From aa5996d58e68edfbefe51061856aecd549dd09c4 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 13 Jan 2015 01:27:08 +0000
Subject: [PATCH] Faster
---
src/convolutional_layer.c | 41 ++++++++++++++++++++++++-----------------
1 files changed, 24 insertions(+), 17 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index bae06d3..fc5cb0e 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -59,17 +59,15 @@
layer->filters = calloc(c*n*size*size, sizeof(float));
layer->filter_updates = calloc(c*n*size*size, sizeof(float));
- layer->filter_momentum = calloc(c*n*size*size, sizeof(float));
layer->biases = calloc(n, sizeof(float));
layer->bias_updates = calloc(n, sizeof(float));
- layer->bias_momentum = calloc(n, sizeof(float));
- float scale = 1./(size*size*c);
- scale = .01;
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*2*(rand_uniform()-.5);
+ float scale = 1./sqrt(size*size*c);
+ //scale = .05;
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal();
for(i = 0; i < n; ++i){
//layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = .5;
+ layer->biases[i] = scale;
}
int out_h = convolutional_out_height(*layer);
int out_w = convolutional_out_width(*layer);
@@ -77,14 +75,13 @@
layer->col_image = calloc(out_h*out_w*size*size*c, sizeof(float));
layer->output = calloc(layer->batch*out_h * out_w * n, sizeof(float));
layer->delta = calloc(layer->batch*out_h * out_w * n, sizeof(float));
+
#ifdef GPU
layer->filters_cl = cl_make_array(layer->filters, c*n*size*size);
layer->filter_updates_cl = cl_make_array(layer->filter_updates, c*n*size*size);
- layer->filter_momentum_cl = cl_make_array(layer->filter_momentum, c*n*size*size);
layer->biases_cl = cl_make_array(layer->biases, n);
layer->bias_updates_cl = cl_make_array(layer->bias_updates, n);
- layer->bias_momentum_cl = cl_make_array(layer->bias_momentum, n);
layer->col_image_cl = cl_make_array(layer->col_image, out_h*out_w*size*size*c);
layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n);
@@ -173,7 +170,9 @@
int n = layer.size*layer.size*layer.c;
int k = convolutional_out_height(layer)*
convolutional_out_width(layer);
+
gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
+
learn_bias_convolutional_layer(layer);
if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
@@ -207,7 +206,7 @@
axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1);
- scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1);
+ axpy_cpu(size, -layer.decay, layer.filters, 1, layer.filter_updates, 1);
axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
scal_cpu(size, layer.momentum, layer.filter_updates, 1);
}
@@ -267,13 +266,18 @@
}
#ifdef GPU
+#define BLOCK 32
+
+#define STR_HELPER(x) #x
+#define STR(x) STR_HELPER(x)
+
cl_kernel get_convolutional_learn_bias_kernel()
{
static int init = 0;
static cl_kernel kernel;
if(!init){
- kernel = get_kernel("src/convolutional_layer.cl", "learn_bias", 0);
+ kernel = get_kernel("src/convolutional_layer.cl", "learn_bias", "-D BLOCK=" STR(BLOCK));
init = 1;
}
return kernel;
@@ -283,7 +287,6 @@
{
int size = convolutional_out_height(layer) * convolutional_out_width(layer);
- cl_setup();
cl_kernel kernel = get_convolutional_learn_bias_kernel();
cl_command_queue queue = cl.queue;
@@ -295,9 +298,10 @@
cl.error = clSetKernelArg(kernel, i++, sizeof(layer.bias_updates_cl), (void*) &layer.bias_updates_cl);
check_error(cl);
- const size_t global_size[] = {layer.n};
+ const size_t global_size[] = {layer.n*BLOCK};
+ const size_t local_size[] = {BLOCK};
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, local_size, 0, 0, 0);
check_error(cl);
}
@@ -306,7 +310,7 @@
static int init = 0;
static cl_kernel kernel;
if(!init){
- kernel = get_kernel("src/convolutional_layer.cl", "bias", 0);
+ kernel = get_kernel("src/convolutional_layer.cl", "bias", "-D BLOCK=" STR(BLOCK));
init = 1;
}
return kernel;
@@ -318,7 +322,6 @@
int out_w = convolutional_out_width(layer);
int size = out_h*out_w;
- cl_setup();
cl_kernel kernel = get_convolutional_bias_kernel();
cl_command_queue queue = cl.queue;
@@ -394,12 +397,16 @@
{
cl_read_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
cl_read_array(layer.biases_cl, layer.biases, layer.n);
+ cl_read_array(layer.filter_updates_cl, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
+ cl_read_array(layer.bias_updates_cl, layer.bias_updates, layer.n);
}
void push_convolutional_layer(convolutional_layer layer)
{
cl_write_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
cl_write_array(layer.biases_cl, layer.biases, layer.n);
+ cl_write_array(layer.filter_updates_cl, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
+ cl_write_array(layer.bias_updates_cl, layer.bias_updates, layer.n);
}
void update_convolutional_layer_gpu(convolutional_layer layer)
@@ -408,10 +415,10 @@
axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
scal_ongpu(layer.n,layer.momentum, layer.bias_updates_cl, 1);
- scal_ongpu(size, 1.-layer.learning_rate*layer.decay, layer.filters_cl, 1);
+ axpy_ongpu(size, -layer.decay, layer.filters_cl, 1, layer.filter_updates_cl, 1);
axpy_ongpu(size, layer.learning_rate, layer.filter_updates_cl, 1, layer.filters_cl, 1);
scal_ongpu(size, layer.momentum, layer.filter_updates_cl, 1);
- pull_convolutional_layer(layer);
+ //pull_convolutional_layer(layer);
}
--
Gitblit v1.10.0