From 79fffcce3ce495bd415dc1284224c915d7194d4c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 11 Dec 2014 21:15:26 +0000
Subject: [PATCH] Better imagenet distributed training
---
src/convolutional_layer.c | 23 +++++++++++------------
1 files changed, 11 insertions(+), 12 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index bae06d3..18d00e6 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -59,17 +59,15 @@
layer->filters = calloc(c*n*size*size, sizeof(float));
layer->filter_updates = calloc(c*n*size*size, sizeof(float));
- layer->filter_momentum = calloc(c*n*size*size, sizeof(float));
layer->biases = calloc(n, sizeof(float));
layer->bias_updates = calloc(n, sizeof(float));
- layer->bias_momentum = calloc(n, sizeof(float));
- float scale = 1./(size*size*c);
- scale = .01;
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*2*(rand_uniform()-.5);
+ float scale = 1./sqrt(size*size*c);
+ //scale = .05;
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal();
for(i = 0; i < n; ++i){
//layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = .5;
+ layer->biases[i] = scale;
}
int out_h = convolutional_out_height(*layer);
int out_w = convolutional_out_width(*layer);
@@ -77,14 +75,13 @@
layer->col_image = calloc(out_h*out_w*size*size*c, sizeof(float));
layer->output = calloc(layer->batch*out_h * out_w * n, sizeof(float));
layer->delta = calloc(layer->batch*out_h * out_w * n, sizeof(float));
+
#ifdef GPU
layer->filters_cl = cl_make_array(layer->filters, c*n*size*size);
layer->filter_updates_cl = cl_make_array(layer->filter_updates, c*n*size*size);
- layer->filter_momentum_cl = cl_make_array(layer->filter_momentum, c*n*size*size);
layer->biases_cl = cl_make_array(layer->biases, n);
layer->bias_updates_cl = cl_make_array(layer->bias_updates, n);
- layer->bias_momentum_cl = cl_make_array(layer->bias_momentum, n);
layer->col_image_cl = cl_make_array(layer->col_image, out_h*out_w*size*size*c);
layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n);
@@ -207,7 +204,7 @@
axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1);
- scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1);
+ axpy_cpu(size, -layer.decay, layer.filters, 1, layer.filter_updates, 1);
axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
scal_cpu(size, layer.momentum, layer.filter_updates, 1);
}
@@ -283,7 +280,6 @@
{
int size = convolutional_out_height(layer) * convolutional_out_width(layer);
- cl_setup();
cl_kernel kernel = get_convolutional_learn_bias_kernel();
cl_command_queue queue = cl.queue;
@@ -318,7 +314,6 @@
int out_w = convolutional_out_width(layer);
int size = out_h*out_w;
- cl_setup();
cl_kernel kernel = get_convolutional_bias_kernel();
cl_command_queue queue = cl.queue;
@@ -394,12 +389,16 @@
{
cl_read_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
cl_read_array(layer.biases_cl, layer.biases, layer.n);
+ cl_read_array(layer.filter_updates_cl, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
+ cl_read_array(layer.bias_updates_cl, layer.bias_updates, layer.n);
}
void push_convolutional_layer(convolutional_layer layer)
{
cl_write_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
cl_write_array(layer.biases_cl, layer.biases, layer.n);
+ cl_write_array(layer.filter_updates_cl, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
+ cl_write_array(layer.bias_updates_cl, layer.bias_updates, layer.n);
}
void update_convolutional_layer_gpu(convolutional_layer layer)
@@ -408,7 +407,7 @@
axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
scal_ongpu(layer.n,layer.momentum, layer.bias_updates_cl, 1);
- scal_ongpu(size, 1.-layer.learning_rate*layer.decay, layer.filters_cl, 1);
+ axpy_ongpu(size, -layer.decay, layer.filters_cl, 1, layer.filter_updates_cl, 1);
axpy_ongpu(size, layer.learning_rate, layer.filter_updates_cl, 1, layer.filters_cl, 1);
scal_ongpu(size, layer.momentum, layer.filter_updates_cl, 1);
pull_convolutional_layer(layer);
--
Gitblit v1.10.0