From 064d4462d3890fd5dda5c7845119e1a3854d477b Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 26 Nov 2015 21:45:12 +0000
Subject: [PATCH] stuff
---
src/network.c | 90 +++++++++++++++++++++++++++++++++++----------
1 files changed, 70 insertions(+), 20 deletions(-)
diff --git a/src/network.c b/src/network.c
index 70bcb58..d9585c4 100644
--- a/src/network.c
+++ b/src/network.c
@@ -8,10 +8,10 @@
#include "crop_layer.h"
#include "connected_layer.h"
+#include "local_layer.h"
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "detection_layer.h"
-#include "region_layer.h"
#include "normalization_layer.h"
#include "maxpool_layer.h"
#include "avgpool_layer.h"
@@ -20,11 +20,60 @@
#include "dropout_layer.h"
#include "route_layer.h"
+int get_current_batch(network net)
+{
+ int batch_num = (*net.seen)/(net.batch*net.subdivisions);
+ return batch_num;
+}
+
+void reset_momentum(network net)
+{
+ if (net.momentum == 0) return;
+ net.learning_rate = 0;
+ net.momentum = 0;
+ net.decay = 0;
+ #ifdef GPU
+ if(gpu_index >= 0) update_network_gpu(net);
+ #endif
+}
+
+float get_current_rate(network net)
+{
+ int batch_num = get_current_batch(net);
+ int i;
+ float rate;
+ switch (net.policy) {
+ case CONSTANT:
+ return net.learning_rate;
+ case STEP:
+ return net.learning_rate * pow(net.scale, batch_num/net.step);
+ case STEPS:
+ rate = net.learning_rate;
+ for(i = 0; i < net.num_steps; ++i){
+ if(net.steps[i] > batch_num) return rate;
+ rate *= net.scales[i];
+ if(net.steps[i] > batch_num - 1) reset_momentum(net);
+ }
+ return rate;
+ case EXP:
+ return net.learning_rate * pow(net.gamma, batch_num);
+ case POLY:
+ return net.learning_rate * pow(1 - (float)batch_num / net.max_batches, net.power);
+ case SIG:
+ return net.learning_rate * (1./(1.+exp(net.gamma*(batch_num - net.step))));
+ default:
+ fprintf(stderr, "Policy is weird!\n");
+ return net.learning_rate;
+ }
+}
+
char *get_layer_string(LAYER_TYPE a)
{
switch(a){
case CONVOLUTIONAL:
return "convolutional";
+ case LOCAL:
+ return "local";
case DECONVOLUTIONAL:
return "deconvolutional";
case CONNECTED:
@@ -37,8 +86,6 @@
return "softmax";
case DETECTION:
return "detection";
- case REGION:
- return "region";
case DROPOUT:
return "dropout";
case CROP:
@@ -60,6 +107,7 @@
network net = {0};
net.n = n;
net.layers = calloc(net.n, sizeof(layer));
+ net.seen = calloc(1, sizeof(int));
#ifdef GPU
net.input_gpu = calloc(1, sizeof(float *));
net.truth_gpu = calloc(1, sizeof(float *));
@@ -79,12 +127,12 @@
forward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
forward_deconvolutional_layer(l, state);
+ } else if(l.type == LOCAL){
+ forward_local_layer(l, state);
} else if(l.type == NORMALIZATION){
forward_normalization_layer(l, state);
} else if(l.type == DETECTION){
forward_detection_layer(l, state);
- } else if(l.type == REGION){
- forward_region_layer(l, state);
} else if(l.type == CONNECTED){
forward_connected_layer(l, state);
} else if(l.type == CROP){
@@ -110,14 +158,17 @@
{
int i;
int update_batch = net.batch*net.subdivisions;
+ float rate = get_current_rate(net);
for(i = 0; i < net.n; ++i){
layer l = net.layers[i];
if(l.type == CONVOLUTIONAL){
- update_convolutional_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
+ update_convolutional_layer(l, update_batch, rate, net.momentum, net.decay);
} else if(l.type == DECONVOLUTIONAL){
- update_deconvolutional_layer(l, net.learning_rate, net.momentum, net.decay);
+ update_deconvolutional_layer(l, rate, net.momentum, net.decay);
} else if(l.type == CONNECTED){
- update_connected_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
+ update_connected_layer(l, update_batch, rate, net.momentum, net.decay);
+ } else if(l.type == LOCAL){
+ update_local_layer(l, update_batch, rate, net.momentum, net.decay);
}
}
}
@@ -143,10 +194,6 @@
sum += net.layers[i].cost[0];
++count;
}
- if(net.layers[i].type == REGION){
- sum += net.layers[i].cost[0];
- ++count;
- }
}
return sum/count;
}
@@ -187,12 +234,12 @@
backward_dropout_layer(l, state);
} else if(l.type == DETECTION){
backward_detection_layer(l, state);
- } else if(l.type == REGION){
- backward_region_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 == LOCAL){
+ backward_local_layer(l, state);
} else if(l.type == COST){
backward_cost_layer(l, state);
} else if(l.type == ROUTE){
@@ -203,6 +250,7 @@
float train_network_datum(network net, float *x, float *y)
{
+ *net.seen += net.batch;
#ifdef GPU
if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
@@ -214,7 +262,7 @@
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);
+ if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);
return error;
}
@@ -227,7 +275,6 @@
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;
@@ -248,7 +295,6 @@
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;
}
@@ -294,6 +340,7 @@
//if(w == net->w && h == net->h) return 0;
net->w = w;
net->h = h;
+ int inputs = 0;
//fprintf(stderr, "Resizing to %d x %d...", w, h);
//fflush(stderr);
for (i = 0; i < net->n; ++i){
@@ -307,9 +354,12 @@
break;
}else if(l.type == NORMALIZATION){
resize_normalization_layer(&l, w, h);
+ }else if(l.type == COST){
+ resize_cost_layer(&l, inputs);
}else{
error("Cannot resize this type of layer");
}
+ inputs = l.outputs;
net->layers[i] = l;
w = l.out_w;
h = l.out_h;
@@ -500,12 +550,12 @@
return acc;
}
-float *network_accuracies(network net, data d)
+float *network_accuracies(network net, data d, int n)
{
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);
+ acc[0] = matrix_topk_accuracy(d.y, guess, 1);
+ acc[1] = matrix_topk_accuracy(d.y, guess, n);
free_matrix(guess);
return acc;
}
--
Gitblit v1.10.0