From 352ae7e65b6a74bcd768aa88b866a44c713284c8 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 26 Oct 2016 15:35:44 +0000
Subject: [PATCH] ADAM
---
src/softmax_layer.c | 72 +++++++++++++++++++++++++----------
1 files changed, 51 insertions(+), 21 deletions(-)
diff --git a/src/softmax_layer.c b/src/softmax_layer.c
index ea22d05..31f3e03 100644
--- a/src/softmax_layer.c
+++ b/src/softmax_layer.c
@@ -19,38 +19,38 @@
l.outputs = inputs;
l.output = calloc(inputs*batch, sizeof(float));
l.delta = calloc(inputs*batch, sizeof(float));
+
+ l.forward = forward_softmax_layer;
+ l.backward = backward_softmax_layer;
#ifdef GPU
+ l.forward_gpu = forward_softmax_layer_gpu;
+ l.backward_gpu = backward_softmax_layer_gpu;
+
l.output_gpu = cuda_make_array(l.output, inputs*batch);
l.delta_gpu = cuda_make_array(l.delta, inputs*batch);
#endif
return l;
}
-void softmax_array(float *input, int n, 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]-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 l, network_state state)
{
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.output+b*inputs);
+ if(l.softmax_tree){
+ for(b = 0; b < batch; ++b){
+ int i;
+ int count = 0;
+ for(i = 0; i < l.softmax_tree->groups; ++i){
+ int group_size = l.softmax_tree->group_size[i];
+ softmax(state.input+b*inputs + count, group_size, l.temperature, l.output+b*inputs + count);
+ count += group_size;
+ }
+ }
+ } else {
+ for(b = 0; b < batch; ++b){
+ softmax(state.input+b*inputs, inputs, l.temperature, l.output+b*inputs);
+ }
}
}
@@ -58,7 +58,37 @@
{
int i;
for(i = 0; i < l.inputs*l.batch; ++i){
- state.delta[i] = l.delta[i];
+ state.delta[i] += l.delta[i];
}
}
+#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