From 0b2be9d76a369e28fd1fc384b3952621921c9ae5 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 12 Jul 2015 05:52:43 +0000
Subject: [PATCH] Small version of YOLO, ~1GB RAM
---
src/connected_layer.c | 213 ++++++++++++++++++++++++++++++++++++----------------
1 files changed, 147 insertions(+), 66 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index d77a10c..55d84ca 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,97 +1,178 @@
#include "connected_layer.h"
+#include "utils.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 inputs, int outputs, ACTIVATION activator)
+connected_layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation)
{
- printf("Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
int i;
- connected_layer *layer = calloc(1, sizeof(connected_layer));
- layer->inputs = inputs;
- layer->outputs = outputs;
+ connected_layer l = {0};
+ l.type = CONNECTED;
- layer->output = calloc(outputs, sizeof(double*));
- layer->delta = calloc(outputs, sizeof(double*));
+ l.inputs = inputs;
+ l.outputs = outputs;
+ l.batch=batch;
- layer->weight_updates = calloc(inputs*outputs, sizeof(double));
- layer->weight_momentum = calloc(inputs*outputs, sizeof(double));
- layer->weights = calloc(inputs*outputs, sizeof(double));
- for(i = 0; i < inputs*outputs; ++i)
- layer->weights[i] = .01*(.5 - (double)rand()/RAND_MAX);
+ l.output = calloc(batch*outputs, sizeof(float*));
+ l.delta = calloc(batch*outputs, sizeof(float*));
- layer->bias_updates = calloc(outputs, sizeof(double));
- layer->bias_momentum = calloc(outputs, sizeof(double));
- layer->biases = calloc(outputs, sizeof(double));
- for(i = 0; i < outputs; ++i)
- layer->biases[i] = 1;
+ l.weight_updates = calloc(inputs*outputs, sizeof(float));
+ l.bias_updates = calloc(outputs, sizeof(float));
- if(activator == SIGMOID){
- layer->activation = sigmoid_activation;
- layer->gradient = sigmoid_gradient;
- }else if(activator == RELU){
- layer->activation = relu_activation;
- layer->gradient = relu_gradient;
- }else if(activator == IDENTITY){
- layer->activation = identity_activation;
- layer->gradient = identity_gradient;
+ l.weights = calloc(inputs*outputs, sizeof(float));
+ l.biases = calloc(outputs, sizeof(float));
+
+
+ //float scale = 1./sqrt(inputs);
+ float scale = sqrt(2./inputs);
+ for(i = 0; i < inputs*outputs; ++i){
+ l.weights[i] = 2*scale*rand_uniform() - scale;
}
- return layer;
+ for(i = 0; i < outputs; ++i){
+ l.biases[i] = scale;
+ }
+
+#ifdef GPU
+ l.weights_gpu = cuda_make_array(l.weights, inputs*outputs);
+ l.biases_gpu = cuda_make_array(l.biases, outputs);
+
+ l.weight_updates_gpu = cuda_make_array(l.weight_updates, inputs*outputs);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, outputs);
+
+ l.output_gpu = cuda_make_array(l.output, outputs*batch);
+ l.delta_gpu = cuda_make_array(l.delta, outputs*batch);
+#endif
+ l.activation = activation;
+ fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
+ return l;
}
-void forward_connected_layer(connected_layer layer, double *input)
+void update_connected_layer(connected_layer l, int batch, float learning_rate, float momentum, float decay)
{
- int i, j;
- for(i = 0; i < layer.outputs; ++i){
- layer.output[i] = layer.biases[i];
- for(j = 0; j < layer.inputs; ++j){
- layer.output[i] += input[j]*layer.weights[i*layer.inputs + j];
- }
- layer.output[i] = layer.activation(layer.output[i]);
- }
+ axpy_cpu(l.outputs, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
+ scal_cpu(l.outputs, momentum, l.bias_updates, 1);
+
+ axpy_cpu(l.inputs*l.outputs, -decay*batch, l.weights, 1, l.weight_updates, 1);
+ axpy_cpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
+ scal_cpu(l.inputs*l.outputs, momentum, l.weight_updates, 1);
}
-void learn_connected_layer(connected_layer layer, double *input)
+void forward_connected_layer(connected_layer l, network_state state)
{
- int i, j;
- for(i = 0; i < layer.outputs; ++i){
- layer.bias_updates[i] += layer.delta[i];
- for(j = 0; j < layer.inputs; ++j){
- layer.weight_updates[i*layer.inputs + j] += layer.delta[i]*input[j];
- }
+ int i;
+ for(i = 0; i < l.batch; ++i){
