From db0397cfaaf488364e3d2e1669dfefae2ee6ea73 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 14 Dec 2015 19:57:10 +0000
Subject: [PATCH] shortcut layers, msr networks
---
src/network.c | 57 +++++++++++++++++++++++++++++++++++++++++++--------------
1 files changed, 43 insertions(+), 14 deletions(-)
diff --git a/src/network.c b/src/network.c
index 80ee291..8dee8cc 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"
@@ -19,6 +19,7 @@
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"
+#include "shortcut_layer.h"
int get_current_batch(network net)
{
@@ -26,6 +27,17 @@
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);
@@ -41,6 +53,7 @@
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:
@@ -60,6 +73,8 @@
switch(a){
case CONVOLUTIONAL:
return "convolutional";
+ case LOCAL:
+ return "local";
case DECONVOLUTIONAL:
return "deconvolutional";
case CONNECTED:
@@ -72,8 +87,6 @@
return "softmax";
case DETECTION:
return "detection";
- case REGION:
- return "region";
case DROPOUT:
return "dropout";
case CROP:
@@ -82,6 +95,8 @@
return "cost";
case ROUTE:
return "route";
+ case SHORTCUT:
+ return "shortcut";
case NORMALIZATION:
return "normalization";
default:
@@ -107,6 +122,7 @@
{
int i;
for(i = 0; i < net.n; ++i){
+ state.index = i;
layer l = net.layers[i];
if(l.delta){
scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
@@ -115,12 +131,12 @@
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){
@@ -137,6 +153,8 @@
forward_dropout_layer(l, state);
} else if(l.type == ROUTE){
forward_route_layer(l, net);
+ } else if(l.type == SHORTCUT){
+ forward_shortcut_layer(l, state);
}
state.input = l.output;
}
@@ -155,6 +173,8 @@
update_deconvolutional_layer(l, rate, net.momentum, net.decay);
} else if(l.type == CONNECTED){
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);
}
}
}
@@ -180,10 +200,6 @@
sum += net.layers[i].cost[0];
++count;
}
- if(net.layers[i].type == REGION){
- sum += net.layers[i].cost[0];
- ++count;
- }
}
return sum/count;
}
@@ -201,6 +217,7 @@
float *original_input = state.input;
float *original_delta = state.delta;
for(i = net.n-1; i >= 0; --i){
+ state.index = i;
if(i == 0){
state.input = original_input;
state.delta = original_delta;
@@ -224,16 +241,18 @@
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){
backward_route_layer(l, net);
+ } else if(l.type == SHORTCUT){
+ backward_shortcut_layer(l, state);
}
}
}
@@ -245,6 +264,8 @@
if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
network_state state;
+ state.index = 0;
+ state.net = net;
state.input = x;
state.delta = 0;
state.truth = y;
@@ -297,6 +318,8 @@
{
int i,j;
network_state state;
+ state.index = 0;
+ state.net = net;
state.train = 1;
state.delta = 0;
float sum = 0;
@@ -330,6 +353,7 @@
//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){
@@ -343,9 +367,12 @@
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;
@@ -429,6 +456,8 @@
#endif
network_state state;
+ state.net = net;
+ state.index = 0;
state.input = input;
state.truth = 0;
state.train = 0;
@@ -536,12 +565,12 @@
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