From 8fd18add6e060a433629fae3fa2a7ef75df4644e Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 04 Nov 2015 03:23:42 +0000
Subject: [PATCH] CVPR Experiments
---
src/convolutional_kernels.cu | 161 ++++++++++++++++++++++++++++++++++++++++++-----------
1 files changed, 127 insertions(+), 34 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index a150c20..60a1879 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -8,21 +8,65 @@
#include "cuda.h"
}
-__global__ void bias_output_kernel(float *output, float *biases, int n, int size)
+__global__ void scale_bias_kernel(float *output, float *biases, int n, int size)
{
int offset = blockIdx.x * blockDim.x + threadIdx.x;
int filter = blockIdx.y;
int batch = blockIdx.z;
- if(offset < size) output[(batch*n+filter)*size + offset] = biases[filter];
+ if(offset < size) output[(batch*n+filter)*size + offset] *= biases[filter];
}
-void bias_output_gpu(float *output, float *biases, int batch, int n, int size)
+void scale_bias_gpu(float *output, float *biases, int batch, int n, int size)
{
dim3 dimGrid((size-1)/BLOCK + 1, n, batch);
dim3 dimBlock(BLOCK, 1, 1);
- bias_output_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
+ scale_bias_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
+ check_error(cudaPeekAtLastError());
+}
+
+__global__ void backward_scale_kernel(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates)
+{
+ __shared__ float part[BLOCK];
+ int i,b;
+ int filter = blockIdx.x;
+ int p = threadIdx.x;
+ float sum = 0;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < size; i += BLOCK){
+ int index = p + i + size*(filter + n*b);
+ sum += (p+i < size) ? delta[index]*x_norm[index] : 0;
+ }
+ }
+ part[p] = sum;
+ __syncthreads();
+ if (p == 0) {
+ for(i = 0; i < BLOCK; ++i) scale_updates[filter] += part[i];
+ }
+}
+
+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);
+ check_error(cudaPeekAtLastError());
+}
+
+__global__ void add_bias_kernel(float *output, float *biases, int n, int size)
+{
+ int offset = blockIdx.x * blockDim.x + threadIdx.x;
+ int filter = blockIdx.y;
+ int batch = blockIdx.z;
+
+ if(offset < size) output[(batch*n+filter)*size + offset] += biases[filter];
+}
+
+void add_bias_gpu(float *output, float *biases, int batch, int n, int size)
+{
+ dim3 dimGrid((size-1)/BLOCK + 1, n, batch);
+ dim3 dimBlock(BLOCK, 1, 1);
+
+ add_bias_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
check_error(cudaPeekAtLastError());
}
@@ -41,7 +85,7 @@
}
part[p] = sum;
__syncthreads();
- if(p == 0){
+ if (p == 0) {
for(i = 0; i < BLOCK; ++i) bias_updates[filter] += part[i];
}
}
@@ -52,53 +96,88 @@
check_error(cudaPeekAtLastError());
}
-void forward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
+void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
{
int i;
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = convolutional_out_height(layer)*
- convolutional_out_width(layer);
+ int m = l.n;
+ int k = l.size*l.size*l.c;
+ int n = convolutional_out_height(l)*
+ convolutional_out_width(l);
- bias_output_gpu(layer.output_gpu, layer.biases_gpu, layer.batch, layer.n, n);
- for(i = 0; i < layer.batch; ++i){
- im2col_ongpu(state.input + 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;
+ fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
+ 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;
+ float * b = l.col_image_gpu;
+ float * c = l.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);
+
+ 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);
+
+ 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);
+ } else {
+ normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.n, l.out_h*l.out_w);
+ }
+
+ scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.n, l.out_h*l.out_w);
+ }
+ 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);
}
-void backward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
+void backward_convolutional_layer_gpu(convolutional_layer l, network_state state)
{
int i;
- int m = layer.n;
- int n = layer.size*layer.size*layer.c;
- int k = convolutional_out_height(layer)*
- convolutional_out_width(layer);
+ int m = l.n;
+ int n = l.size*l.size*l.c;
+ int k = convolutional_out_height(l)*
+ convolutional_out_width(l);
- gradient_array_ongpu(layer.output_gpu, m*k*layer.batch, layer.activation, layer.delta_gpu);
- backward_bias_gpu(layer.bias_updates_gpu, layer.delta_gpu, layer.batch, layer.n, k);
+ gradient_array_ongpu(l.output_gpu, m*k*l.batch, l.activation, l.delta_gpu);
- for(i = 0; i < layer.batch; ++i){
- float * a = layer.delta_gpu;
- float * b = layer.col_image_gpu;
- float * c = layer.filter_updates_gpu;
+ backward_bias_gpu(l.bias_updates_gpu, l.delta_gpu, l.batch, l.n, k);
- im2col_ongpu(state.input + i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_gpu);
+ if(l.batch_normalize){
+ backward_scale_gpu(l.x_norm_gpu, l.delta_gpu, l.batch, l.n, l.out_w*l.out_h, l.scale_updates_gpu);
+
+ 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);
+ 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);
+ }
+
+ for(i = 0; i < l.batch; ++i){
+ float * a = l.delta_gpu;
+ float * b = l.col_image_gpu;
+ float * c = l.filter_updates_gpu;
+
+ 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);
gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
if(state.delta){
-
- float * a = layer.filters_gpu;
- float * b = layer.delta_gpu;
- float * c = layer.col_image_gpu;
+ float * a = l.filters_gpu;
+ float * b = l.delta_gpu;
+ float * c = l.col_image_gpu;
gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
- col2im_ongpu(layer.col_image_gpu, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, state.delta + i*layer.c*layer.h*layer.w);
+ 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);
}
}
}
@@ -109,6 +188,11 @@
cuda_pull_array(layer.biases_gpu, layer.biases, layer.n);
cuda_pull_array(layer.filter_updates_gpu, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
cuda_pull_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
+ if (layer.batch_normalize){
+ cuda_pull_array(layer.scales_gpu, layer.scales, layer.n);
+ cuda_pull_array(layer.rolling_mean_gpu, layer.rolling_mean, layer.n);
+ cuda_pull_array(layer.rolling_variance_gpu, layer.rolling_variance, layer.n);
+ }
}
void push_convolutional_layer(convolutional_layer layer)
@@ -117,6 +201,11 @@
cuda_push_array(layer.biases_gpu, layer.biases, layer.n);
cuda_push_array(layer.filter_updates_gpu, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
+ if (layer.batch_normalize){
+ cuda_push_array(layer.scales_gpu, layer.scales, layer.n);
+ cuda_push_array(layer.rolling_mean_gpu, layer.rolling_mean, layer.n);
+ cuda_push_array(layer.rolling_variance_gpu, layer.rolling_variance, layer.n);
+ }
}
void update_convolutional_layer_gpu(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)
@@ -126,8 +215,12 @@
axpy_ongpu(layer.n, learning_rate/batch, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
scal_ongpu(layer.n, momentum, layer.bias_updates_gpu, 1);
+ axpy_ongpu(layer.n, learning_rate/batch, layer.scale_updates_gpu, 1, layer.scales_gpu, 1);
+ scal_ongpu(layer.n, momentum, layer.scale_updates_gpu, 1);
+
axpy_ongpu(size, -decay*batch, layer.filters_gpu, 1, layer.filter_updates_gpu, 1);
axpy_ongpu(size, learning_rate/batch, layer.filter_updates_gpu, 1, layer.filters_gpu, 1);
scal_ongpu(size, momentum, layer.filter_updates_gpu, 1);
}
+
--
Gitblit v1.10.0