From 18d5e4f39c1441f2c21043ac3204b5cb279f8758 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 24 Aug 2018 15:29:40 +0000
Subject: [PATCH] Fixed yolov3-tiny_xnor.cfg
---
src/maxpool_layer.c | 154 +++++++++++++++++++++++++++++---------------------
1 files changed, 89 insertions(+), 65 deletions(-)
diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c
index ef7176d..5ff7e9d 100644
--- a/src/maxpool_layer.c
+++ b/src/maxpool_layer.c
@@ -1,110 +1,134 @@
#include "maxpool_layer.h"
#include "cuda.h"
+#include "gemm.h"
#include <stdio.h>
-image get_maxpool_image(maxpool_layer layer)
+image get_maxpool_image(maxpool_layer l)
{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- return float_to_image(h,w,c,layer.output);
+ int h = l.out_h;
+ int w = l.out_w;
+ int c = l.c;
+ return float_to_image(w,h,c,l.output);
}
-image get_maxpool_delta(maxpool_layer layer)
+image get_maxpool_delta(maxpool_layer l)
{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- return float_to_image(h,w,c,layer.delta);
+ int h = l.out_h;
+ int w = l.out_w;
+ int c = l.c;
+ 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 *layer = calloc(1, sizeof(maxpool_layer));
- layer->batch = batch;
- layer->h = h;
- layer->w = w;
- layer->c = c;
- layer->size = size;
- layer->stride = stride;
- int output_size = ((h-1)/stride+1) * ((w-1)/stride+1) * c * batch;
- layer->indexes = calloc(output_size, sizeof(int));
- layer->output = calloc(output_size, sizeof(float));
- layer->delta = calloc(output_size, sizeof(float));
+ maxpool_layer l = {0};
+ l.type = MAXPOOL;
+ l.batch = batch;
+ l.h = h;
+ l.w = w;
+ l.c = c;
+ 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;
+ l.size = size;
+ l.stride = stride;
+ int output_size = l.out_h * l.out_w * l.out_c * batch;
+ 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
- layer->indexes_gpu = cuda_make_int_array(output_size);
- layer->output_gpu = cuda_make_array(layer->output, output_size);
- layer->delta_gpu = cuda_make_array(layer->delta, output_size);
+ 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
- return layer;
+ 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 *layer, int h, int w)
+void resize_maxpool_layer(maxpool_layer *l, int w, int h)
{
- layer->h = h;
- layer->w = w;
- int output_size = ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * layer->c * layer->batch;
- layer->output = realloc(layer->output, output_size * sizeof(float));
- layer->delta = realloc(layer->delta, output_size * sizeof(float));
+ l->h = h;
+ l->w = w;
+ l->inputs = h*w*l->c;
+
+ 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;
+
+ l->indexes = realloc(l->indexes, output_size * sizeof(int));
+ l->output = realloc(l->output, output_size * sizeof(float));
+ l->delta = realloc(l->delta, output_size * sizeof(float));
#ifdef GPU
- cuda_free((float *)layer->indexes_gpu);
- cuda_free(layer->output_gpu);
- cuda_free(layer->delta_gpu);
- layer->indexes_gpu = cuda_make_int_array(output_size);
- layer->output_gpu = cuda_make_array(layer->output, output_size);
- layer->delta_gpu = cuda_make_array(layer->delta, output_size);
+ cuda_free((float *)l->indexes_gpu);
+ 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(l->output, output_size);
+ l->delta_gpu = cuda_make_array(l->delta, output_size);
#endif
}
-void forward_maxpool_layer(const maxpool_layer layer, float *input)
+void forward_maxpool_layer(const maxpool_layer l, network_state state)
{
- int b,i,j,k,l,m;
- int w_offset = (-layer.size-1)/2 + 1;
- int h_offset = (-layer.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 = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
+ int b,i,j,k,m,n;
+ int w_offset = -l.pad / 2;
+ int h_offset = -l.pad / 2;
- for(b = 0; b < layer.batch; ++b){
+ int h = l.out_h;
+ int w = l.out_w;
+ int c = l.c;
+
+ for(b = 0; b < l.batch; ++b){
for(k = 0; k < c; ++k){
for(i = 0; i < h; ++i){
for(j = 0; j < w; ++j){
int out_index = j + w*(i + h*(k + c*b));
float max = -FLT_MAX;
int max_i = -1;
- for(l = 0; l < layer.size; ++l){
- for(m = 0; m < layer.size; ++m){
- int cur_h = h_offset + i*layer.stride + l;
- int cur_w = w_offset + j*layer.stride + m;
- int index = cur_w + layer.w*(cur_h + layer.h*(k + b*layer.c));
- int valid = (cur_h >= 0 && cur_h < layer.h &&
- cur_w >= 0 && cur_w < layer.w);
- float val = (valid != 0) ? input[index] : -FLT_MAX;
+ for(n = 0; n < l.size; ++n){
+ for(m = 0; m < l.size; ++m){
+ int cur_h = h_offset + i*l.stride + n;
+ int cur_w = w_offset + j*l.stride + m;
+ int index = cur_w + l.w*(cur_h + l.h*(k + b*l.c));
+ int valid = (cur_h >= 0 && cur_h < l.h &&
+ cur_w >= 0 && cur_w < l.w);
+ float val = (valid != 0) ? state.input[index] : -FLT_MAX;
max_i = (val > max) ? index : max_i;
max = (val > max) ? val : max;
}
}
- layer.output[out_index] = max;
- layer.indexes[out_index] = max_i;
+ l.output[out_index] = max;
+ l.indexes[out_index] = max_i;
}
}
}
}
}
-void backward_maxpool_layer(const maxpool_layer layer, float *delta)
+void backward_maxpool_layer(const maxpool_layer l, network_state state)
{
int i;
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
- for(i = 0; i < h*w*c*layer.batch; ++i){
- int index = layer.indexes[i];
- delta[index] += layer.delta[i];
+ int h = l.out_h;
+ int w = l.out_w;
+ int c = l.c;
+ 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