From f047cfff99e00e28c02eb59b6d32386c122f9af6 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 08 Mar 2015 18:31:12 +0000
Subject: [PATCH] renamed sigmoid to logistic
---
src/softmax_layer.c | 79 ++++++++++++++-------------------------
1 files changed, 29 insertions(+), 50 deletions(-)
diff --git a/src/softmax_layer.c b/src/softmax_layer.c
index b6e9fe9..a200ae5 100644
--- a/src/softmax_layer.c
+++ b/src/softmax_layer.c
@@ -1,80 +1,59 @@
#include "softmax_layer.h"
-#include "mini_blas.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 batch, int inputs)
+softmax_layer *make_softmax_layer(int batch, int groups, int inputs)
{
+ 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*batch, sizeof(float));
layer->delta = calloc(inputs*batch, sizeof(float));
- layer->jacobian = calloc(inputs*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;
}
-/* UNSTABLE!
-void forward_softmax_layer(const softmax_layer layer, float *input)
+void softmax_array(float *input, int n, float *output)
{
int i;
float sum = 0;
- for(i = 0; i < layer.inputs; ++i){
- sum += exp(input[i]);
+ 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 forward_softmax_layer(const softmax_layer layer, float *input)
{
- int i,b;
- for(b = 0; b < layer.batch; ++b){
- float sum = 0;
- float largest = 0;
- for(i = 0; i < layer.inputs; ++i){
- if(input[i+b*layer.inputs] > largest) largest = input[i+b*layer.inputs];
- }
- for(i = 0; i < layer.inputs; ++i){
- sum += exp(input[i+b*layer.inputs]-largest);
- //printf("%f, ", input[i]);
- }
- //printf("\n");
- if(sum) sum = largest+log(sum);
- else sum = largest-100;
- for(i = 0; i < layer.inputs; ++i){
- layer.output[i+b*layer.inputs] = exp(input[i+b*layer.inputs]-sum);
- }
+ int b;
+ int inputs = layer.inputs / layer.groups;
+ int batch = layer.batch * layer.groups;
+ for(b = 0; b < batch; ++b){
+ softmax_array(input+b*inputs, inputs, layer.output+b*inputs);
}
}
-void backward_softmax_layer(const softmax_layer layer, float *input, float *delta)
+void backward_softmax_layer(const softmax_layer layer, float *delta)
{
-/*
- int i,j,b;
- for(b = 0; b < layer.batch; ++b){
- for(i = 0; i < layer.inputs; ++i){
- for(j = 0; j < layer.inputs; ++j){
- int d = (i==j);
- layer.jacobian[b*layer.inputs*layer.inputs + i*layer.inputs + j] =
- layer.output[b*layer.inputs + i] * (d - layer.output[b*layer.inputs + j]);
- }
- }
- }
- for(b = 0; b < layer.batch; ++b){
- int M = layer.inputs;
- int N = 1;
- int K = layer.inputs;
- float *A = layer.jacobian + b*layer.inputs*layer.inputs;
- float *B = layer.delta + b*layer.inputs;
- float *C = delta + b*layer.inputs;
- gemm(0,0,M,N,K,1,A,K,B,N,0,C,N);
- }
- */
-
int i;
for(i = 0; i < layer.inputs*layer.batch; ++i){
delta[i] = layer.delta[i];
--
Gitblit v1.10.0