From 0f1a31648c5292fa49b35eac90a2ee676d6c13e6 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sat, 31 Jan 2015 06:05:23 +0000
Subject: [PATCH] idk, probably something changed
---
src/connected_layer.c | 196 ++++++++++++++++++++++++++++++++++++++----------
1 files changed, 154 insertions(+), 42 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index 72cb3fb..514dff0 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,64 +1,105 @@
#include "connected_layer.h"
#include "utils.h"
-#include "mini_blas.h"
+#include "cuda.h"
+#include "blas.h"
+#include "gemm.h"
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
-connected_layer *make_connected_layer(int batch, int inputs, int outputs, float dropout, ACTIVATION activation)
+connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay)
{
- fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
int i;
connected_layer *layer = calloc(1, sizeof(connected_layer));
+
+ layer->learning_rate = learning_rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
+
layer->inputs = inputs;
layer->outputs = outputs;
layer->batch=batch;
- layer->dropout = dropout;
layer->output = calloc(batch*outputs, sizeof(float*));
layer->delta = calloc(batch*outputs, sizeof(float*));
layer->weight_updates = calloc(inputs*outputs, sizeof(float));
- layer->weight_adapt = calloc(inputs*outputs, sizeof(float));
- layer->weight_momentum = calloc(inputs*outputs, sizeof(float));
- layer->weights = calloc(inputs*outputs, sizeof(float));
- float scale = 1./inputs;
- for(i = 0; i < inputs*outputs; ++i)
- layer->weights[i] = scale*(rand_uniform());
-
layer->bias_updates = calloc(outputs, sizeof(float));
- layer->bias_adapt = calloc(outputs, sizeof(float));
- layer->bias_momentum = calloc(outputs, sizeof(float));
- layer->biases = calloc(outputs, sizeof(float));
- for(i = 0; i < outputs; ++i)
- //layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = 1;
+ layer->weight_prev = calloc(inputs*outputs, sizeof(float));
+ layer->bias_prev = calloc(outputs, sizeof(float));
+
+ layer->weights = calloc(inputs*outputs, sizeof(float));
+ 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();
+ }
+
+ for(i = 0; i < outputs; ++i){
+ layer->biases[i] = scale;
+ }
+
+#ifdef GPU
+ layer->weights_gpu = cuda_make_array(layer->weights, inputs*outputs);
+ layer->biases_gpu = cuda_make_array(layer->biases, outputs);
+
+ layer->weight_updates_gpu = cuda_make_array(layer->weight_updates, inputs*outputs);
+ layer->bias_updates_gpu = cuda_make_array(layer->bias_updates, outputs);
+
+ layer->output_gpu = cuda_make_array(layer->output, outputs*batch);
+ layer->delta_gpu = cuda_make_array(layer->delta, outputs*batch);
+#endif
layer->activation = activation;
+ fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
return layer;
}
-void update_connected_layer(connected_layer layer, float step, float momentum, float decay)
+void secret_update_connected_layer(connected_layer *layer)
{
- int i;
- for(i = 0; i < layer.outputs; ++i){
- layer.bias_momentum[i] = step*(layer.bias_updates[i]) + momentum*layer.bias_momentum[i];
- layer.biases[i] += layer.bias_momentum[i];
- }
- for(i = 0; i < layer.outputs*layer.inputs; ++i){
- layer.weight_momentum[i] = step*(layer.weight_updates[i] - decay*layer.weights[i]) + momentum*layer.weight_momentum[i];
- layer.weights[i] += layer.weight_momentum[i];
- }
- memset(layer.bias_updates, 0, layer.outputs*sizeof(float));
- memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(float));
+ 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);
+
+ 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 forward_connected_layer(connected_layer layer, float *input, int train)
+void update_connected_layer(connected_layer layer)
{
- if(!train) layer.dropout = 0;
- memcpy(layer.output, layer.biases, layer.outputs*sizeof(float));
+ axpy_cpu(layer.outputs, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.outputs, layer.momentum, layer.bias_updates, 1);
+
+ axpy_cpu(layer.inputs*layer.outputs, -layer.decay, layer.weights, 1, layer.weight_updates, 1);
+ axpy_cpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates, 1, layer.weights, 1);
+ scal_cpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates, 1);
