OrgLion-ML
/
darknet
mirror of https://github.com/AlexeyAB/darknet.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Another fix for [batchnorm] layer

This commit is contained in:
AlexeyAB 2019-12-20 22:49:09 +03:00
parent d6eaa2e95f
commit 5f6dced3ce
5 changed files with 67 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
src/batchnorm_layer.c
View File

@ -2,12 +2,13 @@
#include "blas.h"
#include <stdio.h>
layer make_batchnorm_layer(int batch, int w, int h, int c)
layer make_batchnorm_layer(int batch, int w, int h, int c, int train)
{
fprintf(stderr, "Batch Normalization Layer: %d x %d x %d image\n", w,h,c);
layer layer = { (LAYER_TYPE)0 };
layer.type = BATCHNORM;
layer.batch = batch;
layer.train = train;
layer.h = layer.out_h = h;
layer.w = layer.out_w = w;
layer.c = layer.out_c = c;
@ -42,13 +43,19 @@ layer make_batchnorm_layer(int batch, int w, int h, int c)
layer.update_gpu = update_batchnorm_layer_gpu;
layer.output_gpu = cuda_make_array(layer.output, h * w * c * batch);
layer.delta_gpu = cuda_make_array(layer.delta, h * w * c * batch);
layer.biases_gpu = cuda_make_array(layer.biases, c);
layer.bias_updates_gpu = cuda_make_array(layer.bias_updates, c);
layer.scales_gpu = cuda_make_array(layer.scales, c);
layer.scale_updates_gpu = cuda_make_array(layer.scale_updates, c);
if (train) {
layer.delta_gpu = cuda_make_array(layer.delta, h * w * c * batch);
layer.bias_updates_gpu = cuda_make_array(layer.bias_updates, c);
layer.scale_updates_gpu = cuda_make_array(layer.scale_updates, c);
layer.mean_delta_gpu = cuda_make_array(layer.mean, c);
layer.variance_delta_gpu = cuda_make_array(layer.variance, c);
}
layer.mean_gpu = cuda_make_array(layer.mean, c);
layer.variance_gpu = cuda_make_array(layer.variance, c);
@ -56,16 +63,18 @@ layer make_batchnorm_layer(int batch, int w, int h, int c)
layer.rolling_mean_gpu = cuda_make_array(layer.mean, c);
layer.rolling_variance_gpu = cuda_make_array(layer.variance, c);
layer.mean_delta_gpu = cuda_make_array(layer.mean, c);
layer.variance_delta_gpu = cuda_make_array(layer.variance, c);
if (train) {
layer.x_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
#ifndef CUDNN
layer.x_norm_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
#endif // not CUDNN
}
layer.x_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
layer.x_norm_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
#ifdef CUDNN
cudnnCreateTensorDescriptor(&layer.normTensorDesc);
cudnnCreateTensorDescriptor(&layer.normDstTensorDesc);
cudnnSetTensor4dDescriptor(layer.normDstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, layer.batch, layer.out_c, layer.out_h, layer.out_w);
cudnnSetTensor4dDescriptor(layer.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, layer.out_c, 1, 1);
CHECK_CUDNN(cudnnCreateTensorDescriptor(&layer.normTensorDesc));
CHECK_CUDNN(cudnnCreateTensorDescriptor(&layer.normDstTensorDesc));
CHECK_CUDNN(cudnnSetTensor4dDescriptor(layer.normDstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, layer.batch, layer.out_c, layer.out_h, layer.out_w));
CHECK_CUDNN(cudnnSetTensor4dDescriptor(layer.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, layer.out_c, 1, 1));
#endif
#endif
return layer;
@ -129,9 +138,40 @@ void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_del
}
}
void resize_batchnorm_layer(layer *layer, int w, int h)
void resize_batchnorm_layer(layer *l, int w, int h)
{
fprintf(stderr, "Not implemented\n");
l->out_h = l->h = h;
l->out_w = l->w = w;
l->outputs = l->inputs = h*w*l->c;
const int output_size = l->outputs * l->batch;
l->output = (float*)realloc(l->output, output_size * sizeof(float));
l->delta = (float*)realloc(l->delta, output_size * sizeof(float));
#ifdef GPU
cuda_free(l->output_gpu);
l->output_gpu = cuda_make_array(l->output, output_size);
if (l->train) {
cuda_free(l->delta_gpu);
l->delta_gpu = cuda_make_array(l->delta, output_size);
cuda_free(l->x_gpu);
l->x_gpu = cuda_make_array(l->output, output_size);
#ifndef CUDNN
cuda_free(l->x_norm_gpu);
l->x_norm_gpu = cuda_make_array(l->output, output_size);
#endif // not CUDNN
}
#ifdef CUDNN
CHECK_CUDNN(cudnnDestroyTensorDescriptor(l->normDstTensorDesc));
CHECK_CUDNN(cudnnCreateTensorDescriptor(&l->normDstTensorDesc));
CHECK_CUDNN(cudnnSetTensor4dDescriptor(l->normDstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w));
#endif // CUDNN
#endif // GPU
}
void forward_batchnorm_layer(layer l, network_state state)
@ -157,6 +197,7 @@ void forward_batchnorm_layer(layer l, network_state state)
normalize_cpu(l.output, l.rolling_mean, l.rolling_variance, l.batch, l.out_c, l.out_h*l.out_w);
}
scale_bias(l.output, l.scales, l.batch, l.out_c, l.out_h*l.out_w);
add_bias(l.output, l.biases, l.batch, l.out_c, l.out_w*l.out_h);
}
void backward_batchnorm_layer(const layer l, network_state state)
@ -188,12 +229,14 @@ void update_batchnorm_layer(layer l, int batch, float learning_rate, float momen
void pull_batchnorm_layer(layer l)
{
cuda_pull_array(l.biases_gpu, l.biases, l.c);
cuda_pull_array(l.scales_gpu, l.scales, l.c);
cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.c);
cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.c);
}
void push_batchnorm_layer(layer l)
{
cuda_push_array(l.biases_gpu, l.biases, l.c);
cuda_push_array(l.scales_gpu, l.scales, l.c);
cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.c);
cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.c);

