From cd8a3dcb4ca42f22ad8f46a95e00977c92be6bbd Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Thu, 08 Feb 2018 23:22:42 +0000
Subject: [PATCH] Compile fixes
---
src/convolutional_kernels.cu | 89 ++++++++++++++++++++++++++++++--------------
1 files changed, 60 insertions(+), 29 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 2376835..03c9ab7 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -2,6 +2,10 @@
#include "curand.h"
#include "cublas_v2.h"
+#ifdef CUDNN
+#pragma comment(lib, "cudnn.lib")
+#endif
+
extern "C" {
#include "convolutional_layer.h"
#include "batchnorm_layer.h"
@@ -17,7 +21,7 @@
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if (i >= n) return;
- binary[i] = (x[i] > 0) ? 1 : -1;
+ binary[i] = (x[i] >= 0) ? 1 : -1;
}
void binarize_gpu(float *x, int n, float *binary)
@@ -48,39 +52,38 @@
}
-__global__ void binarize_filters_kernel(float *filters, int n, int size, float *binary)
+__global__ void binarize_weights_kernel(float *weights, 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 += abs(weights[f*size + i]);
}
mean = mean / size;
for(i = 0; i < size; ++i){
- binary[f*size + i] = (filters[f*size + i] > 0) ? mean : -mean;
+ binary[f*size + i] = (weights[f*size + i] > 0) ? mean : -mean;
+ //binary[f*size + i] = weights[f*size + i];
}
}
-void binarize_filters_gpu(float *filters, int n, int size, float *binary)
+void binarize_weights_gpu(float *weights, int n, int size, float *binary)
{
- binarize_filters_kernel<<<cuda_gridsize(n), BLOCK>>>(filters, n, size, binary);
+ binarize_weights_kernel<<<cuda_gridsize(n), BLOCK>>>(weights, n, size, binary);
check_error(cudaPeekAtLastError());
}
void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
{
- int i;
-
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);
+ binarize_weights_gpu(l.weights_gpu, l.n, l.c*l.size*l.size, l.binary_weights_gpu);
swap_binary(&l);
}
if(l.xnor){
- binarize_filters_gpu(l.filters_gpu, l.n, l.c*l.size*l.size, l.binary_filters_gpu);
+ binarize_weights_gpu(l.weights_gpu, l.n, l.c*l.size*l.size, l.binary_weights_gpu);
swap_binary(&l);
binarize_gpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input_gpu);
state.input = l.binary_input_gpu;
@@ -92,8 +95,8 @@
&one,
l.srcTensorDesc,
state.input,
- l.filterDesc,
- l.filters_gpu,
+ l.weightDesc,
+ l.weights_gpu,
l.convDesc,
l.fw_algo,
state.workspace,
@@ -103,12 +106,13 @@
l.output_gpu);
#else
+ int i;
int m = l.n;
int k = l.size*l.size*l.c;
int n = l.out_w*l.out_h;
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, state.workspace);
- float * a = l.filters_gpu;
+ float * a = l.weights_gpu;
float * b = state.workspace;
float * c = l.output_gpu;
gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
@@ -123,6 +127,7 @@
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
//if(l.dot > 0) dot_error_gpu(l);
if(l.binary || l.xnor) swap_binary(&l);
+ //cudaDeviceSynchronize(); // for correct profiling of performance
}
void backward_convolutional_layer_gpu(convolutional_layer l, network_state state)
@@ -133,6 +138,9 @@
if(l.batch_normalize){
backward_batchnorm_layer_gpu(l, state);
+ //axpy_ongpu(l.outputs*l.batch, -state.net.decay, l.x_gpu, 1, l.delta_gpu, 1);
+ } else {
+ //axpy_ongpu(l.outputs*l.batch, -state.net.decay, l.output_gpu, 1, l.delta_gpu, 1);
}
float *original_input = state.input;
@@ -150,15 +158,15 @@
state.workspace,
l.workspace_size,
&one,
- l.dfilterDesc,
- l.filter_updates_gpu);
+ l.dweightDesc,
+ l.weight_updates_gpu);
if(state.delta){
if(l.binary || l.xnor) swap_binary(&l);
cudnnConvolutionBackwardData(cudnn_handle(),
&one,
- l.filterDesc,
- l.filters_gpu,
+ l.weightDesc,
+ l.weights_gpu,
l.ddstTensorDesc,
l.delta_gpu,
l.convDesc,
@@ -181,14 +189,14 @@
for(i = 0; i < l.batch; ++i){
float * a = l.delta_gpu;
float * b = state.workspace;
- float * c = l.filter_updates_gpu;
+ float * c = l.weight_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, state.workspace);
gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
if(state.delta){
if(l.binary || l.xnor) swap_binary(&l);
- float * a = l.filters_gpu;
+ float * a = l.weights_gpu;
float * b = l.delta_gpu;
float * c = state.workspace;
@@ -206,43 +214,66 @@
void pull_convolutional_layer(convolutional_layer layer)
{
- cuda_pull_array(layer.filters_gpu, layer.filters, layer.c*layer.n*layer.size*layer.size);
+ cuda_pull_array(layer.weights_gpu, layer.weights, layer.c*layer.n*layer.size*layer.size);
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.weight_updates_gpu, layer.weight_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);
}
+ if (layer.adam){
+ cuda_pull_array(layer.m_gpu, layer.m, layer.c*layer.n*layer.size*layer.size);
+ cuda_pull_array(layer.v_gpu, layer.v, layer.c*layer.n*layer.size*layer.size);
+ }
}
void push_convolutional_layer(convolutional_layer layer)
{
- cuda_push_array(layer.filters_gpu, layer.filters, layer.c*layer.n*layer.size*layer.size);
+ cuda_push_array(layer.weights_gpu, layer.weights, layer.c*layer.n*layer.size*layer.size);
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.weight_updates_gpu, layer.weight_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);
}
+ if (layer.adam){
+ cuda_push_array(layer.m_gpu, layer.m, layer.c*layer.n*layer.size*layer.size);
+ cuda_push_array(layer.v_gpu, layer.v, layer.c*layer.n*layer.size*layer.size);
+ }
}
void update_convolutional_layer_gpu(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)
{
int size = layer.size*layer.size*layer.c*layer.n;
-
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);
+ if(layer.scales_gpu){
+ 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);
+ if(layer.adam){
+ scal_ongpu(size, layer.B1, layer.m_gpu, 1);
+ scal_ongpu(size, layer.B2, layer.v_gpu, 1);
+
+ axpy_ongpu(size, -decay*batch, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
+
+ axpy_ongpu(size, -(1-layer.B1), layer.weight_updates_gpu, 1, layer.m_gpu, 1);
+ mul_ongpu(size, layer.weight_updates_gpu, 1, layer.weight_updates_gpu, 1);
+ axpy_ongpu(size, (1-layer.B2), layer.weight_updates_gpu, 1, layer.v_gpu, 1);
+
+ adam_gpu(size, layer.weights_gpu, layer.m_gpu, layer.v_gpu, layer.B1, layer.B2, learning_rate/batch, layer.eps, layer.t+1);
+ fill_ongpu(size, 0, layer.weight_updates_gpu, 1);
+ }else{
+ axpy_ongpu(size, -decay*batch, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
+ axpy_ongpu(size, learning_rate/batch, layer.weight_updates_gpu, 1, layer.weights_gpu, 1);
+ scal_ongpu(size, momentum, layer.weight_updates_gpu, 1);
+ }
}
--
Gitblit v1.10.0