+ copy_cpu(l.outputs, l.biases, 1, l.output + i*l.outputs, 1);
}
+ int m = l.batch;
+ int k = l.inputs;
+ int n = l.outputs;
+ float *a = state.input;
+ float *b = l.weights;
+ float *c = l.output;
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ activate_array(l.output, l.outputs*l.batch, l.activation);
}
-void update_connected_layer(connected_layer layer, double step, double momentum, double decay)
+void backward_connected_layer(connected_layer l, network_state state)
{
- int i,j;
- for(i = 0; i < layer.outputs; ++i){
- layer.bias_momentum[i] = step*(layer.bias_updates[i] - decay*layer.biases[i]) + momentum*layer.bias_momentum[i];
- layer.biases[i] += layer.bias_momentum[i];
- for(j = 0; j < layer.inputs; ++j){
- int index = i*layer.inputs+j;
- layer.weight_momentum[index] = step*(layer.weight_updates[index] - decay*layer.weights[index]) + momentum*layer.weight_momentum[index];
- layer.weights[index] += layer.weight_momentum[index];
- }
+ int i;
+ gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);
+ for(i = 0; i < l.batch; ++i){
+ axpy_cpu(l.outputs, 1, l.delta + i*l.outputs, 1, l.bias_updates, 1);
}
- memset(layer.bias_updates, 0, layer.outputs*sizeof(double));
- memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(double));
+ int m = l.inputs;
+ int k = l.batch;
+ int n = l.outputs;
+ float *a = state.input;
+ float *b = l.delta;
+ float *c = l.weight_updates;
+ gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
+
+ m = l.batch;
+ k = l.outputs;
+ n = l.inputs;
+
+ a = l.delta;
+ b = l.weights;
+ c = state.delta;
+
+ if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
}
-void backward_connected_layer(connected_layer layer, double *input, double *delta)
+#ifdef GPU
+
+void pull_connected_layer(connected_layer l)
{
- int i, j;
-
- for(j = 0; j < layer.inputs; ++j){
- double grad = layer.gradient(input[j]);
- delta[j] = 0;
- for(i = 0; i < layer.outputs; ++i){
- delta[j] += layer.delta[i]*layer.weights[i*layer.inputs + j];
- }
- delta[j] *= grad;
- }
+ cuda_pull_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
+ cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
+ cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
+ cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
}
+void push_connected_layer(connected_layer l)
+{
+ cuda_push_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
+ cuda_push_array(l.biases_gpu, l.biases, l.outputs);
+ cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
+ cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+}
+
+void update_connected_layer_gpu(connected_layer l, int batch, float learning_rate, float momentum, float decay)
+{
+ axpy_ongpu(l.outputs, learning_rate/batch, l.bias_updates_gpu, 1, l.biases_gpu, 1);
+ scal_ongpu(l.outputs, momentum, l.bias_updates_gpu, 1);
+
+ axpy_ongpu(l.inputs*l.outputs, -decay*batch, l.weights_gpu, 1, l.weight_updates_gpu, 1);
+ axpy_ongpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates_gpu, 1, l.weights_gpu, 1);
+ scal_ongpu(l.inputs*l.outputs, momentum, l.weight_updates_gpu, 1);
+}
+
+void forward_connected_layer_gpu(connected_layer l, network_state state)
+{
+ int i;
+ for(i = 0; i < l.batch; ++i){
+ copy_ongpu_offset(l.outputs, l.biases_gpu, 0, 1, l.output_gpu, i*l.outputs, 1);
+ }
+ int m = l.batch;
+ int k = l.inputs;
+ int n = l.outputs;
+ float * a = state.input;
+ float * b = l.weights_gpu;
+ float * c = l.output_gpu;
+ gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
+}
+
+void backward_connected_layer_gpu(connected_layer l, network_state state)
+{
+ int i;
+ gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
+ for(i = 0; i < l.batch; ++i){
+ axpy_ongpu_offset(l.outputs, 1, l.delta_gpu, i*l.outputs, 1, l.bias_updates_gpu, 0, 1);
+ }
+ int m = l.inputs;
+ int k = l.batch;
+ int n = l.outputs;
+ float * a = state.input;
+ float * b = l.delta_gpu;
+ float * c = l.weight_updates_gpu;
+ gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);
+
+ m = l.batch;
+ k = l.outputs;
+ n = l.inputs;
+
+ a = l.delta_gpu;
+ b = l.weights_gpu;
+ c = state.delta;
+
+ if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,0,c,n);
+}
+#endif
--
Gitblit v1.10.0