From cd8d53df21f3ad2810add2a8cff766c745f55a17 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 09 May 2014 22:14:52 +0000
Subject: [PATCH] So there WAS this huge bug. Gone now
---
src/convolutional_layer.c | 271 +++++++++++++++++++++++++++++++++++++++++++-----------
1 files changed, 216 insertions(+), 55 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 7478158..5aa76ee 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,95 +1,256 @@
#include "convolutional_layer.h"
+#include "utils.h"
+#include "mini_blas.h"
+#include <stdio.h>
-double convolution_activation(double x)
+int convolutional_out_height(convolutional_layer layer)
{
- return x*(x>0);
+ return (layer.h-layer.size)/layer.stride + 1;
}
-double convolution_gradient(double x)
+int convolutional_out_width(convolutional_layer layer)
{
- return (x>=0);
+ return (layer.w-layer.size)/layer.stride + 1;
}
-convolutional_layer *make_convolutional_layer(int h, int w, int c, int n, int size, int stride)
+image get_convolutional_image(convolutional_layer layer)
+{
+ int h,w,c;
+ h = convolutional_out_height(layer);
+ w = convolutional_out_width(layer);
+ c = layer.n;
+ return float_to_image(h,w,c,layer.output);
+}
+
+image get_convolutional_delta(convolutional_layer layer)
+{
+ int h,w,c;
+ h = convolutional_out_height(layer);
+ w = convolutional_out_width(layer);
+ c = layer.n;
+ return float_to_image(h,w,c,layer.delta);
+}
+
+convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, ACTIVATION activation)
{
int i;
+ size = 2*(size/2)+1; //HA! And you thought you'd use an even sized filter...
convolutional_layer *layer = calloc(1, sizeof(convolutional_layer));
+ layer->h = h;
+ layer->w = w;
+ layer->c = c;
layer->n = n;
+ layer->batch = batch;
layer->stride = stride;
- layer->kernels = calloc(n, sizeof(image));
- layer->kernel_updates = calloc(n, sizeof(image));
+ layer->size = size;
+
+ 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);
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform());
for(i = 0; i < n; ++i){
- layer->kernels[i] = make_random_kernel(size, c);
- layer->kernel_updates[i] = make_random_kernel(size, c);
+ //layer->biases[i] = rand_normal()*scale + scale;
+ layer->biases[i] = .5;
}
- layer->output = make_image((h-1)/stride+1, (w-1)/stride+1, n);
- layer->upsampled = make_image(h,w,n);
+ int out_h = convolutional_out_height(*layer);
+ int out_w = convolutional_out_width(*layer);
+
+ layer->col_image = calloc(layer->batch*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
+ #endif
+ layer->activation = activation;
+
+ fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n);
+ srand(0);
+
return layer;
}
-void run_convolutional_layer(const image input, const convolutional_layer layer)
+void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c)
{
- int i;
- for(i = 0; i < layer.n; ++i){
- convolve(input, layer.kernels[i], layer.stride, i, layer.output);
- }
- for(i = 0; i < layer.output.h*layer.output.w*layer.output.c; ++i){
- layer.output.data[i] = convolution_activation(layer.output.data[i]);
- }
+ layer->h = h;
+ layer->w = w;
+ layer->c = c;
+ int out_h = convolutional_out_height(*layer);
+ int out_w = convolutional_out_width(*layer);
+
+ layer->col_image = realloc(layer->col_image,
+ layer->batch*out_h*out_w*layer->size*layer->size*layer->c*sizeof(float));
+ layer->output = realloc(layer->output,
+ layer->batch*out_h * out_w * layer->n*sizeof(float));
+ layer->delta = realloc(layer->delta,
+ layer->batch*out_h * out_w * layer->n*sizeof(float));
}
-void backpropagate_convolutional_layer(image input, convolutional_layer layer)
-{
- int i;
- zero_image(input);
- for(i = 0; i < layer.n; ++i){
- back_convolve(input, layer.kernels[i], layer.stride, i, layer.output);
- }
-}
-
-void backpropagate_convolutional_layer_convolve(image input, convolutional_layer layer)
+void bias_output(const convolutional_layer layer)
{
int i,j;
+ int out_h = convolutional_out_height(layer);
+ int out_w = convolutional_out_width(layer);
for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
+ for(j = 0; j < out_h*out_w; ++j){
+ layer.output[i*out_h*out_w + j] = layer.biases[i];
+ }
}
+}
- zero_image(input);
- upsample_image(layer.output, layer.stride, layer.upsampled);
- for(j = 0; j < input.c; ++j){
+void forward_convolutional_layer(const convolutional_layer layer, float *in)
+{
+ int out_h = convolutional_out_height(layer);
+ int out_w = convolutional_out_width(layer);
+
+ int m = layer.n;
+ int k = layer.size*layer.size*layer.c;
+ int n = out_h*out_w*layer.batch;
+
+ float *a = layer.filters;
+ float *b = layer.col_image;
+ float *c = layer.output;
+ im2col_cpu(in,layer.batch, layer.c, layer.h, layer.w,
