From 0b2be9d76a369e28fd1fc384b3952621921c9ae5 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 12 Jul 2015 05:52:43 +0000
Subject: [PATCH] Small version of YOLO, ~1GB RAM
---
src/maxpool_layer.c | 206 +++++++++++++++++++++++----------------------------
1 files changed, 94 insertions(+), 112 deletions(-)
diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c
index 08c9f2f..bc3aa68 100644
--- a/src/maxpool_layer.c
+++ b/src/maxpool_layer.c
@@ -1,141 +1,123 @@
#include "maxpool_layer.h"
+#include "cuda.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)
{
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;
- layer->output = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * c*batch, sizeof(float));
- layer->delta = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * c*batch, sizeof(float));
- return layer;
+ 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.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));
+ #ifdef 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 l;
}
-void resize_maxpool_layer(maxpool_layer *layer, int h, int w, int c)
+void resize_maxpool_layer(maxpool_layer *l, int w, int h)
{
- layer->h = h;
- layer->w = w;
- layer->c = c;
- layer->output = realloc(layer->output, ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * c * layer->batch* sizeof(float));
- layer->delta = realloc(layer->delta, ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * c * layer->batch*sizeof(float));
+ int stride = l->stride;
+ l->h = h;
+ l->w = w;
+
+ l->out_w = (w-1)/stride + 1;
+ l->out_h = (h-1)/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 *)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(0, output_size);
+ l->delta_gpu = cuda_make_array(0, output_size);
+ #endif
}
-float get_max_region(image im, int h, int w, int c, int size)
+void forward_maxpool_layer(const maxpool_layer l, network_state state)
{
- int i,j;
- int lower = (-size-1)/2 + 1;
- int upper = size/2 + 1;
-
- int lh = (h-lower < 0) ? 0 : h-lower;
- int uh = (h+upper > im.h) ? im.h : h+upper;
+ int b,i,j,k,m,n;
+ int w_offset = (-l.size-1)/2 + 1;
+ int h_offset = (-l.size-1)/2 + 1;
- int lw = (w-lower < 0) ? 0 : w-lower;
- int uw = (w+upper > im.w) ? im.w : w+upper;
-
- //printf("%d\n", -3/2);
- //printf("%d %d\n", lower, upper);
- //printf("%d %d %d %d\n", lh, uh, lw, uw);
-
- float max = -FLT_MAX;
- for(i = lh; i < uh; ++i){
- for(j = lw; j < uw; ++j){
- float val = get_pixel(im, i, j, c);
- if (val > max) max = val;
- }
- }
- return max;
-}
+ int h = (l.h-1)/l.stride + 1;
+ int w = (l.w-1)/l.stride + 1;
+ int c = l.c;
-void forward_maxpool_layer(const maxpool_layer layer, float *in)
-{
- int b;
- for(b = 0; b < layer.batch; ++b){
- image input = float_to_image(layer.h, layer.w, layer.c, in+b*layer.h*layer.w*layer.c);
-
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- image output = float_to_image(h,w,c,layer.output+b*h*w*c);
-
- int i,j,k;
- for(k = 0; k < input.c; ++k){
- for(i = 0; i < input.h; i += layer.stride){
- for(j = 0; j < input.w; j += layer.stride){
- float max = get_max_region(input, i, j, k, layer.size);
- set_pixel(output, i/layer.stride, j/layer.stride, k, max);
+ 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(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;
+ }
+ }
+ l.output[out_index] = max;
+ l.indexes[out_index] = max_i;
}
}
}
}
}
-float set_max_region_delta(image im, image delta, int h, int w, int c, int size, float max, float error)
+void backward_maxpool_layer(const maxpool_layer l, network_state state)
{
- int i,j;
- int lower = (-size-1)/2 + 1;
- int upper = size/2 + 1;
-
- int lh = (h-lower < 0) ? 0 : h-lower;
- int uh = (h+upper > im.h) ? im.h : h+upper;
-
- int lw = (w-lower < 0) ? 0 : w-lower;
- int uw = (w+upper > im.w) ? im.w : w+upper;
-
- for(i = lh; i < uh; ++i){
- for(j = lw; j < uw; ++j){
- float val = get_pixel(im, i, j, c);
- if (val == max){
- add_pixel(delta, i, j, c, error);
- }
- }
- }
- return max;
-}
-
-void backward_maxpool_layer(const maxpool_layer layer, float *in, float *delta)
-{
- int b;
- for(b = 0; b < layer.batch; ++b){
- image input = float_to_image(layer.h, layer.w, layer.c, in+b*layer.h*layer.w*layer.c);
- image input_delta = float_to_image(layer.h, layer.w, layer.c, delta+b*layer.h*layer.w*layer.c);
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- image output = float_to_image(h,w,c,layer.output+b*h*w*c);
- image output_delta = float_to_image(h,w,c,layer.delta+b*h*w*c);
- zero_image(input_delta);
-
- int i,j,k;
- for(k = 0; k < input.c; ++k){
- for(i = 0; i < input.h; i += layer.stride){
- for(j = 0; j < input.w; j += layer.stride){
- float max = get_pixel(output, i/layer.stride, j/layer.stride, k);
- float error = get_pixel(output_delta, i/layer.stride, j/layer.stride, k);
- set_max_region_delta(input, input_delta, i, j, k, layer.size, max, error);
- }
- }
- }
+ int i;
+ int h = (l.h-1)/l.stride + 1;
+ int w = (l.w-1)/l.stride + 1;
+ 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