From 5b6be00d4b1ffd671c20c4c72d2239c924eaa3d4 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Thu, 23 Aug 2018 12:28:34 +0000
Subject: [PATCH] Added yolov3-tiny_xnor.cfg
---
src/maxpool_layer.c | 45 ++++++++++++++++++++++++++++-----------------
1 files changed, 28 insertions(+), 17 deletions(-)
diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c
index bc3aa68..5ff7e9d 100644
--- a/src/maxpool_layer.c
+++ b/src/maxpool_layer.c
@@ -1,5 +1,6 @@
#include "maxpool_layer.h"
#include "cuda.h"
+#include "gemm.h"
#include <stdio.h>
image get_maxpool_image(maxpool_layer l)
@@ -18,17 +19,17 @@
return float_to_image(w,h,c,l.delta);
}
-maxpool_layer make_maxpool_layer(int batch, int h, int w, int c, int size, int stride)
+maxpool_layer make_maxpool_layer(int batch, int h, int w, int c, int size, int stride, int padding)
{
- fprintf(stderr, "Maxpool Layer: %d x %d x %d image, %d size, %d stride\n", h,w,c,size,stride);
maxpool_layer l = {0};
l.type = MAXPOOL;
l.batch = batch;
l.h = h;
l.w = w;
l.c = c;
- l.out_w = (w-1)/stride + 1;
- l.out_h = (h-1)/stride + 1;
+ l.pad = padding;
+ l.out_w = (w + padding - size) / stride + 1;
+ l.out_h = (h + padding - size) / stride + 1;
l.out_c = c;
l.outputs = l.out_h * l.out_w * l.out_c;
l.inputs = h*w*c;
@@ -38,22 +39,28 @@
l.indexes = calloc(output_size, sizeof(int));
l.output = calloc(output_size, sizeof(float));
l.delta = calloc(output_size, sizeof(float));
+ l.forward = forward_maxpool_layer;
+ l.backward = backward_maxpool_layer;
#ifdef GPU
+ l.forward_gpu = forward_maxpool_layer_gpu;
+ l.backward_gpu = backward_maxpool_layer_gpu;
l.indexes_gpu = cuda_make_int_array(output_size);
l.output_gpu = cuda_make_array(l.output, output_size);
l.delta_gpu = cuda_make_array(l.delta, output_size);
#endif
+ l.bflops = (l.size*l.size*l.c * l.out_h*l.out_w) / 1000000000.;
+ fprintf(stderr, "max %d x %d / %d %4d x%4d x%4d -> %4d x%4d x%4d %5.3f BF\n", size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
return l;
}
void resize_maxpool_layer(maxpool_layer *l, int w, int h)
{
- int stride = l->stride;
l->h = h;
l->w = w;
+ l->inputs = h*w*l->c;
- l->out_w = (w-1)/stride + 1;
- l->out_h = (h-1)/stride + 1;
+ l->out_w = (w + l->pad - l->size) / l->stride + 1;
+ l->out_h = (h + l->pad - l->size) / l->stride + 1;
l->outputs = l->out_w * l->out_h * l->c;
int output_size = l->outputs * l->batch;
@@ -66,19 +73,24 @@
cuda_free(l->output_gpu);
cuda_free(l->delta_gpu);
l->indexes_gpu = cuda_make_int_array(output_size);
- l->output_gpu = cuda_make_array(0, output_size);
- l->delta_gpu = cuda_make_array(0, output_size);
+ l->output_gpu = cuda_make_array(l->output, output_size);
+ l->delta_gpu = cuda_make_array(l->delta, output_size);
#endif
}
void forward_maxpool_layer(const maxpool_layer l, network_state state)
{
- int b,i,j,k,m,n;
- int w_offset = (-l.size-1)/2 + 1;
- int h_offset = (-l.size-1)/2 + 1;
+ if (!state.train) {
+ forward_maxpool_layer_avx(state.input, l.output, l.indexes, l.size, l.w, l.h, l.out_w, l.out_h, l.c, l.pad, l.stride, l.batch);
+ return;
+ }
- int h = (l.h-1)/l.stride + 1;
- int w = (l.w-1)/l.stride + 1;
+ int b,i,j,k,m,n;
+ int w_offset = -l.pad / 2;
+ int h_offset = -l.pad / 2;
+
+ int h = l.out_h;
+ int w = l.out_w;
int c = l.c;
for(b = 0; b < l.batch; ++b){
@@ -111,10 +123,9 @@
void backward_maxpool_layer(const maxpool_layer l, network_state state)
{
int i;
- int h = (l.h-1)/l.stride + 1;
- int w = (l.w-1)/l.stride + 1;
+ int h = l.out_h;
+ int w = l.out_w;
int c = l.c;
- memset(state.delta, 0, l.batch*l.h*l.w*l.c*sizeof(float));
for(i = 0; i < h*w*c*l.batch; ++i){
int index = l.indexes[i];
state.delta[index] += l.delta[i];
--
Gitblit v1.10.0