From 516f019ba6fb88de7218dd3b4eaeadb1cf676518 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 11 May 2015 20:46:49 +0000
Subject: [PATCH] route handles input images well....ish
---
src/maxpool_layer.c | 201 ++++++++++++++++++--------------------------------
1 files changed, 73 insertions(+), 128 deletions(-)
diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c
index c05e939..c7739f1 100644
--- a/src/maxpool_layer.c
+++ b/src/maxpool_layer.c
@@ -1,171 +1,116 @@
#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.h-1)/l.stride + 1;
+ int w = (l.w-1)/l.stride + 1;
+ 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.h-1)/l.stride + 1;
+ int w = (l.w-1)/l.stride + 1;
+ 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;
+ maxpool_layer l = {0};
+ l.type = MAXPOOL;
+ l.batch = batch;
+ l.h = h;
+ l.w = w;
+ l.c = c;
+ l.out_h = h;
+ l.out_w = w;
+ l.out_c = c;
+ l.outputs = l.out_h * l.out_w * l.out_c;
+ l.inputs = l.outputs;
+ l.size = size;
+ l.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));
+ l.indexes = calloc(output_size, sizeof(int));
+ l.output = calloc(output_size, sizeof(float));
+ l.delta = calloc(output_size, sizeof(float));
#ifdef GPU
- layer->indexes_cl = cl_make_int_array(layer->indexes, output_size);
- layer->output_cl = cl_make_array(layer->output, output_size);
- layer->delta_cl = cl_make_array(layer->delta, output_size);
+ 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;
+ return l;
}
-void resize_maxpool_layer(maxpool_layer *layer, int h, int w, int c)
+void resize_maxpool_layer(maxpool_layer *l, int h, int w)
{
- 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));
+ l->h = h;
+ l->w = w;
+ int output_size = ((h-1)/l->stride+1) * ((w-1)/l->stride+1) * l->c * l->batch;
+ 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(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;
+ int b,i,j,k,m,n;
+ int w_offset = (-l.size-1)/2 + 1;
+ int h_offset = (-l.size-1)/2 + 1;
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
+ int h = (l.h-1)/l.stride + 1;
+ int w = (l.w-1)/l.stride + 1;
+ int c = l.c;
- for(b = 0; b < layer.batch; ++b){
+ 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.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];
}
}
-#ifdef GPU
-cl_kernel get_forward_kernel()
-{
- static int init = 0;
- static cl_kernel kernel;
- if(!init){
- kernel = get_kernel("src/maxpool_layer.cl", "forward", 0);
- init = 1;
- }
- return kernel;
-}
-
-void forward_maxpool_layer_gpu(maxpool_layer layer, cl_mem input)
-{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- cl_kernel kernel = get_forward_kernel();
- cl_command_queue queue = cl.queue;
-
- cl_uint i = 0;
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.h), (void*) &layer.h);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.w), (void*) &layer.w);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.c), (void*) &layer.c);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.stride), (void*) &layer.stride);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.size), (void*) &layer.size);
- cl.error = clSetKernelArg(kernel, i++, sizeof(input), (void*) &input);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.indexes_cl), (void*) &layer.indexes_cl);
- check_error(cl);
-
- const size_t global_size[] = {h*w*c*layer.batch};
-
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
- check_error(cl);
-}
-
-cl_kernel get_backward_kernel()
-{
- static int init = 0;
- static cl_kernel kernel;
- if(!init){
- kernel = get_kernel("src/maxpool_layer.cl", "backward", 0);
- init = 1;
- }
- return kernel;
-}
-
-void backward_maxpool_layer_gpu(maxpool_layer layer, cl_mem delta)
-{
- cl_kernel kernel = get_backward_kernel();
- cl_command_queue queue = cl.queue;
-
- cl_uint i = 0;
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.h), (void*) &layer.h);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.w), (void*) &layer.w);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.c), (void*) &layer.c);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.stride), (void*) &layer.stride);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.size), (void*) &layer.size);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.delta_cl), (void*) &layer.delta_cl);
- cl.error = clSetKernelArg(kernel, i++, sizeof(delta), (void*) &delta);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.indexes_cl), (void*) &layer.indexes_cl);
- check_error(cl);
-
- const size_t global_size[] = {layer.h*layer.w*layer.c*layer.batch};
-
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
- check_error(cl);
-}
-
-#endif
--
Gitblit v1.10.0