From dc0d7bb8a8779dc194ddaa57260815c1195d398e Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 08 May 2015 17:37:39 +0000
Subject: [PATCH] forgot route layer
---
src/network_kernels.cu | 52 ++++++++++++++++++++++------------------------------
1 files changed, 22 insertions(+), 30 deletions(-)
diff --git a/src/network_kernels.cu b/src/network_kernels.cu
index 03cb149..7ff5d15 100644
--- a/src/network_kernels.cu
+++ b/src/network_kernels.cu
@@ -18,17 +18,17 @@
#include "normalization_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
+#include "route_layer.h"
}
-extern "C" float * get_network_output_gpu_layer(network net, int i);
-extern "C" float * get_network_delta_gpu_layer(network net, int i);
-float *get_network_output_gpu(network net);
+float * get_network_output_gpu_layer(network net, int i);
+float * get_network_delta_gpu_layer(network net, int i);
+float * get_network_output_gpu(network net);
void forward_network_gpu(network net, network_state state)
{
int i;
for(i = 0; i < net.n; ++i){
-//clock_t time = clock();
if(net.types[i] == CONVOLUTIONAL){
forward_convolutional_layer_gpu(*(convolutional_layer *)net.layers[i], state);
}
@@ -56,10 +56,10 @@
else if(net.types[i] == CROP){
forward_crop_layer_gpu(*(crop_layer *)net.layers[i], state);
}
+ else if(net.types[i] == ROUTE){
+ forward_route_layer_gpu(*(route_layer *)net.layers[i], net);
+ }
state.input = get_network_output_gpu_layer(net, i);
-//cudaDeviceSynchronize();
-//printf("forw %d: %s %f\n", i, get_layer_string(net.types[i]), sec(clock() - time));
-//time = clock();
}
}
@@ -68,7 +68,6 @@
int i;
float * original_input = state.input;
for(i = net.n-1; i >= 0; --i){
-//clock_t time = clock();
if(i == 0){
state.input = original_input;
state.delta = 0;
@@ -76,6 +75,7 @@
state.input = get_network_output_gpu_layer(net, i-1);
state.delta = get_network_delta_gpu_layer(net, i-1);
}
+
if(net.types[i] == CONVOLUTIONAL){
backward_convolutional_layer_gpu(*(convolutional_layer *)net.layers[i], state);
}
@@ -100,19 +100,20 @@
else if(net.types[i] == SOFTMAX){
backward_softmax_layer_gpu(*(softmax_layer *)net.layers[i], state);
}
-//cudaDeviceSynchronize();
-//printf("back %d: %s %f\n", i, get_layer_string(net.types[i]), sec(clock() - time));
-//time = clock();
+ else if(net.types[i] == ROUTE){
+ backward_route_layer_gpu(*(route_layer *)net.layers[i], net);
+ }
}
}
void update_network_gpu(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_gpu(layer, net.learning_rate, net.momentum, net.decay);
+ update_convolutional_layer_gpu(layer, update_batch, net.learning_rate, net.momentum, net.decay);
}
else if(net.types[i] == DECONVOLUTIONAL){
deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
@@ -120,7 +121,7 @@
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer_gpu(layer, net.learning_rate, net.momentum, net.decay);
+ update_connected_layer_gpu(layer, update_batch, net.learning_rate, net.momentum, net.decay);
}
}
}
@@ -148,6 +149,9 @@
else if(net.types[i] == SOFTMAX){
return ((softmax_layer *)net.layers[i]) -> output_gpu;
}
+ else if(net.types[i] == ROUTE){
+ return ((route_layer *)net.layers[i]) -> output_gpu;
+ }
else if(net.types[i] == DROPOUT){
return get_network_output_gpu_layer(net, i-1);
}
@@ -176,6 +180,10 @@
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.delta_gpu;
}
+ else if(net.types[i] == ROUTE){
+ route_layer layer = *(route_layer *)net.layers[i];
+ return layer.delta_gpu;
+ }
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.delta_gpu;
@@ -188,7 +196,6 @@
float train_network_datum_gpu(network net, float *x, float *y)
{
- // clock_t time = clock();
network_state state;
int x_size = get_network_input_size(net)*net.batch;
int y_size = get_network_output_size(net)*net.batch;
@@ -202,26 +209,11 @@
state.input = *net.input_gpu;
state.truth = *net.truth_gpu;
state.train = 1;
-//cudaDeviceSynchronize();
-//printf("trans %f\n", sec(clock() - time));
-//time = clock();
forward_network_gpu(net, state);
-//cudaDeviceSynchronize();
-//printf("forw %f\n", sec(clock() - time));
-//time = clock();
backward_network_gpu(net, state);
-//cudaDeviceSynchronize();
-//printf("back %f\n", sec(clock() - time));
-//time = clock();
- update_network_gpu(net);
float error = get_network_cost(net);
+ if ((net.seen / net.batch) % net.subdivisions == 0) update_network_gpu(net);
- //print_letters(y, 50);
- //float *out = get_network_output_gpu(net);
- //print_letters(out, 50);
-//cudaDeviceSynchronize();
-//printf("updt %f\n", sec(clock() - time));
-//time = clock();
return error;
}
--
Gitblit v1.10.0