+ layer.size, layer.stride, b);
+ bias_output(layer);
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ activate_array(layer.output, m*n, layer.activation, 0.);
+}
+
+#ifdef GPU
+void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
+{
+ int m = layer.n;
+ int k = layer.size*layer.size*layer.c;
+ int n = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
+
+ cl_write_array(layer.filters_cl, layer.filters, m*k);
+ cl_mem a = layer.filters_cl;
+ cl_mem b = layer.col_image_cl;
+ cl_mem c = layer.output_cl;
+ im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, b);
+ gemm_ongpu(0,0,m,n,k,1,a,k,b,n,0,c,n);
+ activate_array_ongpu(layer.output_cl, m*n, layer.activation, 0.);
+ cl_read_array(layer.output_cl, layer.output, m*n);
+}
+#endif
+
+void learn_bias_convolutional_layer(convolutional_layer layer)
+{
+ int i,b;
+ int size = convolutional_out_height(layer)
+ *convolutional_out_width(layer);
+ for(b = 0; b < layer.batch; ++b){
for(i = 0; i < layer.n; ++i){
- two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, input, j);
+ layer.bias_updates[i] += mean_array(layer.delta+size*(i+b*layer.n), size);
}
}
-
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
}
-void learn_convolutional_layer(image input, convolutional_layer layer)
+void backward_convolutional_layer(convolutional_layer layer, float *delta)
{
- int i;
- for(i = 0; i < layer.n; ++i){
- kernel_update(input, layer.kernel_updates[i], layer.stride, i, layer.output);
+ int m = layer.n;
+ int n = layer.size*layer.size*layer.c;
+ int k = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
+ gradient_array(layer.output, m*k, layer.activation, layer.delta);
+ learn_bias_convolutional_layer(layer);
+
+ float *a = layer.delta;
+ float *b = layer.col_image;
+ float *c = layer.filter_updates;
+
+ gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+
+ if(delta){
+ int i;
+ m = layer.size*layer.size*layer.c;
+ k = layer.n;
+ n = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
+
+ a = layer.filters;
+ b = layer.delta;
+ c = layer.col_image;
+
+ gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
+
+ memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+ for(i = 0; i < layer.batch; ++i){
+ col2im_cpu(c+i*n/layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, delta+i*n/layer.batch);
+ }
}
- image old_input = copy_image(input);
- backpropagate_convolutional_layer(input, layer);
- for(i = 0; i < input.h*input.w*input.c; ++i){
- input.data[i] *= convolution_gradient(old_input.data[i]);
- }
- free_image(old_input);
}
-void update_convolutional_layer(convolutional_layer layer, double step)
+void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
{
- int i,j;
- for(i = 0; i < layer.n; ++i){
- int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c;
- for(j = 0; j < pixels; ++j){
- layer.kernels[i].data[j] += step*layer.kernel_updates[i].data[j];
+ int size = layer.size*layer.size*layer.c*layer.n;
+ axpy_cpu(layer.n, step, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.n, momentum, layer.bias_updates, 1);
+
+ scal_cpu(size, 1.-step*decay, layer.filters, 1);
+ axpy_cpu(size, step, layer.filter_updates, 1, layer.filters, 1);
+ scal_cpu(size, momentum, layer.filter_updates, 1);
+}
+
+
+image get_convolutional_filter(convolutional_layer layer, int i)
+{
+ int h = layer.size;
+ int w = layer.size;
+ int c = layer.c;
+ return float_to_image(h,w,c,layer.filters+i*h*w*c);
+}
+
+image *weighted_sum_filters(convolutional_layer layer, image *prev_filters)
+{
+ image *filters = calloc(layer.n, sizeof(image));
+ int i,j,k,c;
+ if(!prev_filters){
+ for(i = 0; i < layer.n; ++i){
+ filters[i] = copy_image(get_convolutional_filter(layer, i));
}
- zero_image(layer.kernel_updates[i]);
}
+ else{
+ image base = prev_filters[0];
+ for(i = 0; i < layer.n; ++i){
+ image filter = get_convolutional_filter(layer, i);
+ filters[i] = make_image(base.h, base.w, base.c);
+ for(j = 0; j < layer.size; ++j){
+ for(k = 0; k < layer.size; ++k){
+ for(c = 0; c < layer.c; ++c){
+ float weight = get_pixel(filter, j, k, c);
+ image prev_filter = copy_image(prev_filters[c]);
+ scale_image(prev_filter, weight);
+ add_into_image(prev_filter, filters[i], 0,0);
+ free_image(prev_filter);
+ }
+ }
+ }
+ }
+ }
+ return filters;
+}
+
+image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters)
+{
+ image *single_filters = weighted_sum_filters(layer, 0);
+ show_images(single_filters, layer.n, window);
+
+ image delta = get_convolutional_image(layer);
+ image dc = collapse_image_layers(delta, 1);
+ char buff[256];
+ sprintf(buff, "%s: Output", window);
+ show_image(dc, buff);
+ save_image(dc, buff);
+ free_image(dc);
+ return single_filters;
}
--
Gitblit v1.10.0