From 76e258520edb50e8bb897ba15aa9467579e70a6a Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Wed, 20 Jun 2018 10:28:25 +0000
Subject: [PATCH] Minor fix
---
src/softmax_layer.c | 64 +++++++++++++++++++++++--------
1 files changed, 47 insertions(+), 17 deletions(-)
diff --git a/src/softmax_layer.c b/src/softmax_layer.c
index 20bc07f..27f73fd 100644
--- a/src/softmax_layer.c
+++ b/src/softmax_layer.c
@@ -10,7 +10,7 @@
softmax_layer make_softmax_layer(int batch, int inputs, int groups)
{
assert(inputs%groups == 0);
- fprintf(stderr, "Softmax Layer: %d inputs\n", inputs);
+ fprintf(stderr, "softmax %4d\n", inputs);
softmax_layer l = {0};
l.type = SOFTMAX;
l.batch = batch;
@@ -32,21 +32,17 @@
return l;
}
-void softmax_array(float *input, int n, float temp, float *output)
+void softmax_tree(float *input, int batch, int inputs, float temp, tree *hierarchy, float *output)
{
- int 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 < n; ++i){
- sum += exp(input[i]/temp-largest/temp);
- }
- if(sum) sum = largest/temp+log(sum);
- else sum = largest-100;
- for(i = 0; i < n; ++i){
- output[i] = exp(input[i]/temp-sum);
+ int b;
+ for(b = 0; b < batch; ++b){
+ int i;
+ int count = 0;
+ for(i = 0; i < hierarchy->groups; ++i){
+ int group_size = hierarchy->group_size[i];
+ softmax(input+b*inputs + count, group_size, temp, output+b*inputs + count, 1);
+ count += group_size;
+ }
}
}
@@ -55,8 +51,12 @@
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);
+ if(l.softmax_tree){
+ softmax_tree(state.input, batch, inputs, l.temperature, l.softmax_tree, l.output);
+ } else {
+ for(b = 0; b < batch; ++b){
+ softmax(state.input+b*inputs, inputs, l.temperature, l.output+b*inputs, 1);
+ }
}
}
@@ -68,3 +68,33 @@
}
}
+#ifdef GPU
+
+void pull_softmax_layer_output(const softmax_layer layer)
+{
+ cuda_pull_array(layer.output_gpu, layer.output, layer.inputs*layer.batch);
+}
+
+void forward_softmax_layer_gpu(const softmax_layer l, network_state state)
+{
+ int inputs = l.inputs / l.groups;
+ int batch = l.batch * l.groups;
+ if(l.softmax_tree){
+ int i;
+ int count = 0;
+ for (i = 0; i < l.softmax_tree->groups; ++i) {
+ int group_size = l.softmax_tree->group_size[i];
+ softmax_gpu(state.input+count, group_size, inputs, batch, l.temperature, l.output_gpu + count);
+ count += group_size;
+ }
+ } else {
+ softmax_gpu(state.input, inputs, inputs, batch, l.temperature, l.output_gpu);
+ }
+}
+
+void backward_softmax_layer_gpu(const softmax_layer layer, network_state state)
+{
+ axpy_ongpu(layer.batch*layer.inputs, 1, layer.delta_gpu, 1, state.delta, 1);
+}
+
+#endif
--
Gitblit v1.10.0