, y, 1);
+ backward_network(net, x);
+ float error = get_network_cost(net);
update_network(net);
- //return (y[class]?1:0);
return error;
}
@@ -269,7 +312,7 @@
int i;
float sum = 0;
for(i = 0; i < n; ++i){
- get_batch(d, batch, X, y);
+ get_random_batch(d, batch, X, y);
float err = train_network_datum(net, X, y);
sum += err;
}
@@ -277,6 +320,26 @@
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);
+ 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;
@@ -287,37 +350,91 @@
int index = rand()%d.X.rows;
float *x = d.X.vals[index];
float *y = d.y.vals[index];
- forward_network(net, x, 1);
- sum += backward_network(net, x, y);
+ forward_network(net, x, y, 1);
+ backward_network(net, x);
+ sum += get_network_cost(net);
}
update_network(net);
}
return (float)sum/(n*batch);
}
-
-void train_network(network net, data d)
+void set_learning_network(network *net, float rate, float momentum, float 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]);
- if(i%100 == 0){
- visualize_network(net);
- cvWaitKey(10);
+ net->learning_rate=rate;
+ net->momentum = momentum;
+ net->decay = decay;
+ for(i = 0; i < net->n; ++i){
+ if(net->types[i] == CONVOLUTIONAL){
+ convolutional_layer *layer = (convolutional_layer *)net->layers[i];
+ layer->learning_rate=rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
+ }
+ else if(net->types[i] == CONNECTED){
+ connected_layer *layer = (connected_layer *)net->layers[i];
+ layer->learning_rate=rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
}
}
- visualize_network(net);
- cvWaitKey(100);
- fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}
+
+void set_batch_network(network *net, int b)
+{
+ net->batch = b;
+ int i;
+ for(i = 0; i < net->n; ++i){
+ if(net->types[i] == CONVOLUTIONAL){
+ convolutional_layer *layer = (convolutional_layer *)net->layers[i];
+ layer->batch = b;
+ }else if(net->types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer *layer = (deconvolutional_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == MAXPOOL){
+ maxpool_layer *layer = (maxpool_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == CONNECTED){
+ connected_layer *layer = (connected_layer *)net->layers[i];
+ layer->batch = b;
+ } else if(net->types[i] == DROPOUT){
+ dropout_layer *layer = (dropout_layer *) net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == FREEWEIGHT){
+ freeweight_layer *layer = (freeweight_layer *) net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == SOFTMAX){
+ softmax_layer *layer = (softmax_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == COST){
+ cost_layer *layer = (cost_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == CROP){
+ crop_layer *layer = (crop_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ }
+}
+
+
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;
}
+ if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_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;
@@ -328,11 +445,19 @@
} else if(net.types[i] == DROPOUT){
dropout_layer layer = *(dropout_layer *) net.layers[i];
return layer.inputs;
+ } else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *) net.layers[i];
+ return layer.c*layer.h*layer.w;
+ }
+ 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;
}
+ printf("Can't find input size\n");
return 0;
}
@@ -343,22 +468,37 @@
image output = get_convolutional_image(layer);
return output.h*output.w*output.c;
}
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ image output = get_deconvolutional_image(layer);
+ return output.h*output.w*output.c;
+ }
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] == CROP){
+ crop_layer layer = *(crop_layer *) net.layers[i];
+ return layer.c*layer.crop_height*layer.crop_width;
+ }
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
- } else if(net.types[i] == DROPOUT){
+ }
+ 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;
}
+ printf("Can't find output size\n");
return 0;
}
@@ -368,21 +508,31 @@
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);
+ 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, c);
+ 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 if(net.types[i] == NORMALIZATION){
normalization_layer *layer = (normalization_layer *)net.layers[i];
- resize_normalization_layer(layer, h, w, c);
+ resize_normalization_layer(layer, h, w);
image output = get_normalization_image(*layer);
h = output.h;
w = output.w;
@@ -396,7 +546,8 @@
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);
}
@@ -411,6 +562,10 @@
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
return get_convolutional_image(layer);
}
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ return get_deconvolutional_image(layer);
+ }
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return get_maxpool_image(layer);
@@ -419,6 +574,9 @@
normalization_layer layer = *(normalization_layer *)net.layers[i];
return get_normalization_image(layer);
}
+ else if(net.types[i] == DROPOUT){
+ return get_network_image_layer(net, i-1);
+ }
else if(net.types[i] == CROP){
crop_layer layer = *(crop_layer *)net.layers[i];
return get_crop_image(layer);
@@ -441,7 +599,7 @@
image *prev = 0;
int i;
char buff[256];
- show_image(get_network_image_layer(net, 0), "Crop");
+ //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){
@@ -455,9 +613,21 @@
}
}
+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)
{
- forward_network(net, input, 0);
+ #ifdef GPU
+ if(gpu_index >= 0) return network_predict_gpu(net, input);
+ #endif
+
+ forward_network(net, input, 0, 0);
float *out = get_network_output(net);
return out;
}
@@ -492,7 +662,7 @@
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.rows, sizeof(float));
+ 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;
@@ -554,18 +724,54 @@
}
}
+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_accuracy(d.y, guess);
+ 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_accuracy(d.y, guess);
+ float acc = matrix_topk_accuracy(d.y, guess,1);
free_matrix(guess);
return acc;
}
--
Gitblit v1.10.0