From bb48b9c992e2f48bff2432a0387681cad0c98dec Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Thu, 23 Aug 2018 00:00:48 +0000
Subject: [PATCH] Merge branch 'master' of github.com:AlexeyAB/darknet
---
src/connected_layer.c | 184 +++++++++++++++++++++++++++++++++++++++++----
1 files changed, 165 insertions(+), 19 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index 2d83dd9..e6dc759 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,4 +1,5 @@
#include "connected_layer.h"
+#include "batchnorm_layer.h"
#include "utils.h"
#include "cuda.h"
#include "blas.h"
@@ -9,7 +10,7 @@
#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 +19,16 @@
l.inputs = inputs;
l.outputs = outputs;
l.batch=batch;
+ l.batch_normalize = batch_normalize;
+ l.h = 1;
+ l.w = 1;
+ l.c = inputs;
+ l.out_h = 1;
+ l.out_w = 1;
+ l.out_c = outputs;
- 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));
@@ -28,18 +36,44 @@
l.weights = calloc(outputs*inputs, sizeof(float));
l.biases = calloc(outputs, sizeof(float));
+ l.forward = forward_connected_layer;
+ l.backward = backward_connected_layer;
+ l.update = update_connected_layer;
//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;
+ l.biases[i] = 0;
+ }
+
+ 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.forward_gpu = forward_connected_layer_gpu;
+ l.backward_gpu = backward_connected_layer_gpu;
+ l.update_gpu = update_connected_layer_gpu;
+
l.weights_gpu = cuda_make_array(l.weights, outputs*inputs);
l.biases_gpu = cuda_make_array(l.biases, outputs);
@@ -48,9 +82,31 @@
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);
+#ifdef CUDNN
+ cudnnCreateTensorDescriptor(&l.normTensorDesc);
+ cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+ cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
+ cudnnSetTensor4dDescriptor(l.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l.out_c, 1, 1);
+#endif
+ }
#endif
l.activation = activation;
- fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
+ fprintf(stderr, "connected %4d -> %4d\n", inputs, outputs);
return l;
}
@@ -59,6 +115,11 @@
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 +128,7 @@
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 +136,27 @@
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 +167,16 @@
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;
@@ -106,6 +196,41 @@
if(c) gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
}
+
+void denormalize_connected_layer(layer l)
+{
+ int i, j;
+ for(i = 0; i < l.outputs; ++i){
+ float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .000001);
+ for(j = 0; j < l.inputs; ++j){
+ l.weights[i*l.inputs + j] *= scale;
+ }
+ l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
+ }
+}
+
+
+void statistics_connected_layer(layer l)
+{
+ if(l.batch_normalize){
+ printf("Scales ");
+ print_statistics(l.scales, l.outputs);
+ /*
+ printf("Rolling Mean ");
+ print_statistics(l.rolling_mean, l.outputs);
+ printf("Rolling Variance ");
+ print_statistics(l.rolling_variance, l.outputs);
+ */
+ }
+ printf("Biases ");
+ print_statistics(l.biases, l.outputs);
+ printf("Weights ");
+ print_statistics(l.weights, l.outputs);
+}
+
#ifdef GPU
void pull_connected_layer(connected_layer l)
@@ -114,6 +239,11 @@
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 +252,11 @@
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 +264,11 @@
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 +277,8 @@
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);
+
int m = l.batch;
int k = l.inputs;
int n = l.outputs;
@@ -147,22 +286,29 @@
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) {
+ forward_batchnorm_layer_gpu(l, state);
+ }
+ else {
+ add_bias_gpu(l.output_gpu, l.biases_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);
- */
}
void backward_connected_layer_gpu(connected_layer l, network_state state)
{
int i;
+ constrain_ongpu(l.outputs*l.batch, 1, l.delta_gpu, 1);
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_batchnorm_layer_gpu(l, state);
+ }
+
int m = l.outputs;
int k = l.batch;
int n = l.inputs;
--
Gitblit v1.10.0