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 | 41 +++++++++++++++--------------------------
1 files changed, 15 insertions(+), 26 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index bcf307f..9f0a2f8 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -48,13 +48,11 @@
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());
}
-extern "C" void forward_convolutional_layer_gpu(convolutional_layer layer, float *in)
+extern "C" void forward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
{
int i;
int m = layer.n;
@@ -63,9 +61,8 @@
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(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(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;
@@ -74,9 +71,8 @@
activate_array_ongpu(layer.output_gpu, m*n*layer.batch, layer.activation);
}
-extern "C" void backward_convolutional_layer_gpu(convolutional_layer layer, float *in, float *delta_gpu)
+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;
@@ -86,17 +82,17 @@
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);
- if(delta_gpu) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, delta_gpu, 1);
+ if(state.delta) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, state.delta, 1);
for(i = 0; i < layer.batch; ++i){
float * a = layer.delta_gpu;
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,alpha,a + i*m*k,k,b,k,1,c,n);
+ 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,1,a + i*m*k,k,b,k,1,c,n);
- if(delta_gpu){
+ if(state.delta){
float * a = layer.filters_gpu;
float * b = layer.delta_gpu;
@@ -104,7 +100,7 @@
gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
- col2im_ongpu(layer.col_image_gpu, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_gpu + i*layer.c*layer.h*layer.w);
+ 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);
}
}
}
@@ -125,22 +121,15 @@
cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
}
-extern "C" void update_convolutional_layer_gpu(convolutional_layer layer)
+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;
-/*
- 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, 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, layer.learning_rate, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
- scal_ongpu(layer.n,layer.momentum, layer.bias_updates_gpu, 1);
-
- axpy_ongpu(size, -layer.decay, layer.filters_gpu, 1, layer.filter_updates_gpu, 1);
- axpy_ongpu(size, layer.learning_rate, layer.filter_updates_gpu, 1, layer.filters_gpu, 1);
- scal_ongpu(size, layer.momentum, layer.filter_updates_gpu, 1);
- //pull_convolutional_layer(layer);
+ 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