From e3ee7b7cd64894144a594c0caaa202f7aeac6723 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 15 Mar 2016 02:35:36 +0000
Subject: [PATCH] Hi Harley
---
src/connected_layer.c | 153 +++++++++++++++++++++++++++++++++++++++++++++-----
1 files changed, 137 insertions(+), 16 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index 2d83dd9..df78e67 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -9,7 +9,7 @@
#include <stdlib.h>
#include <string.h>
-connected_layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation)
+connected_layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize)
{
int i;
connected_layer l = {0};
@@ -18,9 +18,10 @@
l.inputs = inputs;
l.outputs = outputs;
l.batch=batch;
+ l.batch_normalize = batch_normalize;
- l.output = calloc(batch*outputs, sizeof(float*));
- l.delta = calloc(batch*outputs, sizeof(float*));
+ l.output = calloc(batch*outputs, sizeof(float));
+ l.delta = calloc(batch*outputs, sizeof(float));
l.weight_updates = calloc(inputs*outputs, sizeof(float));
l.bias_updates = calloc(outputs, sizeof(float));
@@ -32,13 +33,32 @@
//float scale = 1./sqrt(inputs);
float scale = sqrt(2./inputs);
for(i = 0; i < outputs*inputs; ++i){
- l.weights[i] = 2*scale*rand_uniform() - scale;
+ l.weights[i] = scale*rand_uniform(-1, 1);
}
for(i = 0; i < outputs; ++i){
l.biases[i] = scale;
}
+ if(batch_normalize){
+ l.scales = calloc(outputs, sizeof(float));
+ l.scale_updates = calloc(outputs, sizeof(float));
+ for(i = 0; i < outputs; ++i){
+ l.scales[i] = 1;
+ }
+
+ l.mean = calloc(outputs, sizeof(float));
+ l.mean_delta = calloc(outputs, sizeof(float));
+ l.variance = calloc(outputs, sizeof(float));
+ l.variance_delta = calloc(outputs, sizeof(float));
+
+ l.rolling_mean = calloc(outputs, sizeof(float));
+ l.rolling_variance = calloc(outputs, sizeof(float));
+
+ l.x = calloc(batch*outputs, sizeof(float));
+ l.x_norm = calloc(batch*outputs, sizeof(float));
+ }
+
#ifdef GPU
l.weights_gpu = cuda_make_array(l.weights, outputs*inputs);
l.biases_gpu = cuda_make_array(l.biases, outputs);
@@ -48,6 +68,22 @@
l.output_gpu = cuda_make_array(l.output, outputs*batch);
l.delta_gpu = cuda_make_array(l.delta, outputs*batch);
+ if(batch_normalize){
+ l.scales_gpu = cuda_make_array(l.scales, outputs);
+ l.scale_updates_gpu = cuda_make_array(l.scale_updates, outputs);
+
+ l.mean_gpu = cuda_make_array(l.mean, outputs);
+ l.variance_gpu = cuda_make_array(l.variance, outputs);
+
+ l.rolling_mean_gpu = cuda_make_array(l.mean, outputs);
+ l.rolling_variance_gpu = cuda_make_array(l.variance, outputs);
+
+ l.mean_delta_gpu = cuda_make_array(l.mean, outputs);
+ l.variance_delta_gpu = cuda_make_array(l.variance, outputs);
+
+ l.x_gpu = cuda_make_array(l.output, l.batch*outputs);
+ l.x_norm_gpu = cuda_make_array(l.output, l.batch*outputs);
+ }
#endif
l.activation = activation;
fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
@@ -59,6 +95,11 @@
axpy_cpu(l.outputs, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
scal_cpu(l.outputs, momentum, l.bias_updates, 1);
+ if(l.batch_normalize){
+ axpy_cpu(l.outputs, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
+ scal_cpu(l.outputs, momentum, l.scale_updates, 1);
+ }
+
axpy_cpu(l.inputs*l.outputs, -decay*batch, l.weights, 1, l.weight_updates, 1);
axpy_cpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
scal_cpu(l.inputs*l.outputs, momentum, l.weight_updates, 1);
@@ -67,9 +108,7 @@
void forward_connected_layer(connected_layer l, network_state state)
{
int i;
- for(i = 0; i < l.batch; ++i){
- copy_cpu(l.outputs, l.biases, 1, l.output + i*l.outputs, 1);
- }
+ fill_cpu(l.outputs*l.batch, 0, l.output, 1);
int m = l.batch;
int k = l.inputs;
int n = l.outputs;
@@ -77,6 +116,27 @@
float *b = l.weights;
float *c = l.output;
gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+ if(l.batch_normalize){
+ if(state.train){
+ mean_cpu(l.output, l.batch, l.outputs, 1, l.mean);
+ variance_cpu(l.output, l.mean, l.batch, l.outputs, 1, l.variance);
+
+ scal_cpu(l.outputs, .95, l.rolling_mean, 1);
+ axpy_cpu(l.outputs, .05, l.mean, 1, l.rolling_mean, 1);
+ scal_cpu(l.outputs, .95, l.rolling_variance, 1);
+ axpy_cpu(l.outputs, .05, l.variance, 1, l.rolling_variance, 1);
+
+ copy_cpu(l.outputs*l.batch, l.output, 1, l.x, 1);
+ normalize_cpu(l.output, l.mean, l.variance, l.batch, l.outputs, 1);
+ copy_cpu(l.outputs*l.batch, l.output, 1, l.x_norm, 1);
+ } else {
+ normalize_cpu(l.output, l.rolling_mean, l.rolling_variance, l.batch, l.outputs, 1);
+ }
+ scale_bias(l.output, l.scales, l.batch, l.outputs, 1);
+ }
+ for(i = 0; i < l.batch; ++i){
+ axpy_cpu(l.outputs, 1, l.biases, 1, l.output + i*l.outputs, 1);
+ }
