From 5c067dc44785a761a0243d8cd634e3ac17d548ad Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 12 Sep 2016 20:55:20 +0000
Subject: [PATCH] good chance I didn't break anything
---
src/network_kernels.cu | 180 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 178 insertions(+), 2 deletions(-)
diff --git a/src/network_kernels.cu b/src/network_kernels.cu
index 730634e..3e0c2b6 100644
--- a/src/network_kernels.cu
+++ b/src/network_kernels.cu
@@ -16,14 +16,18 @@
#include "crop_layer.h"
#include "connected_layer.h"
#include "rnn_layer.h"
+#include "gru_layer.h"
#include "crnn_layer.h"
#include "detection_layer.h"
+#include "region_layer.h"
#include "convolutional_layer.h"
#include "activation_layer.h"
#include "deconvolutional_layer.h"
#include "maxpool_layer.h"
+#include "reorg_layer.h"
#include "avgpool_layer.h"
#include "normalization_layer.h"
+#include "batchnorm_layer.h"
#include "cost_layer.h"
#include "local_layer.h"
#include "softmax_layer.h"
@@ -39,6 +43,7 @@
void forward_network_gpu(network net, network_state state)
{
+ state.workspace = net.workspace;
int i;
for(i = 0; i < net.n; ++i){
state.index = i;
@@ -56,10 +61,14 @@
forward_local_layer_gpu(l, state);
} else if(l.type == DETECTION){
forward_detection_layer_gpu(l, state);
+ } else if(l.type == REGION){
+ forward_region_layer_gpu(l, state);
} else if(l.type == CONNECTED){
forward_connected_layer_gpu(l, state);
} else if(l.type == RNN){
forward_rnn_layer_gpu(l, state);
+ } else if(l.type == GRU){
+ forward_gru_layer_gpu(l, state);
} else if(l.type == CRNN){
forward_crnn_layer_gpu(l, state);
} else if(l.type == CROP){
@@ -70,8 +79,12 @@
forward_softmax_layer_gpu(l, state);
} else if(l.type == NORMALIZATION){
forward_normalization_layer_gpu(l, state);
+ } else if(l.type == BATCHNORM){
+ forward_batchnorm_layer_gpu(l, state);
} else if(l.type == MAXPOOL){
forward_maxpool_layer_gpu(l, state);
+ } else if(l.type == REORG){
+ forward_reorg_layer_gpu(l, state);
} else if(l.type == AVGPOOL){
forward_avgpool_layer_gpu(l, state);
} else if(l.type == DROPOUT){
@@ -87,6 +100,7 @@
void backward_network_gpu(network net, network_state state)
{
+ state.workspace = net.workspace;
int i;
float * original_input = state.input;
float * original_delta = state.delta;
@@ -111,20 +125,28 @@
backward_local_layer_gpu(l, state);
} else if(l.type == MAXPOOL){
if(i != 0) backward_maxpool_layer_gpu(l, state);
+ } else if(l.type == REORG){
+ backward_reorg_layer_gpu(l, state);
} else if(l.type == AVGPOOL){
if(i != 0) backward_avgpool_layer_gpu(l, state);
} else if(l.type == DROPOUT){
backward_dropout_layer_gpu(l, state);
} else if(l.type == DETECTION){
backward_detection_layer_gpu(l, state);
+ } else if(l.type == REGION){
+ backward_region_layer_gpu(l, state);
} else if(l.type == NORMALIZATION){
backward_normalization_layer_gpu(l, state);
+ } else if(l.type == BATCHNORM){
+ backward_batchnorm_layer_gpu(l, state);
} else if(l.type == SOFTMAX){
if(i != 0) backward_softmax_layer_gpu(l, state);
} else if(l.type == CONNECTED){
backward_connected_layer_gpu(l, state);
} else if(l.type == RNN){
backward_rnn_layer_gpu(l, state);
+ } else if(l.type == GRU){
+ backward_gru_layer_gpu(l, state);
} else if(l.type == CRNN){
backward_crnn_layer_gpu(l, state);
} else if(l.type == COST){
@@ -150,6 +172,8 @@
update_deconvolutional_layer_gpu(l, rate, net.momentum, net.decay);
} else if(l.type == CONNECTED){
update_connected_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
+ } else if(l.type == GRU){
+ update_gru_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
} else if(l.type == RNN){
update_rnn_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
} else if(l.type == CRNN){
@@ -160,14 +184,14 @@
}
}
-float train_network_datum_gpu(network net, float *x, float *y)
+void forward_backward_network_gpu(network net, float *x, float *y)
{
network_state state;
state.index = 0;
state.net = net;
int x_size = get_network_input_size(net)*net.batch;
int y_size = get_network_output_size(net)*net.batch;
- if(net.layers[net.n-1].type == DETECTION) y_size = net.layers[net.n-1].truths*net.batch;
+ if(net.layers[net.n-1].truths) y_size = net.layers[net.n-1].truths*net.batch;
if(!*net.input_gpu){
*net.input_gpu = cuda_make_array(x, x_size);
*net.truth_gpu = cuda_make_array(y, y_size);
@@ -181,12 +205,164 @@
state.train = 1;
forward_network_gpu(net, state);
backward_network_gpu(net, state);
+}
+
+float train_network_datum_gpu(network net, float *x, float *y)
+{
+ *net.seen += net.batch;
