From 09fd5c8c84eeae711f49d3a52d8bf4b65f43970b Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 15 Mar 2016 06:11:02 +0000
Subject: [PATCH] I hate deepmind
---
src/convolutional_layer.c | 112 +++++++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 95 insertions(+), 17 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index b9fd3c9..159951d 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -41,7 +41,65 @@
return float_to_image(w,h,c,l.delta);
}
-convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize)
+void backward_scale_cpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates)
+{
+ int i,b,f;
+ for(f = 0; f < n; ++f){
+ float sum = 0;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < size; ++i){
+ int index = i + size*(f + n*b);
+ sum += delta[index] * x_norm[index];
+ }
+ }
+ scale_updates[f] += sum;
+ }
+}
+
+void mean_delta_cpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
+{
+
+ int i,j,k;
+ for(i = 0; i < filters; ++i){
+ mean_delta[i] = 0;
+ for (j = 0; j < batch; ++j) {
+ for (k = 0; k < spatial; ++k) {
+ int index = j*filters*spatial + i*spatial + k;
+ mean_delta[i] += delta[index];
+ }
+ }
+ mean_delta[i] *= (-1./sqrt(variance[i] + .00001f));
+ }
+}
+void variance_delta_cpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta)
+{
+
+ int i,j,k;
+ for(i = 0; i < filters; ++i){
+ variance_delta[i] = 0;
+ for(j = 0; j < batch; ++j){
+ for(k = 0; k < spatial; ++k){
+ int index = j*filters*spatial + i*spatial + k;
+ variance_delta[i] += delta[index]*(x[index] - mean[i]);
+ }
+ }
+ variance_delta[i] *= -.5 * pow(variance[i] + .00001f, (float)(-3./2.));
+ }
+}
+void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
+{
+ int f, j, k;
+ for(j = 0; j < batch; ++j){
+ for(f = 0; f < filters; ++f){
+ for(k = 0; k < spatial; ++k){
+ int index = j*filters*spatial + f*spatial + k;
+ delta[index] = delta[index] * 1./(sqrt(variance[f]) + .00001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
+ }
+ }
+ }
+}
+
+convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize, int binary)
{
int i;
convolutional_layer l = {0};
@@ -51,6 +109,7 @@
l.w = w;
l.c = c;
l.n = n;
+ l.binary = binary;
l.batch = batch;
l.stride = stride;
l.size = size;
@@ -65,7 +124,7 @@
// float scale = 1./sqrt(size*size*c);
float scale = sqrt(2./(size*size*c));
- for(i = 0; i < c*n*size*size; ++i) l.filters[i] = 2*scale*rand_uniform() - scale;
+ for(i = 0; i < c*n*size*size; ++i) l.filters[i] = scale*rand_uniform(-1, 1);
int out_h = convolutional_out_height(l);
int out_w = convolutional_out_width(l);
l.out_h = out_h;
@@ -78,6 +137,10 @@
l.output = calloc(l.batch*out_h * out_w * n, sizeof(float));
l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float));
+ if(binary){
+ l.binary_filters = calloc(c*n*size*size, sizeof(float));
+ }
+
if(batch_normalize){
l.scales = calloc(n, sizeof(float));
l.scale_updates = calloc(n, sizeof(float));
@@ -86,9 +149,8 @@
}
l.mean = calloc(n, sizeof(float));
- l.spatial_mean = calloc(n*l.batch, sizeof(float));
-
l.variance = calloc(n, sizeof(float));
+
l.rolling_mean = calloc(n, sizeof(float));
l.rolling_variance = calloc(n, sizeof(float));
}
@@ -107,6 +169,10 @@
l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n);
l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ if(binary){
+ l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
+ }
+
if(batch_normalize){
l.mean_gpu = cuda_make_array(l.mean, n);
l.variance_gpu = cuda_make_array(l.variance, n);
@@ -114,12 +180,6 @@
l.rolling_mean_gpu = cuda_make_array(l.mean, n);
l.rolling_variance_gpu = cuda_make_array(l.variance, n);
- l.spatial_mean_gpu = cuda_make_array(l.spatial_mean, n*l.batch);
- l.spatial_variance_gpu = cuda_make_array(l.spatial_mean, n*l.batch);
-
- l.spatial_mean_delta_gpu = cuda_make_array(l.spatial_mean, n*l.batch);
- l.spatial_variance_delta_gpu = cuda_make_array(l.spatial_mean, n*l.batch);
-
l.mean_delta_gpu = cuda_make_array(l.mean, n);
l.variance_delta_gpu = cuda_make_array(l.variance, n);
@@ -148,7 +208,7 @@
void test_convolutional_layer()
{
- convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1);
+ convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 0);
l.batch_normalize = 1;
float data[] = {1,1,1,1,1,
1,1,1,1,1,
@@ -201,13 +261,25 @@
#endif
}
-void bias_output(float *output, float *biases, int batch, int n, int size)
+void add_bias(float *output, float *biases, int batch, int n, int size)
{
int i,j,b;
for(b = 0; b < batch; ++b){
for(i = 0; i < n; ++i){
for(j = 0; j < size; ++j){
- output[(b*n + i)*size + j] = biases[i];
+ output[(b*n + i)*size + j] += biases[i];
+ }
+ }
+ }
+}
+
+void scale_bias(float *output, float *scales, int batch, int n, int size)
+{
+ int i,j,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ for(j = 0; j < size; ++j){
+ output[(b*n + i)*size + j] *= scales[i];
}
}
}
@@ -229,7 +301,7 @@
int out_w = convolutional_out_width(l);
int i;
- bias_output(l.output, l.biases, l.batch, l.n, out_h*out_w);
+ fill_cpu(l.outputs*l.batch, 0, l.output, 1);
int m = l.n;
int k = l.size*l.size*l.c;
@@ -248,10 +320,16 @@
}
if(l.batch_normalize){
- mean_cpu(l.output, l.batch, l.n, l.out_h*l.out_w, l.mean);
- variance_cpu(l.output, l.mean, l.batch, l.n, l.out_h*l.out_w, l.variance);
- normalize_cpu(l.output, l.mean, l.variance, l.batch, l.n, l.out_h*l.out_w);
+ if(state.train){
+ mean_cpu(l.output, l.batch, l.n, l.out_h*l.out_w, l.mean);
+ variance_cpu(l.output, l.mean, l.batch, l.n, l.out_h*l.out_w, l.variance);
+ normalize_cpu(l.output, l.mean, l.variance, l.batch, l.n, l.out_h*l.out_w);
+ } else {
+ normalize_cpu(l.output, l.rolling_mean, l.rolling_variance, l.batch, l.n, l.out_h*l.out_w);
+ }
+ scale_bias(l.output, l.scales, l.batch, l.n, out_h*out_w);
}
+ add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
activate_array(l.output, m*n*l.batch, l.activation);
}
--
Gitblit v1.10.0