From 989ab8c38a02fa7ea9c25108151736c62e81c972 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 24 Apr 2015 17:27:50 +0000
Subject: [PATCH] IOU loss function
---
src/network.c | 235 +++++++++++++++++++++-------------------------------------
1 files changed, 86 insertions(+), 149 deletions(-)
diff --git a/src/network.c b/src/network.c
index b60f059..3247a31 100644
--- a/src/network.c
+++ b/src/network.c
@@ -4,6 +4,7 @@
#include "image.h"
#include "data.h"
#include "utils.h"
+#include "params.h"
#include "crop_layer.h"
#include "connected_layer.h"
@@ -13,7 +14,6 @@
#include "maxpool_layer.h"
#include "cost_layer.h"
#include "normalization_layer.h"
-#include "freeweight_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
@@ -36,8 +36,6 @@
return "normalization";
case DROPOUT:
return "dropout";
- case FREEWEIGHT:
- return "freeweight";
case CROP:
return "crop";
case COST:
@@ -48,16 +46,18 @@
return "none";
}
-network make_network(int n, int batch)
+network make_network(int n)
{
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;
net.seen = 0;
+ net.batch = 0;
+ net.inputs = 0;
+ net.h = net.w = net.c = 0;
#ifdef GPU
net.input_gpu = calloc(1, sizeof(float *));
net.truth_gpu = calloc(1, sizeof(float *));
@@ -65,86 +65,60 @@
return net;
}
-void forward_network(network net, float *input, float *truth, int train)
+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;
+ forward_convolutional_layer(*(convolutional_layer *)net.layers[i], state);
}
else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- forward_deconvolutional_layer(layer, input);
- input = layer.output;
+ forward_deconvolutional_layer(*(deconvolutional_layer *)net.layers[i], state);
}
else if(net.types[i] == DETECTION){
- detection_layer layer = *(detection_layer *)net.layers[i];
- forward_detection_layer(layer, input, truth);
- input = layer.output;
+ forward_detection_layer(*(detection_layer *)net.layers[i], state);
}
else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- forward_connected_layer(layer, input);
- input = layer.output;
+ forward_connected_layer(*(connected_layer *)net.layers[i], state);
}
else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_layer(layer, train, input);
- input = layer.output;
+ forward_crop_layer(*(crop_layer *)net.layers[i], state);
}
else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- forward_cost_layer(layer, input, truth);
+ forward_cost_layer(*(cost_layer *)net.layers[i], state);
}
else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- forward_softmax_layer(layer, input);
- input = layer.output;
+ forward_softmax_layer(*(softmax_layer *)net.layers[i], state);
}
else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- forward_maxpool_layer(layer, input);
- input = layer.output;
+ forward_maxpool_layer(*(maxpool_layer *)net.layers[i], state);
}
else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- forward_normalization_layer(layer, input);
- input = layer.output;
+ forward_normalization_layer(*(normalization_layer *)net.layers[i], state);
}
else if(net.types[i] == DROPOUT){
- if(!train) continue;
- dropout_layer layer = *(dropout_layer *)net.layers[i];
- forward_dropout_layer(layer, input);
- input = layer.output;
+ forward_dropout_layer(*(dropout_layer *)net.layers[i], state);
}
- 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);
+ state.input = get_network_output_layer(net, i);
}
}
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);
+ update_convolutional_layer(layer, update_batch, net.learning_rate, net.momentum, net.decay);
}
else if(net.types[i] == DECONVOLUTIONAL){
deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- update_deconvolutional_layer(layer);
+ update_deconvolutional_layer(layer, net.learning_rate, net.momentum, net.decay);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer(layer);
+ update_connected_layer(layer, update_batch, net.learning_rate, net.momentum, net.decay);
}
}
}
@@ -152,37 +126,27 @@
float *get_network_output_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output;
+ return ((convolutional_layer *)net.layers[i]) -> output;
} else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- return layer.output;
+ return ((deconvolutional_layer *)net.layers[i]) -> output;
} else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output;
+ return ((maxpool_layer *)net.layers[i]) -> output;
} else if(net.types[i] == DETECTION){
- detection_layer layer = *(detection_layer *)net.layers[i];
- return layer.output;
+ return ((detection_layer *)net.layers[i]) -> output;
} else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.output;
+ return ((softmax_layer *)net.layers[i]) -> 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;
+ return ((connected_layer *)net.layers[i]) -> output;
} else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- return layer.output;
+ return ((crop_layer *)net.layers[i]) -> output;