activate_array(l.output, l.outputs*l.batch, l.activation);
}
@@ -87,6 +147,16 @@
for(i = 0; i < l.batch; ++i){
axpy_cpu(l.outputs, 1, l.delta + i*l.outputs, 1, l.bias_updates, 1);
}
+ if(l.batch_normalize){
+ backward_scale_cpu(l.x_norm, l.delta, l.batch, l.outputs, 1, l.scale_updates);
+
+ scale_bias(l.delta, l.scales, l.batch, l.outputs, 1);
+
+ mean_delta_cpu(l.delta, l.variance, l.batch, l.outputs, 1, l.mean_delta);
+ variance_delta_cpu(l.x, l.delta, l.mean, l.variance, l.batch, l.outputs, 1, l.variance_delta);
+ normalize_delta_cpu(l.x, l.mean, l.variance, l.mean_delta, l.variance_delta, l.batch, l.outputs, 1, l.delta);
+ }
+
int m = l.outputs;
int k = l.batch;
int n = l.inputs;
@@ -114,6 +184,11 @@
cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+ if (l.batch_normalize){
+ cuda_pull_array(l.scales_gpu, l.scales, l.outputs);
+ cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
+ cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
+ }
}
void push_connected_layer(connected_layer l)
@@ -122,6 +197,11 @@
cuda_push_array(l.biases_gpu, l.biases, l.outputs);
cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+ if (l.batch_normalize){
+ cuda_push_array(l.scales_gpu, l.scales, l.outputs);
+ cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
+ cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
+ }
}
void update_connected_layer_gpu(connected_layer l, int batch, float learning_rate, float momentum, float decay)
@@ -129,6 +209,11 @@
axpy_ongpu(l.outputs, learning_rate/batch, l.bias_updates_gpu, 1, l.biases_gpu, 1);
scal_ongpu(l.outputs, momentum, l.bias_updates_gpu, 1);
+ if(l.batch_normalize){
+ axpy_ongpu(l.outputs, learning_rate/batch, l.scale_updates_gpu, 1, l.scales_gpu, 1);
+ scal_ongpu(l.outputs, momentum, l.scale_updates_gpu, 1);
+ }
+
axpy_ongpu(l.inputs*l.outputs, -decay*batch, l.weights_gpu, 1, l.weight_updates_gpu, 1);
axpy_ongpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates_gpu, 1, l.weights_gpu, 1);
scal_ongpu(l.inputs*l.outputs, momentum, l.weight_updates_gpu, 1);
@@ -137,9 +222,12 @@
void forward_connected_layer_gpu(connected_layer l, network_state state)
{
int i;
- for(i = 0; i < l.batch; ++i){
- copy_ongpu_offset(l.outputs, l.biases_gpu, 0, 1, l.output_gpu, i*l.outputs, 1);
- }
+ fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
+ /*
+ for(i = 0; i < l.batch; ++i){
+ copy_ongpu_offset(l.outputs, l.biases_gpu, 0, 1, l.output_gpu, i*l.outputs, 1);
+ }
+ */
int m = l.batch;
int k = l.inputs;
int n = l.outputs;
@@ -147,13 +235,35 @@
float * b = l.weights_gpu;
float * c = l.output_gpu;
gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
+ if(l.batch_normalize){
+ if(state.train){
+ fast_mean_gpu(l.output_gpu, l.batch, l.outputs, 1, l.mean_gpu);
+ fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.outputs, 1, l.variance_gpu);
+
+ scal_ongpu(l.outputs, .95, l.rolling_mean_gpu, 1);
+ axpy_ongpu(l.outputs, .05, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
+ scal_ongpu(l.outputs, .95, l.rolling_variance_gpu, 1);
+ axpy_ongpu(l.outputs, .05, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
+
+ 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.outputs, 1);
+ 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.outputs, 1);
+ }
+
+ scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.outputs, 1);
+ }
+ for(i = 0; i < l.batch; ++i){
+ axpy_ongpu(l.outputs, 1, l.biases_gpu, 1, l.output_gpu + i*l.outputs, 1);
+ }
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
-/*
- cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
- float avg = mean_array(l.output, l.outputs*l.batch);
- printf("%f\n", avg);
- */
+ /*
+ cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
+ float avg = mean_array(l.output, l.outputs*l.batch);
+ printf("%f\n", avg);
+ */
}
void backward_connected_layer_gpu(connected_layer l, network_state state)
@@ -161,8 +271,19 @@
int i;
gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
for(i = 0; i < l.batch; ++i){
- axpy_ongpu_offset(l.outputs, 1, l.delta_gpu, i*l.outputs, 1, l.bias_updates_gpu, 0, 1);
+ axpy_ongpu(l.outputs, 1, l.delta_gpu + i*l.outputs, 1, l.bias_updates_gpu, 1);
}
+
+ if(l.batch_normalize){
+ backward_scale_gpu(l.x_norm_gpu, l.delta_gpu, l.batch, l.outputs, 1, l.scale_updates_gpu);
+
+ scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.outputs, 1);
+
+ fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.outputs, 1, l.mean_delta_gpu);
+ fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.outputs, 1, 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.outputs, 1, l.delta_gpu);
+ }
+
int m = l.outputs;
int k = l.batch;
int n = l.inputs;
--
Gitblit v1.10.0