+ forward_backward_network_gpu(net, x, y);
float error = get_network_cost(net);
if (((*net.seen) / net.batch) % net.subdivisions == 0) update_network_gpu(net);
return error;
}
+typedef struct {
+ network net;
+ float *X;
+ float *y;
+} train_args;
+
+void *train_thread(void *ptr)
+{
+ train_args args = *(train_args*)ptr;
+
+ cuda_set_device(args.net.gpu_index);
+ forward_backward_network_gpu(args.net, args.X, args.y);
+ free(ptr);
+ return 0;
+}
+
+pthread_t train_network_in_thread(network net, float *X, float *y)
+{
+ pthread_t thread;
+ train_args *ptr = (train_args *)calloc(1, sizeof(train_args));
+ ptr->net = net;
+ ptr->X = X;
+ ptr->y = y;
+ if(pthread_create(&thread, 0, train_thread, ptr)) error("Thread creation failed");
+ return thread;
+}
+
+void pull_updates(layer l)
+{
+#ifdef GPU
+ if(l.type == CONVOLUTIONAL){
+ cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.n);
+ cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.n*l.size*l.size*l.c);
+ if(l.scale_updates) cuda_pull_array(l.scale_updates_gpu, l.scale_updates, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+ cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.outputs*l.inputs);
+ }
+#endif
+}
+
+void push_updates(layer l)
+{
+#ifdef GPU
+ if(l.type == CONVOLUTIONAL){
+ cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.n);
+ cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.n*l.size*l.size*l.c);
+ if(l.scale_updates) cuda_push_array(l.scale_updates_gpu, l.scale_updates, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+ cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.outputs*l.inputs);
+ }
+#endif
+}
+
+void merge_updates(layer l, layer base)
+{
+ if (l.type == CONVOLUTIONAL) {
+ axpy_cpu(l.n, 1, l.bias_updates, 1, base.bias_updates, 1);
+ axpy_cpu(l.n*l.size*l.size*l.c, 1, l.weight_updates, 1, base.weight_updates, 1);
+ if (l.scale_updates) {
+ axpy_cpu(l.n, 1, l.scale_updates, 1, base.scale_updates, 1);
+ }
+ } else if(l.type == CONNECTED) {
+ axpy_cpu(l.outputs, 1, l.bias_updates, 1, base.bias_updates, 1);
+ axpy_cpu(l.outputs*l.inputs, 1, l.weight_updates, 1, base.weight_updates, 1);
+ }
+}
+
+void distribute_updates(layer l, layer base)
+{
+ if (l.type == CONVOLUTIONAL) {
+ copy_cpu(l.n, base.bias_updates, 1, l.bias_updates, 1);
+ copy_cpu(l.n*l.size*l.size*l.c, base.weight_updates, 1, l.weight_updates, 1);
+ if (l.scale_updates) {
+ copy_cpu(l.n, base.scale_updates, 1, l.scale_updates, 1);
+ }
+ } else if(l.type == CONNECTED) {
+ copy_cpu(l.outputs, base.bias_updates, 1, l.bias_updates, 1);
+ copy_cpu(l.outputs*l.inputs, base.weight_updates, 1, l.weight_updates, 1);
+ }
+}
+
+void sync_updates(network *nets, int n)
+{
+ int i,j;
+ int layers = nets[0].n;
+ network net = nets[0];
+ for (j = 0; j < layers; ++j) {
+ layer base = net.layers[j];
+ cuda_set_device(net.gpu_index);
+ pull_updates(base);
+ for (i = 1; i < n; ++i) {
+ cuda_set_device(nets[i].gpu_index);
+ layer l = nets[i].layers[j];
+ pull_updates(l);
+ merge_updates(l, base);
+ }
+ for (i = 1; i < n; ++i) {
+ cuda_set_device(nets[i].gpu_index);
+ layer l = nets[i].layers[j];
+ distribute_updates(l, base);
+ push_updates(l);
+ }
+ cuda_set_device(net.gpu_index);
+ push_updates(base);
+ }
+ for (i = 0; i < n; ++i) {
+ cuda_set_device(nets[i].gpu_index);
+ if(i > 0) nets[i].momentum = 0;
+ update_network_gpu(nets[i]);
+ }
+}
+
+float train_networks(network *nets, int n, data d)
+{
+ int batch = nets[0].batch;
+ assert(batch * n == d.X.rows);
+ assert(nets[0].subdivisions % n == 0);
+ float **X = (float **) calloc(n, sizeof(float *));
+ float **y = (float **) calloc(n, sizeof(float *));
+ pthread_t *threads = (pthread_t *) calloc(n, sizeof(pthread_t));
+
+ int i;
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ X[i] = (float *) calloc(batch*d.X.cols, sizeof(float));
+ y[i] = (float *) calloc(batch*d.y.cols, sizeof(float));
+ get_next_batch(d, batch, i*batch, X[i], y[i]);
+ threads[i] = train_network_in_thread(nets[i], X[i], y[i]);
+ }
+ for(i = 0; i < n; ++i){
+ pthread_join(threads[i], 0);
+ *nets[i].seen += n*nets[i].batch;
+ printf("%f\n", get_network_cost(nets[i]) / batch);
+ sum += get_network_cost(nets[i]);
+ free(X[i]);
+ free(y[i]);
+ }
+ if (((*nets[0].seen) / nets[0].batch) % nets[0].subdivisions == 0) sync_updates(nets, n);
+ free(X);
+ free(y);
+ free(threads);
+ return (float)sum/(n*batch);
+}
+
float *get_network_output_layer_gpu(network net, int i)
{
layer l = net.layers[i];
--
Gitblit v1.10.0