} else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- return layer.output;
+ return ((normalization_layer *)net.layers[i]) -> output;
}
return 0;
}
+
float *get_network_output(network net)
{
int i;
@@ -210,8 +174,6 @@
} 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];
return layer.delta;
@@ -224,6 +186,9 @@
if(net.types[net.n-1] == COST){
return ((cost_layer *)net.layers[net.n-1])->output[0];
}
+ if(net.types[net.n-1] == DETECTION){
+ return ((detection_layer *)net.layers[net.n-1])->cost[0];
+ }
return 0;
}
@@ -232,24 +197,6 @@
return get_network_delta_layer(net, net.n-1);
}
-float calculate_error_network(network net, float *truth)
-{
- 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)
{
float *out = get_network_output(net);
@@ -257,54 +204,53 @@
return max_index(out, k);
}
-void backward_network(network net, float *input, float *truth)
+void backward_network(network net, network_state state)
{
int i;
- float *prev_input;
- float *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);
+ state.input = get_network_output_layer(net, i-1);
+ state.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_input, prev_delta);
+ backward_convolutional_layer(layer, state);
} else if(net.types[i] == DECONVOLUTIONAL){
deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- backward_deconvolutional_layer(layer, prev_input, prev_delta);
+ backward_deconvolutional_layer(layer, state);
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- if(i != 0) backward_maxpool_layer(layer, prev_delta);
+ if(i != 0) backward_maxpool_layer(layer, state);
}
else if(net.types[i] == DROPOUT){
dropout_layer layer = *(dropout_layer *)net.layers[i];
- backward_dropout_layer(layer, prev_delta);
+ backward_dropout_layer(layer, state);
}
else if(net.types[i] == DETECTION){
detection_layer layer = *(detection_layer *)net.layers[i];
- backward_detection_layer(layer, prev_input, prev_delta);
+ backward_detection_layer(layer, state);
}
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);
+ if(i != 0) backward_normalization_layer(layer, state);
}
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
- if(i != 0) backward_softmax_layer(layer, prev_delta);
+ if(i != 0) backward_softmax_layer(layer, state);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- backward_connected_layer(layer, prev_input, prev_delta);
+ backward_connected_layer(layer, state);
}
else if(net.types[i] == COST){
cost_layer layer = *(cost_layer *)net.layers[i];
- backward_cost_layer(layer, prev_input, prev_delta);
+ backward_cost_layer(layer, state);
}
}
}
@@ -314,10 +260,14 @@
#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, y);
+ 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);
- update_network(net);
+ if((net.seen/net.batch)%net.subdivisions == 0) update_network(net);
return error;
}
@@ -330,6 +280,7 @@
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;
@@ -350,6 +301,7 @@
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;
}
@@ -361,15 +313,17 @@
float train_network_batch(network net, data d, int n)
{
int i,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;
- float *x = d.X.vals[index];
- float *y = d.y.vals[index];
- forward_network(net, x, y, 1);
- backward_network(net, x, y);
+ 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);
}
update_network(net);
@@ -377,28 +331,6 @@
return (float)sum/(n*batch);
}
-void set_learning_network(network *net, float rate, float momentum, float decay)
-{
- int i;
- 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;
- }
- }
-}
-
void set_batch_network(network *net, int b)
{
net->batch = b;
@@ -425,10 +357,6 @@
detection_layer *layer = (detection_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;
@@ -472,15 +400,11 @@
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");
+ fprintf(stderr, "Can't find input size\n");
return 0;
}
@@ -505,7 +429,7 @@
image output = get_maxpool_image(layer);
return output.h*output.w*output.c;
}
- else if(net.types[i] == CROP){
+ else if(net.types[i] == CROP){
crop_layer layer = *(crop_layer *) net.layers[i];
return layer.c*layer.crop_height*layer.crop_width;
}
@@ -517,15 +441,11 @@
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");
+ fprintf(stderr, "Can't find output size\n");
return 0;
}
@@ -583,6 +503,18 @@
return get_network_input_size_layer(net, 0);
}
+detection_layer *get_network_detection_layer(network net)
+{
+ int i;
+ for(i = 0; i < net.n; ++i){
+ if(net.types[i] == DETECTION){
+ detection_layer *layer = (detection_layer *)net.layers[i];
+ return layer;
+ }
+ }
+ return 0;
+}
+
image get_network_image_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
@@ -650,11 +582,16 @@
float *network_predict(network net, float *input)
{
- #ifdef GPU
+#ifdef GPU
if(gpu_index >= 0) return network_predict_gpu(net, input);
- #endif
+#endif
- forward_network(net, input, 0, 0);
+ 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;
}
--
Gitblit v1.10.0