From 729ce43e6ec45cfdb58e06e227428a0f81c5de0f Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 10 Jun 2016 00:20:31 +0000
Subject: [PATCH] stuff

---
 src/softmax_layer.c |   99 ++++++++++++++++++++-----------------------------
 1 files changed, 40 insertions(+), 59 deletions(-)

diff --git a/src/softmax_layer.c b/src/softmax_layer.c
index b6e9fe9..e189701 100644
--- a/src/softmax_layer.c
+++ b/src/softmax_layer.c
@@ -1,83 +1,64 @@
 #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 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->inputs = inputs;
-    layer->output = calloc(inputs*batch, sizeof(float));
-    layer->delta = calloc(inputs*batch, sizeof(float));
-    layer->jacobian = calloc(inputs*inputs*batch, sizeof(float));
-    return layer;
+    softmax_layer l = {0};
+    l.type = SOFTMAX;
+    l.batch = batch;
+    l.groups = groups;
+    l.inputs = inputs;
+    l.outputs = inputs;
+    l.output = calloc(inputs*batch, sizeof(float));
+    l.delta = calloc(inputs*batch, sizeof(float));
+    #ifdef GPU
+    l.output_gpu = cuda_make_array(l.output, inputs*batch); 
+    l.delta_gpu = cuda_make_array(l.delta, inputs*batch); 
+    #endif
+    return l;
 }
 
-/* UNSTABLE!
-void forward_softmax_layer(const softmax_layer layer, float *input)
+void softmax_array(float *input, int n, float temp, 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]/temp-largest/temp);
     }
-}
-*/
-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);
-        }
+    if(sum) sum = largest/temp+log(sum);
+    else sum = largest-100;
+    for(i = 0; i < n; ++i){
+        output[i] = exp(input[i]/temp-sum);
     }
 }
 
-void backward_softmax_layer(const softmax_layer layer, float *input, float *delta)
+void forward_softmax_layer(const softmax_layer l, network_state state)
 {
-/*
-    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]);
-            }
-        }
+    int b;
+    int inputs = l.inputs / l.groups;
+    int batch = l.batch * l.groups;
+    for(b = 0; b < batch; ++b){
+        softmax_array(state.input+b*inputs, inputs, l.temperature, l.output+b*inputs);
     }
-    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);
-    }
-    */
+}
 
+void backward_softmax_layer(const softmax_layer l, network_state state)
+{
     int i;
-    for(i = 0; i < layer.inputs*layer.batch; ++i){
-        delta[i] = layer.delta[i];
+    for(i = 0; i < l.inputs*l.batch; ++i){
+        state.delta[i] += l.delta[i];
     }
 }
 

--
Gitblit v1.10.0