From 1edcf73a73d2007afc61289245763f5cf0c29e10 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 04 Dec 2014 07:20:29 +0000
Subject: [PATCH] Detection good, split up col images
---
src/convolutional_layer.c | 216 ++++++++++++++++++++++++++++++++++++++++++------------
1 files changed, 168 insertions(+), 48 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 6c7f947..bae06d3 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -2,6 +2,7 @@
#include "utils.h"
#include "mini_blas.h"
#include <stdio.h>
+#include <time.h>
int convolutional_out_height(convolutional_layer layer)
{
@@ -64,8 +65,8 @@
layer->bias_updates = calloc(n, sizeof(float));
layer->bias_momentum = calloc(n, sizeof(float));
float scale = 1./(size*size*c);
- //scale = .0001;
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform()-.5);
+ scale = .01;
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*2*(rand_uniform()-.5);
for(i = 0; i < n; ++i){
//layer->biases[i] = rand_normal()*scale + scale;
layer->biases[i] = .5;
@@ -73,7 +74,7 @@
int out_h = convolutional_out_height(*layer);
int out_w = convolutional_out_width(*layer);
- layer->col_image = calloc(layer->batch*out_h*out_w*size*size*c, sizeof(float));
+ 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
@@ -85,7 +86,7 @@
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, layer->batch*out_h*out_w*size*size*c);
+ 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);
layer->output_cl = cl_make_array(layer->output, layer->batch*out_h*out_w*n);
#endif
@@ -105,7 +106,7 @@
int out_w = convolutional_out_width(*layer);
layer->col_image = realloc(layer->col_image,
- layer->batch*out_h*out_w*layer->size*layer->size*layer->c*sizeof(float));
+ out_h*out_w*layer->size*layer->size*layer->c*sizeof(float));
layer->output = realloc(layer->output,
layer->batch*out_h * out_w * layer->n*sizeof(float));
layer->delta = realloc(layer->delta,
@@ -142,20 +143,14 @@
float *b = layer.col_image;
float *c = layer.output;
- im2col_cpu(in, layer.batch, layer.c, layer.h, layer.w,
- layer.size, layer.stride, layer.pad, b);
for(i = 0; i < layer.batch; ++i){
+ im2col_cpu(in, layer.c, layer.h, layer.w,
+ layer.size, layer.stride, layer.pad, b);
gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
c += n*m;
- in += layer.h*layer.w*layer.c;
- b += k*n;
+ in += layer.c*layer.h*layer.w;
}
- /*
- int i;
- for(i = 0; i < m*n; ++i) printf("%f, ", layer.output[i]);
- printf("\n");
- */
activate_array(layer.output, m*n*layer.batch, layer.activation);
}
@@ -166,12 +161,12 @@
*convolutional_out_width(layer);
for(b = 0; b < layer.batch; ++b){
for(i = 0; i < layer.n; ++i){
- layer.bias_updates[i] += mean_array(layer.delta+size*(i+b*layer.n), size);
+ layer.bias_updates[i] += sum_array(layer.delta+size*(i+b*layer.n), size);
}
}
}
-void backward_convolutional_layer(convolutional_layer layer, float *delta)
+void backward_convolutional_layer(convolutional_layer layer, float *in, float *delta)
{
int i;
int m = layer.n;
@@ -181,33 +176,27 @@
gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
learn_bias_convolutional_layer(layer);
- float *a = layer.delta;
- float *b = layer.col_image;
- float *c = layer.filter_updates;
+ if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
for(i = 0; i < layer.batch; ++i){
+ float *a = layer.delta + i*m*k;
+ float *b = layer.col_image;
+ float *c = layer.filter_updates;
+
+ float *im = in+i*layer.c*layer.h*layer.w;
+
+ im2col_cpu(im, layer.c, layer.h, layer.w,
+ layer.size, layer.stride, layer.pad, b);
gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
- a += m*k;
- b += k*n;
- }
- if(delta){
- m = layer.size*layer.size*layer.c;
- k = layer.n;
- n = convolutional_out_height(layer)*
- convolutional_out_width(layer);
+ if(delta){
+ a = layer.filters;
+ b = layer.delta + i*m*k;
+ c = layer.col_image;
- a = layer.filters;
- b = layer.delta;
- c = layer.col_image;
+ gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
- memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
-
- for(i = 0; i < layer.batch; ++i){
- gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
- col2im_cpu(c, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta);
- c += k*n;
- delta += layer.h*layer.w*layer.c;