+}
+
+void forward_connected_layer(connected_layer layer, float *input)
+{
+ int i;
+ for(i = 0; i < layer.batch; ++i){
+ copy_cpu(layer.outputs, layer.biases, 1, layer.output + i*layer.outputs, 1);
+ }
int m = layer.batch;
int k = layer.inputs;
int n = layer.outputs;
@@ -66,15 +107,15 @@
float *b = layer.weights;
float *c = layer.output;
gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
- activate_array(layer.output, layer.outputs*layer.batch, layer.activation, layer.dropout);
+ activate_array(layer.output, layer.outputs*layer.batch, layer.activation);
}
void backward_connected_layer(connected_layer layer, float *input, float *delta)
{
int i;
- for(i = 0; i < layer.outputs*layer.batch; ++i){
- layer.delta[i] *= gradient(layer.output[i], layer.activation);
- layer.bias_updates[i%layer.batch] += layer.delta[i];
+ gradient_array(layer.output, layer.outputs*layer.batch, layer.activation, layer.delta);
+ for(i = 0; i < layer.batch; ++i){
+ axpy_cpu(layer.outputs, 1, layer.delta + i*layer.outputs, 1, layer.bias_updates, 1);
}
int m = layer.inputs;
int k = layer.batch;
@@ -82,16 +123,87 @@
float *a = input;
float *b = layer.delta;
float *c = layer.weight_updates;
- gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
- m = layer.inputs;
+ m = layer.batch;
k = layer.outputs;
- n = layer.batch;
+ n = layer.inputs;
- a = layer.weights;
- b = layer.delta;
+ a = layer.delta;
+ b = layer.weights;
c = delta;
- if(c) gemm(0,0,m,n,k,1,a,k,b,n,0,c,n);
+ if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
}
+#ifdef GPU
+
+void pull_connected_layer(connected_layer layer)
+{
+ cuda_pull_array(layer.weights_gpu, layer.weights, layer.inputs*layer.outputs);
+ cuda_pull_array(layer.biases_gpu, layer.biases, layer.outputs);
+ cuda_pull_array(layer.weight_updates_gpu, layer.weight_updates, layer.inputs*layer.outputs);
+ cuda_pull_array(layer.bias_updates_gpu, layer.bias_updates, layer.outputs);
+}
+
+void push_connected_layer(connected_layer layer)
+{
+ cuda_push_array(layer.weights_gpu, layer.weights, layer.inputs*layer.outputs);
+ cuda_push_array(layer.biases_gpu, layer.biases, layer.outputs);
+ cuda_push_array(layer.weight_updates_gpu, layer.weight_updates, layer.inputs*layer.outputs);
+ cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.outputs);
+}
+
+void update_connected_layer_gpu(connected_layer layer)
+{
+ axpy_ongpu(layer.outputs, layer.learning_rate, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
+ scal_ongpu(layer.outputs, layer.momentum, layer.bias_updates_gpu, 1);
+
+ axpy_ongpu(layer.inputs*layer.outputs, -layer.decay, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
+ axpy_ongpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates_gpu, 1, layer.weights_gpu, 1);
+ scal_ongpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates_gpu, 1);
+ //pull_connected_layer(layer);
+}
+
+void forward_connected_layer_gpu(connected_layer layer, float * input)
+{
+ int i;
+ for(i = 0; i < layer.batch; ++i){
+ copy_ongpu_offset(layer.outputs, layer.biases_gpu, 0, 1, layer.output_gpu, i*layer.outputs, 1);
+ }
+ int m = layer.batch;
+ int k = layer.inputs;
+ int n = layer.outputs;
+ float * a = input;
+ float * b = layer.weights_gpu;
+ float * c = layer.output_gpu;
+ gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ activate_array_ongpu(layer.output_gpu, layer.outputs*layer.batch, layer.activation);
+}
+
+void backward_connected_layer_gpu(connected_layer layer, float * input, float * delta)
+{
+ int i;
+ gradient_array_ongpu(layer.output_gpu, layer.outputs*layer.batch, layer.activation, layer.delta_gpu);
+ for(i = 0; i < layer.batch; ++i){
+ axpy_ongpu_offset(layer.outputs, 1, layer.delta_gpu, i*layer.outputs, 1, layer.bias_updates_gpu, 0, 1);
+ }
+ int m = layer.inputs;
+ int k = layer.batch;
+ int n = layer.outputs;
+ float * a = input;
+ float * b = layer.delta_gpu;
+ float * c = layer.weight_updates_gpu;
+ gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);
+
+ m = layer.batch;
+ k = layer.outputs;
+ n = layer.inputs;
+
+ a = layer.delta_gpu;
+ b = layer.weights_gpu;
+ c = delta;
+
+ if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,0,c,n);
+}
+#endif
--
Gitblit v1.10.0