4
src/batchnorm_layer.h
View File

@ -8,11 +8,13 @@
#ifdef __cplusplus
extern "C" {
#endif
layer make_batchnorm_layer(int batch, int w, int h, int c);
layer make_batchnorm_layer(int batch, int w, int h, int c, int train);
void forward_batchnorm_layer(layer l, network_state state);
void backward_batchnorm_layer(layer l, network_state state);
void update_batchnorm_layer(layer l, int batch, float learning_rate, float momentum, float decay);
void resize_batchnorm_layer(layer *l, int w, int h);
#ifdef GPU
void forward_batchnorm_layer_gpu(layer l, network_state state);
void backward_batchnorm_layer_gpu(layer l, network_state state);

1
src/blas.h
View File

@ -37,6 +37,7 @@ void mean_cpu(float *x, int batch, int filters, int spatial, float *mean);
void variance_cpu(float *x, float *mean, int batch, int filters, int spatial, float *variance);
void normalize_cpu(float *x, float *mean, float *variance, int batch, int filters, int spatial);
void add_bias(float *output, float *biases, int batch, int n, int size);
void scale_bias(float *output, float *scales, int batch, int n, int size);
void backward_scale_cpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates);
void mean_delta_cpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta);

4
src/network.c
View File

@ -527,6 +527,8 @@ int resize_network(network *net, int w, int h)
resize_maxpool_layer(&l, w, h);
}else if (l.type == LOCAL_AVGPOOL) {
resize_maxpool_layer(&l, w, h);
}else if (l.type == BATCHNORM) {
resize_batchnorm_layer(&l, w, h);
}else if(l.type == REGION){
resize_region_layer(&l, w, h);
}else if (l.type == YOLO) {
@ -1079,7 +1081,6 @@ void fuse_conv_batchnorm(network net)
int f;
for (f = 0; f < l->n; ++f)
{
//l->biases[f] = l->biases[f] - (double)l->scales[f] * l->rolling_mean[f] / (sqrt((double)l->rolling_variance[f]) + .000001f);
l->biases[f] = l->biases[f] - (double)l->scales[f] * l->rolling_mean[f] / (sqrt((double)l->rolling_variance[f] + .000001));
const size_t filter_size = l->size*l->size*l->c / l->groups;
@ -1087,7 +1088,6 @@ void fuse_conv_batchnorm(network net)
for (i = 0; i < filter_size; ++i) {
int w_index = f*filter_size + i;
//l->weights[w_index] = (double)l->weights[w_index] * l->scales[f] / (sqrt((double)l->rolling_variance[f]) + .000001f);
l->weights[w_index] = (double)l->weights[w_index] * l->scales[f] / (sqrt((double)l->rolling_variance[f] + .000001));
}
}

4
src/parser.c
View File

@ -805,7 +805,7 @@ layer parse_normalization(list *options, size_params params)
layer parse_batchnorm(list *options, size_params params)
{
layer l = make_batchnorm_layer(params.batch, params.w, params.h, params.c);
layer l = make_batchnorm_layer(params.batch, params.w, params.h, params.c, params.train);
return l;
}
@ -1507,6 +1507,7 @@ void save_batchnorm_weights(layer l, FILE *fp)
pull_batchnorm_layer(l);
}
#endif
fwrite(l.biases, sizeof(float), l.c, fp);
fwrite(l.scales, sizeof(float), l.c, fp);
fwrite(l.rolling_mean, sizeof(float), l.c, fp);
fwrite(l.rolling_variance, sizeof(float), l.c, fp);
@ -1652,6 +1653,7 @@ void load_connected_weights(layer l, FILE *fp, int transpose)
void load_batchnorm_weights(layer l, FILE *fp)
{
fread(l.biases, sizeof(float), l.c, fp);
fread(l.scales, sizeof(float), l.c, fp);
fread(l.rolling_mean, sizeof(float), l.c, fp);
fread(l.rolling_variance, sizeof(float), l.c, fp);