From f047cfff99e00e28c02eb59b6d32386c122f9af6 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 08 Mar 2015 18:31:12 +0000
Subject: [PATCH] renamed sigmoid to logistic
---
src/convolutional_kernels.cu | 68 ++++++++++++---------------------
1 files changed, 25 insertions(+), 43 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 6461aff..bcf307f 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -8,7 +8,7 @@
#include "cuda.h"
}
-__global__ void bias(int n, int size, float *biases, float *output)
+__global__ void bias_output_kernel(float *output, float *biases, int n, int size)
{
int offset = blockIdx.x * blockDim.x + threadIdx.x;
int filter = blockIdx.y;
@@ -17,22 +17,20 @@
if(offset < size) output[(batch*n+filter)*size + offset] = biases[filter];
}
-extern "C" void bias_output_gpu(const convolutional_layer layer)
+extern "C" void bias_output_gpu(float *output, float *biases, int batch, int n, int size)
{
- int size = convolutional_out_height(layer)*convolutional_out_width(layer);
-
dim3 dimBlock(BLOCK, 1, 1);
- dim3 dimGrid((size-1)/BLOCK + 1, layer.n, layer.batch);
+ dim3 dimGrid((size-1)/BLOCK + 1, n, batch);
- bias<<<dimGrid, dimBlock>>>(layer.n, size, layer.biases_gpu, layer.output_gpu);
+ bias_output_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
check_error(cudaPeekAtLastError());
}
-__global__ void learn_bias(int batch, int n, int size, float *delta, float *bias_updates)
+__global__ void backward_bias_kernel(float *bias_updates, float *delta, int batch, int n, int size, float scale)
{
__shared__ float part[BLOCK];
int i,b;
- int filter = (blockIdx.x + blockIdx.y*gridDim.x);
+ int filter = blockIdx.x;
int p = threadIdx.x;
float sum = 0;
for(b = 0; b < batch; ++b){
@@ -44,40 +42,18 @@
part[p] = sum;
__syncthreads();
if(p == 0){
- for(i = 0; i < BLOCK; ++i) bias_updates[filter] += part[i];
+ for(i = 0; i < BLOCK; ++i) bias_updates[filter] += scale * part[i];
}
}
-extern "C" void learn_bias_convolutional_layer_ongpu(convolutional_layer layer)
+extern "C" void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
{
- int size = convolutional_out_height(layer)*convolutional_out_width(layer);
+ float alpha = 1./batch;
-
- learn_bias<<<cuda_gridsize(layer.n), BLOCK>>>(layer.batch, layer.n, size, layer.delta_gpu, layer.bias_updates_gpu);
+ backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size, alpha);
check_error(cudaPeekAtLastError());
}
-extern "C" void test_learn_bias(convolutional_layer l)
-{
- int i;
- int size = convolutional_out_height(l) * convolutional_out_width(l);
- for(i = 0; i < size*l.batch*l.n; ++i){
- l.delta[i] = rand_uniform();
- }
- for(i = 0; i < l.n; ++i){
- l.bias_updates[i] = rand_uniform();
- }
- cuda_push_array(l.delta_gpu, l.delta, size*l.batch*l.n);
- cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.n);
- float *gpu = (float *) calloc(l.n, sizeof(float));
- cuda_pull_array(l.bias_updates_gpu, gpu, l.n);
- for(i = 0; i < l.n; ++i) printf("%.9g %.9g\n", l.bias_updates[i], gpu[i]);
- learn_bias_convolutional_layer_ongpu(l);
- learn_bias_convolutional_layer(l);
- cuda_pull_array(l.bias_updates_gpu, gpu, l.n);
- for(i = 0; i < l.n; ++i) printf("%.9g %.9g\n", l.bias_updates[i], gpu[i]);
-}
-
extern "C" void forward_convolutional_layer_gpu(convolutional_layer layer, float *in)
{
int i;
@@ -86,30 +62,29 @@
int n = convolutional_out_height(layer)*
convolutional_out_width(layer);
- bias_output_gpu(layer);
+ bias_output_gpu(layer.output_gpu, layer.biases_gpu, layer.batch, layer.n, n);
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_gpu);
+ 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_gpu);
float * a = layer.filters_gpu;
float * b = layer.col_image_gpu;
float * c = layer.output_gpu;
gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
}
activate_array_ongpu(layer.output_gpu, m*n*layer.batch, layer.activation);
- cuda_pull_array(layer.output_gpu, layer.output, m*n*layer.batch);
- //for(i = 0; i < m*n*layer.batch; ++i) printf("%f, ", layer.output[i]);
- //printf("\n");
}
extern "C" void backward_convolutional_layer_gpu(convolutional_layer layer, float *in, float *delta_gpu)
{
+ float alpha = 1./layer.batch;
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_gpu, m*k*layer.batch, layer.activation, layer.delta_gpu);
- learn_bias_convolutional_layer_ongpu(layer);
+ backward_bias_gpu(layer.bias_updates_gpu, layer.delta_gpu, layer.batch, layer.n, k);
if(delta_gpu) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, delta_gpu, 1);
@@ -118,8 +93,8 @@
float * b = layer.col_image_gpu;
float * c = layer.filter_updates_gpu;
- 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_gpu);
- gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
+ 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_gpu);
+ gemm_ongpu(0,1,m,n,k,alpha,a + i*m*k,k,b,k,1,c,n);
if(delta_gpu){
@@ -129,7 +104,7 @@
gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
- col2im_ongpu(layer.col_image_gpu, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_gpu);
+ col2im_ongpu(layer.col_image_gpu, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_gpu + i*layer.c*layer.h*layer.w);
}
}
}
@@ -153,6 +128,13 @@
extern "C" void update_convolutional_layer_gpu(convolutional_layer layer)
{
int size = layer.size*layer.size*layer.c*layer.n;
+
+/*
+ cuda_pull_array(layer.filter_updates_gpu, layer.filter_updates, size);
+ cuda_pull_array(layer.filters_gpu, layer.filters, size);
+ printf("Filter: %f updates: %f\n", mag_array(layer.filters, size), layer.learning_rate*mag_array(layer.filter_updates, size));
+ */
+
axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
scal_ongpu(layer.n,layer.momentum, layer.bias_updates_gpu, 1);
--
Gitblit v1.10.0