+ col2im_cpu(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta+i*layer.c*layer.h*layer.w);
}
}
}
@@ -216,7 +205,7 @@
{
int size = layer.size*layer.size*layer.c*layer.n;
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(layer.n, layer.momentum, layer.bias_updates, 1);
scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1);
axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
@@ -278,22 +267,153 @@
}
#ifdef GPU
+
+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);
+ init = 1;
+ }
+ return kernel;
+}
+
+void learn_bias_convolutional_layer_ongpu(convolutional_layer layer)
+{
+ 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;
+
+ cl_uint i = 0;
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.batch), (void*) &layer.batch);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.delta_cl), (void*) &layer.delta_cl);
+ 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};
+
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
+ check_error(cl);
+}
+
+cl_kernel get_convolutional_bias_kernel()
+{
+ static int init = 0;
+ static cl_kernel kernel;
+ if(!init){
+ kernel = get_kernel("src/convolutional_layer.cl", "bias", 0);
+ init = 1;
+ }
+ return kernel;
+}
+
+void bias_output_gpu(const convolutional_layer layer)
+{
+ int out_h = convolutional_out_height(layer);
+ 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;
+
+ cl_uint i = 0;
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.biases_cl), (void*) &layer.biases_cl);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
+ check_error(cl);
+
+ const size_t global_size[] = {layer.n*size, layer.batch};
+
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 2, 0, global_size, 0, 0, 0, 0);
+ check_error(cl);
+}
+
+//#define TIMEIT
+
void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
{
+ int i;
int m = layer.n;
int k = layer.size*layer.size*layer.c;
int n = convolutional_out_height(layer)*
- convolutional_out_width(layer)*
- layer.batch;
+ convolutional_out_width(layer);
- cl_write_array(layer.filters_cl, layer.filters, m*k);
- cl_mem a = layer.filters_cl;
- cl_mem b = layer.col_image_cl;
- cl_mem c = layer.output_cl;
- im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, b);
- gemm_ongpu(0,0,m,n,k,1,a,k,b,n,0,c,n);
- activate_array_ongpu(layer.output_cl, m*n, layer.activation);
- cl_read_array(layer.output_cl, layer.output, m*n);
+ bias_output_gpu(layer);
+
+ for(i = 0; i < layer.batch; ++i){
+ im2col_ongpu(in, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
+ cl_mem a = layer.filters_cl;
+ cl_mem b = layer.col_image_cl;
+ cl_mem c = layer.output_cl;
+ gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,0,n,1.,c,i*m*n,n);
+ }
+ activate_array_ongpu(layer.output_cl, m*n*layer.batch, layer.activation);
}
+
+void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in, cl_mem delta_cl)
+{
+ int i;
+ int m = layer.n;
+ int n = layer.size*layer.size*layer.c;
+ int k = convolutional_out_height(layer)*
+ convolutional_out_width(layer);
+ gradient_array_ongpu(layer.output_cl, m*k*layer.batch, layer.activation, layer.delta_cl);
+ learn_bias_convolutional_layer_ongpu(layer);
+
+ if(delta_cl) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, delta_cl, 1);
+
+ for(i = 0; i < layer.batch; ++i){
+ cl_mem a = layer.delta_cl;
+ cl_mem b = layer.col_image_cl;
+ cl_mem c = layer.filter_updates_cl;
+
+ im2col_ongpu(in, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
+ gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,0,k,1,c,0,n);
+
+ if(delta_cl){
+
+ cl_mem a = layer.filters_cl;
+ cl_mem b = layer.delta_cl;
+ cl_mem c = layer.col_image_cl;
+
+ gemm_ongpu_offset(1,0,n,k,m,1,a,0,n,b,i*k*m,k,0,c,0,k);
+
+ col2im_ongpu(layer.col_image_cl, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl);
+ }
+ }
+}
+
+void pull_convolutional_layer(convolutional_layer layer)
+{
+ 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);
+}
+
+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);
+}
+
+void update_convolutional_layer_gpu(convolutional_layer layer)
+{
+ int size = layer.size*layer.size*layer.c*layer.n;
+ 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.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);
+}
+
+
#endif
--
Gitblit v1.10.0