From 6553b3f0e3e55fc30a99c7d4b5798aa86d18a114 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 30 Mar 2015 02:31:47 +0000
Subject: [PATCH] no comment
---
src/softmax_layer.c | 51 +++++++++++++++++++++++++++++++++++++++------------
1 files changed, 39 insertions(+), 12 deletions(-)
diff --git a/src/softmax_layer.c b/src/softmax_layer.c
index b213e5b..e344d16 100644
--- a/src/softmax_layer.c
+++ b/src/softmax_layer.c
@@ -1,35 +1,62 @@
#include "softmax_layer.h"
+#include "blas.h"
+#include "cuda.h"
+#include <float.h>
#include <math.h>
#include <stdlib.h>
#include <stdio.h>
+#include <assert.h>
-softmax_layer *make_softmax_layer(int inputs)
+softmax_layer *make_softmax_layer(int batch, int inputs, int groups)
{
+ assert(inputs%groups == 0);
fprintf(stderr, "Softmax Layer: %d inputs\n", inputs);
softmax_layer *layer = calloc(1, sizeof(softmax_layer));
+ layer->batch = batch;
+ layer->groups = groups;
layer->inputs = inputs;
- layer->output = calloc(inputs, sizeof(double));
- layer->delta = calloc(inputs, sizeof(double));
+ layer->output = calloc(inputs*batch, sizeof(float));
+ layer->delta = calloc(inputs*batch, sizeof(float));
+ #ifdef GPU
+ layer->output_gpu = cuda_make_array(layer->output, inputs*batch);
+ layer->delta_gpu = cuda_make_array(layer->delta, inputs*batch);
+ #endif
return layer;
}
-void forward_softmax_layer(const softmax_layer layer, double *input)
+void softmax_array(float *input, int n, float *output)
{
int i;
- double sum = 0;
- for(i = 0; i < layer.inputs; ++i){
- sum += exp(input[i]);
+ float sum = 0;
+ float largest = -FLT_MAX;
+ for(i = 0; i < n; ++i){
+ if(input[i] > largest) largest = input[i];
}
- for(i = 0; i < layer.inputs; ++i){
- layer.output[i] = exp(input[i])/sum;
+ for(i = 0; i < n; ++i){
+ sum += exp(input[i]-largest);
+ }
+ if(sum) sum = largest+log(sum);
+ else sum = largest-100;
+ for(i = 0; i < n; ++i){
+ output[i] = exp(input[i]-sum);
}
}
-void backward_softmax_layer(const softmax_layer layer, double *input, double *delta)
+void forward_softmax_layer(const softmax_layer layer, network_state state)
+{
+ int b;
+ int inputs = layer.inputs / layer.groups;
+ int batch = layer.batch * layer.groups;
+ for(b = 0; b < batch; ++b){
+ softmax_array(state.input+b*inputs, inputs, layer.output+b*inputs);
+ }
+}
+
+void backward_softmax_layer(const softmax_layer layer, network_state state)
{
int i;
- for(i = 0; i < layer.inputs; ++i){
- delta[i] = layer.delta[i];
+ for(i = 0; i < layer.inputs*layer.batch; ++i){
+ state.delta[i] = layer.delta[i];
}
}
--
Gitblit v1.10.0