From d0b9326a352ed2fbc3ae66fdef40b4533a2f211d Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 11 Aug 2015 06:22:27 +0000
Subject: [PATCH] Hacks to get nightmare to not break gridsizing
---
src/network.c | 96 +++++++++++++++++++++++++++++-------------------
1 files changed, 58 insertions(+), 38 deletions(-)
diff --git a/src/network.c b/src/network.c
index 68790e5..ff5cd61 100644
--- a/src/network.c
+++ b/src/network.c
@@ -4,13 +4,16 @@
#include "image.h"
#include "data.h"
#include "utils.h"
+#include "blas.h"
#include "crop_layer.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "detection_layer.h"
+#include "normalization_layer.h"
#include "maxpool_layer.h"
+#include "avgpool_layer.h"
#include "cost_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
@@ -27,6 +30,8 @@
return "connected";
case MAXPOOL:
return "maxpool";
+ case AVGPOOL:
+ return "avgpool";
case SOFTMAX:
return "softmax";
case DETECTION:
@@ -39,6 +44,8 @@
return "cost";
case ROUTE:
return "route";
+ case NORMALIZATION:
+ return "normalization";
default:
break;
}
@@ -62,10 +69,15 @@
int i;
for(i = 0; i < net.n; ++i){
layer l = net.layers[i];
+ if(l.delta){
+ scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
+ }
if(l.type == CONVOLUTIONAL){
forward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
forward_deconvolutional_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 == CONNECTED){
@@ -78,6 +90,8 @@
forward_softmax_layer(l, state);
} else if(l.type == MAXPOOL){
forward_maxpool_layer(l, state);
+ } else if(l.type == AVGPOOL){
+ forward_avgpool_layer(l, state);
} else if(l.type == DROPOUT){
forward_dropout_layer(l, state);
} else if(l.type == ROUTE){
@@ -112,13 +126,20 @@
float get_network_cost(network net)
{
- if(net.layers[net.n-1].type == COST){
- return net.layers[net.n-1].output[0];
+ int i;
+ float sum = 0;
+ int count = 0;
+ for(i = 0; i < net.n; ++i){
+ if(net.layers[net.n-1].type == COST){
+ sum += net.layers[net.n-1].output[0];
+ ++count;
+ }
+ if(net.layers[net.n-1].type == DETECTION){
+ sum += net.layers[net.n-1].cost[0];
+ ++count;
+ }
}
- if(net.layers[net.n-1].type == DETECTION){
- return net.layers[net.n-1].cost[0];
- }
- return 0;
+ return sum/count;
}
int get_predicted_class_network(network net)
@@ -132,10 +153,11 @@
{
int i;
float *original_input = state.input;
+ float *original_delta = state.delta;
for(i = net.n-1; i >= 0; --i){
if(i == 0){
state.input = original_input;
- state.delta = 0;
+ state.delta = original_delta;
}else{
layer prev = net.layers[i-1];
state.input = prev.output;
@@ -146,8 +168,12 @@
backward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
backward_deconvolutional_layer(l, state);
+ } else if(l.type == NORMALIZATION){
+ backward_normalization_layer(l, state);
} else if(l.type == MAXPOOL){
if(i != 0) backward_maxpool_layer(l, state);
+ } else if(l.type == AVGPOOL){
+ backward_avgpool_layer(l, state);
} else if(l.type == DROPOUT){
backward_dropout_layer(l, state);
} else if(l.type == DETECTION){
@@ -166,11 +192,12 @@
float train_network_datum(network net, float *x, float *y)
{
- #ifdef GPU
+#ifdef GPU
if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
- #endif
+#endif
network_state state;
state.input = x;
+ state.delta = 0;
state.truth = y;
state.train = 1;
forward_network(net, state);
@@ -224,6 +251,7 @@
int i,j;
network_state state;
state.train = 1;
+ state.delta = 0;
float sum = 0;
int batch = 2;
for(i = 0; i < n; ++i){
@@ -249,43 +277,35 @@
}
}
-/*
-int resize_network(network net, int h, int w, int c)
+int resize_network(network *net, int w, int h)
{
- 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);
+ //if(w == net->w && h == net->h) return 0;
+ net->w = w;
+ net->h = h;
+ //fprintf(stderr, "Resizing to %d x %d...", w, h);
+ //fflush(stderr);
+ for (i = 0; i < net->n; ++i){
+ layer l = net->layers[i];
+ if(l.type == CONVOLUTIONAL){
+ resize_convolutional_layer(&l, w, h);
+ }else if(l.type == MAXPOOL){
+ resize_maxpool_layer(&l, w, h);
+ }else if(l.type == AVGPOOL){
+ resize_avgpool_layer(&l, w, h);
+ break;
+ }else if(l.type == NORMALIZATION){
+ resize_normalization_layer(&l, w, h);
}else{
error("Cannot resize this type of layer");
}
+ net->layers[i] = l;
+ w = l.out_w;
+ h = l.out_h;
}
+ //fprintf(stderr, " Done!\n");
return 0;
}
-*/
int get_network_output_size(network net)
{
--
Gitblit v1.10.0