From 68213b835b9f15cb449ad2037a8b51c17a3de07b Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 14 Mar 2016 22:10:14 +0000
Subject: [PATCH] Makefile
---
src/convolutional_kernels.cu | 96 +++++++++++++++++++++++++++++++++++++++++++----
1 files changed, 87 insertions(+), 9 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 60a1879..85b92df 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -1,3 +1,7 @@
+#include "cuda_runtime.h"
+#include "curand.h"
+#include "cublas_v2.h"
+
extern "C" {
#include "convolutional_layer.h"
#include "gemm.h"
@@ -8,6 +12,21 @@
#include "cuda.h"
}
+__global__ void binarize_filters_kernel(float *filters, int n, int size, float *binary)
+{
+ int f = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if (f >= n) return;
+ int i = 0;
+ float mean = 0;
+ for(i = 0; i < size; ++i){
+ mean += abs(filters[f*size + i]);
+ }
+ mean = mean / size;
+ for(i = 0; i < size; ++i){
+ binary[f*size + i] = (filters[f*size + i] > 0) ? mean : -mean;
+ }
+}
+
__global__ void scale_bias_kernel(float *output, float *biases, int n, int size)
{
int offset = blockIdx.x * blockDim.x + threadIdx.x;
@@ -46,6 +65,12 @@
}
}
+void binarize_filters_gpu(float *filters, int n, int size, float *mean)
+{
+ binarize_filters_kernel<<<cuda_gridsize(n), BLOCK>>>(filters, n, size, mean);
+ check_error(cudaPeekAtLastError());
+}
+
void backward_scale_gpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates)
{
backward_scale_kernel<<<n, BLOCK>>>(x_norm, delta, batch, n, size, scale_updates);
@@ -90,12 +115,59 @@
}
}
+__global__ void dot_kernel(float *output, float scale, int batch, int n, int size, float *delta)
+{
+ int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ int f1 = index / n;
+ int f2 = index % n;
+ if (f2 <= f1) return;
+
+ float sum = 0;
+ float norm1 = 0;
+ float norm2 = 0;
+ int b, i;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < size; ++i){
+ int i1 = b * size * n + f1 * size + i;
+ int i2 = b * size * n + f2 * size + i;
+ sum += output[i1] * output[i2];
+ norm1 += output[i1] * output[i1];
+ norm2 += output[i2] * output[i2];
+ }
+ }
+ norm1 = sqrt(norm1);
+ norm2 = sqrt(norm2);
+ float norm = norm1 * norm2;
+ sum = sum / norm;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < size; ++i){
+ int i1 = b * size * n + f1 * size + i;
+ int i2 = b * size * n + f2 * size + i;
+ delta[i1] += - scale * sum * output[i2] / norm;
+ delta[i2] += - scale * sum * output[i1] / norm;
+ }
+ }
+}
+
+void dot_error_gpu(layer l)
+{
+ dot_kernel<<<cuda_gridsize(l.n*l.n), BLOCK>>>(l.output_gpu, l.dot, l.batch, l.n, l.out_w * l.out_h, l.delta_gpu);
+ check_error(cudaPeekAtLastError());
+}
+
void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
{
backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size);
check_error(cudaPeekAtLastError());
}
+void swap_binary(convolutional_layer *l)
+{
+ float *swap = l->filters_gpu;
+ l->filters_gpu = l->binary_filters_gpu;
+ l->binary_filters_gpu = swap;
+}
+
void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
{
int i;
@@ -105,6 +177,11 @@
convolutional_out_width(l);
fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
+ if(l.binary){
+ binarize_filters_gpu(l.filters_gpu, l.n, l.c*l.size*l.size, l.binary_filters_gpu);
+ swap_binary(&l);
+ }
+
for(i = 0; i < l.batch; ++i){
im2col_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, l.col_image_gpu);
float * a = l.filters_gpu;
@@ -113,19 +190,16 @@
gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
}
- if(l.batch_normalize){
- if(state.train){
- fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.spatial_mean_gpu, l.mean_gpu);
- fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.spatial_variance_gpu, l.variance_gpu);
+ if (l.batch_normalize) {
+ if (state.train) {
+ fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.mean_gpu);
+ fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.variance_gpu);
scal_ongpu(l.n, .95, l.rolling_mean_gpu, 1);
axpy_ongpu(l.n, .05, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
scal_ongpu(l.n, .95, l.rolling_variance_gpu, 1);
axpy_ongpu(l.n, .05, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
- // cuda_pull_array(l.variance_gpu, l.mean, l.n);
- // printf("%f\n", l.mean[0]);
-
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_h*l.out_w);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
@@ -138,6 +212,8 @@
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, n);
activate_array_ongpu(l.output_gpu, m*n*l.batch, l.activation);
+ if(l.dot > 0) dot_error_gpu(l);
+ if(l.binary) swap_binary(&l);
}
void backward_convolutional_layer_gpu(convolutional_layer l, network_state state)
@@ -157,8 +233,8 @@
scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.n, l.out_h*l.out_w);
- fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.spatial_mean_delta_gpu, l.mean_delta_gpu);
- fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.spatial_variance_delta_gpu, l.variance_delta_gpu);
+ fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.mean_delta_gpu);
+ fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.variance_delta_gpu);
normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.n, l.out_w*l.out_h, l.delta_gpu);
}
@@ -171,6 +247,7 @@
gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
if(state.delta){
+ if(l.binary) swap_binary(&l);
float * a = l.filters_gpu;
float * b = l.delta_gpu;
float * c = l.col_image_gpu;
@@ -178,6 +255,7 @@
gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
col2im_ongpu(l.col_image_gpu, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta + i*l.c*l.h*l.w);
+ if(l.binary) swap_binary(&l);
}
}
}
--
Gitblit v1.10.0