From aa5996d58e68edfbefe51061856aecd549dd09c4 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 13 Jan 2015 01:27:08 +0000
Subject: [PATCH] Faster
---
src/connected_layer.c | 44 ++++++++++++++++++++++++++++++++++++++------
1 files changed, 38 insertions(+), 6 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index 85bf5c8..e29df77 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -24,17 +24,23 @@
layer->delta = calloc(batch*outputs, sizeof(float*));
layer->weight_updates = calloc(inputs*outputs, sizeof(float));
+ layer->bias_updates = calloc(outputs, sizeof(float));
+
+ layer->weight_prev = calloc(inputs*outputs, sizeof(float));
+ layer->bias_prev = calloc(outputs, sizeof(float));
+
layer->weights = calloc(inputs*outputs, sizeof(float));
- float scale = 1./inputs;
- scale = .01;
+ layer->biases = calloc(outputs, sizeof(float));
+
+
+ float scale = 1./sqrt(inputs);
+ //scale = .01;
for(i = 0; i < inputs*outputs; ++i){
layer->weights[i] = scale*rand_normal();
}
- layer->bias_updates = calloc(outputs, sizeof(float));
- layer->biases = calloc(outputs, sizeof(float));
for(i = 0; i < outputs; ++i){
- layer->biases[i] = .01;
+ layer->biases[i] = scale;
}
#ifdef GPU
@@ -52,6 +58,32 @@
return layer;
}
+void secret_update_connected_layer(connected_layer *layer)
+{
+ int n = layer->outputs*layer->inputs;
+ float dot = dot_cpu(n, layer->weight_updates, 1, layer->weight_prev, 1);
+ float mag = sqrt(dot_cpu(n, layer->weight_updates, 1, layer->weight_updates, 1))
+ * sqrt(dot_cpu(n, layer->weight_prev, 1, layer->weight_prev, 1));
+ float cos = dot/mag;
+ if(cos > .3) layer->learning_rate *= 1.1;
+ else if (cos < -.3) layer-> learning_rate /= 1.1;
+
+ scal_cpu(n, layer->momentum, layer->weight_prev, 1);
+ axpy_cpu(n, 1, layer->weight_updates, 1, layer->weight_prev, 1);
+ scal_cpu(n, 0, layer->weight_updates, 1);
+
+ scal_cpu(layer->outputs, layer->momentum, layer->bias_prev, 1);
+ axpy_cpu(layer->outputs, 1, layer->bias_updates, 1, layer->bias_prev, 1);
+ scal_cpu(layer->outputs, 0, layer->bias_updates, 1);
+
+ //printf("rate: %f\n", layer->learning_rate);
+
+ axpy_cpu(layer->outputs, layer->learning_rate, layer->bias_prev, 1, layer->biases, 1);
+
+ axpy_cpu(layer->inputs*layer->outputs, -layer->decay, layer->weights, 1, layer->weight_prev, 1);
+ axpy_cpu(layer->inputs*layer->outputs, layer->learning_rate, layer->weight_prev, 1, layer->weights, 1);
+}
+
void update_connected_layer(connected_layer layer)
{
axpy_cpu(layer.outputs, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
@@ -130,7 +162,7 @@
axpy_ongpu(layer.inputs*layer.outputs, -layer.decay, layer.weights_cl, 1, layer.weight_updates_cl, 1);
axpy_ongpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates_cl, 1, layer.weights_cl, 1);
scal_ongpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates_cl, 1);
- pull_connected_layer(layer);
+ //pull_connected_layer(layer);
}
void forward_connected_layer_gpu(connected_layer layer, cl_mem input)
--
Gitblit v1.10.0