From 89354d0a0ce6fbb22ff262658045cdb8796ff6fd Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 04 May 2018 20:52:05 +0000
Subject: [PATCH] Fixed memory leaks. And fixes for Web-camera and IP-camera.
---
src/connected_layer.c | 109 +++++++++++++++++++++++++++++++++---------------------
1 files changed, 67 insertions(+), 42 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index df78e67..e6dc759 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,4 +1,5 @@
#include "connected_layer.h"
+#include "batchnorm_layer.h"
#include "utils.h"
#include "cuda.h"
#include "blas.h"
@@ -19,6 +20,12 @@
l.outputs = outputs;
l.batch=batch;
l.batch_normalize = batch_normalize;
+ l.h = 1;
+ l.w = 1;
+ l.c = inputs;
+ l.out_h = 1;
+ l.out_w = 1;
+ l.out_c = outputs;
l.output = calloc(batch*outputs, sizeof(float));
l.delta = calloc(batch*outputs, sizeof(float));
@@ -29,6 +36,9 @@
l.weights = calloc(outputs*inputs, sizeof(float));
l.biases = calloc(outputs, sizeof(float));
+ l.forward = forward_connected_layer;
+ l.backward = backward_connected_layer;
+ l.update = update_connected_layer;
//float scale = 1./sqrt(inputs);
float scale = sqrt(2./inputs);
@@ -37,7 +47,7 @@
}
for(i = 0; i < outputs; ++i){
- l.biases[i] = scale;
+ l.biases[i] = 0;
}
if(batch_normalize){
@@ -60,6 +70,10 @@
}
#ifdef GPU
+ l.forward_gpu = forward_connected_layer_gpu;
+ l.backward_gpu = backward_connected_layer_gpu;
+ l.update_gpu = update_connected_layer_gpu;
+
l.weights_gpu = cuda_make_array(l.weights, outputs*inputs);
l.biases_gpu = cuda_make_array(l.biases, outputs);
@@ -83,10 +97,16 @@
l.x_gpu = cuda_make_array(l.output, l.batch*outputs);
l.x_norm_gpu = cuda_make_array(l.output, l.batch*outputs);
+#ifdef CUDNN
+ cudnnCreateTensorDescriptor(&l.normTensorDesc);
+ cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+ cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
+ cudnnSetTensor4dDescriptor(l.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l.out_c, 1, 1);
+#endif
}
#endif
l.activation = activation;
- fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
+ fprintf(stderr, "connected %4d -> %4d\n", inputs, outputs);
return l;
}
@@ -176,6 +196,41 @@
if(c) gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
}
+
+void denormalize_connected_layer(layer l)
+{
+ int i, j;
+ for(i = 0; i < l.outputs; ++i){
+ float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .000001);
+ for(j = 0; j < l.inputs; ++j){
+ l.weights[i*l.inputs + j] *= scale;
+ }
+ l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
+ }
+}
+
+
+void statistics_connected_layer(layer l)
+{
+ if(l.batch_normalize){
+ printf("Scales ");
+ print_statistics(l.scales, l.outputs);
+ /*
+ printf("Rolling Mean ");
+ print_statistics(l.rolling_mean, l.outputs);
+ printf("Rolling Variance ");
+ print_statistics(l.rolling_variance, l.outputs);
+ */
+ }
+ printf("Biases ");
+ print_statistics(l.biases, l.outputs);
+ printf("Weights ");
+ print_statistics(l.weights, l.outputs);
+}
+
#ifdef GPU
void pull_connected_layer(connected_layer l)
@@ -223,11 +278,7 @@
{
int i;
fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
- /*
- 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;
@@ -235,53 +286,27 @@
float * b = l.weights_gpu;
float * c = l.output_gpu;
gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
- if(l.batch_normalize){
- if(state.train){
- fast_mean_gpu(l.output_gpu, l.batch, l.outputs, 1, l.mean_gpu);
- fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.outputs, 1, l.variance_gpu);
-
- scal_ongpu(l.outputs, .95, l.rolling_mean_gpu, 1);
- axpy_ongpu(l.outputs, .05, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
- scal_ongpu(l.outputs, .95, l.rolling_variance_gpu, 1);
- axpy_ongpu(l.outputs, .05, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
-
- copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
- normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.outputs, 1);
- copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
- } else {
- normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.outputs, 1);
- }
-
- scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.outputs, 1);
- }
- for(i = 0; i < l.batch; ++i){
- axpy_ongpu(l.outputs, 1, l.biases_gpu, 1, l.output_gpu + i*l.outputs, 1);
- }
+ if (l.batch_normalize) {
+ forward_batchnorm_layer_gpu(l, state);
+ }
+ else {
+ add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.outputs, 1);
+ }
+ //for(i = 0; i < l.batch; ++i) axpy_ongpu(l.outputs, 1, l.biases_gpu, 1, l.output_gpu + i*l.outputs, 1);
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
-
- /*
- cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
- float avg = mean_array(l.output, l.outputs*l.batch);
- printf("%f\n", avg);
- */
}
void backward_connected_layer_gpu(connected_layer l, network_state state)
{
int i;
+ constrain_ongpu(l.outputs*l.batch, 1, l.delta_gpu, 1);
gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
for(i = 0; i < l.batch; ++i){
axpy_ongpu(l.outputs, 1, l.delta_gpu + i*l.outputs, 1, l.bias_updates_gpu, 1);
}
if(l.batch_normalize){
- backward_scale_gpu(l.x_norm_gpu, l.delta_gpu, l.batch, l.outputs, 1, l.scale_updates_gpu);
-
- scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.outputs, 1);
-
- fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.outputs, 1, l.mean_delta_gpu);
- fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.outputs, 1, l.variance_delta_gpu);
- normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.outputs, 1, l.delta_gpu);
+ backward_batchnorm_layer_gpu(l, state);
}
int m = l.outputs;
--
Gitblit v1.10.0