From 664c5dd2f2d1c4ad177d5122df6ce3e2900c6648 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 22 Mar 2015 16:56:40 +0000
Subject: [PATCH] Subdivisions for batches
---
src/convolutional_kernels.cu | 16 ++++++----------
1 files changed, 6 insertions(+), 10 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 77304aa..9f0a2f8 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -48,9 +48,7 @@
extern "C" void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
{
- float alpha = 1./batch;
-
- backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size, alpha);
+ backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size, 1);
check_error(cudaPeekAtLastError());
}
@@ -63,7 +61,6 @@
convolutional_out_width(layer);
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;
@@ -76,7 +73,6 @@
extern "C" void backward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
{
- float alpha = 1./layer.batch;
int i;
int m = layer.n;
int n = layer.size*layer.size*layer.c;
@@ -94,7 +90,7 @@
float * c = layer.filter_updates_gpu;
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);
- gemm_ongpu(0,1,m,n,k,alpha,a + i*m*k,k,b,k,1,c,n);
+ gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
if(state.delta){
@@ -125,15 +121,15 @@
cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
}
-extern "C" void update_convolutional_layer_gpu(convolutional_layer layer, float learning_rate, float momentum, float decay)
+extern "C" 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, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
+ 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(size, -decay, layer.filters_gpu, 1, layer.filter_updates_gpu, 1);
- axpy_ongpu(size, learning_rate, layer.filter_updates_gpu, 1, layer.filters_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