From cbc9984a17b3452da4fee397aac912f1e9f7a4c3 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 10 Jun 2015 07:11:41 +0000
Subject: [PATCH] NIPS
---
src/convolutional_layer.c | 287 ++++++++++++++++++++++++++++++++++++++++++++++-----------
1 files changed, 232 insertions(+), 55 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index f83622b..e29d995 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,86 +1,263 @@
#include "convolutional_layer.h"
+#include "utils.h"
+#include "im2col.h"
+#include "col2im.h"
+#include "blas.h"
+#include "gemm.h"
+#include <stdio.h>
+#include <time.h>
-double convolution_activation(double x)
+int convolutional_out_height(convolutional_layer l)
{
- return x*(x>0);
+ int h = l.h;
+ if (!l.pad) h -= l.size;
+ else h -= 1;
+ return h/l.stride + 1;
}
-double convolution_gradient(double x)
+int convolutional_out_width(convolutional_layer l)
{
- return (x>=0);
+ int w = l.w;
+ if (!l.pad) w -= l.size;
+ else w -= 1;
+ return w/l.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 l)
+{
+ int h,w,c;
+ h = convolutional_out_height(l);
+ w = convolutional_out_width(l);
+ c = l.n;
+ return float_to_image(w,h,c,l.output);
+}
+
+image get_convolutional_delta(convolutional_layer l)
+{
+ int h,w,c;
+ h = convolutional_out_height(l);
+ w = convolutional_out_width(l);
+ c = l.n;
+ 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 i;
- convolutional_layer layer;
- layer.n = n;
- layer.stride = stride;
- layer.kernels = calloc(n, sizeof(image));
- layer.kernel_updates = calloc(n, sizeof(image));
+ convolutional_layer l = {0};
+ l.type = CONVOLUTIONAL;
+
+ l.h = h;
+ l.w = w;
+ l.c = c;
+ l.n = n;
+ l.batch = batch;
+ l.stride = stride;
+ l.size = size;
+ l.pad = pad;
+
+ l.filters = calloc(c*n*size*size, sizeof(float));
+ l.filter_updates = calloc(c*n*size*size, sizeof(float));
+
+ l.biases = calloc(n, sizeof(float));
+ l.bias_updates = calloc(n, sizeof(float));
+ //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 < n; ++i){
- layer.kernels[i] = make_random_kernel(size, c);
- layer.kernel_updates[i] = make_random_kernel(size, c);
+ l.biases[i] = scale;
}
- layer.output = make_image((h-1)/stride+1, (w-1)/stride+1, n);
- layer.upsampled = make_image(h,w,n);
- return layer;
+ int out_h = convolutional_out_height(l);
+ int out_w = convolutional_out_width(l);
+ l.out_h = out_h;
+ l.out_w = out_w;
+ l.out_c = n;
+ l.outputs = l.out_h * l.out_w * l.out_c;
+ l.inputs = l.w * l.h * l.c;
+
+ l.col_image = calloc(out_h*out_w*size*size*c, sizeof(float));
+ l.output = calloc(l.batch*out_h * out_w * n, sizeof(float));
+ l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float));
+
+ #ifdef GPU
+ l.filters_gpu = cuda_make_array(l.filters, c*n*size*size);
+ l.filter_updates_gpu = cuda_make_array(l.filter_updates, c*n*size*size);
+
+ l.biases_gpu = cuda_make_array(l.biases, n);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
+
+ l.col_image_gpu = cuda_make_array(l.col_image, out_h*out_w*size*size*c);
+ 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);
+ #endif
+ l.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);
+
+ return l;
}
-void run_convolutional_layer(const image input, const convolutional_layer layer)
+void resize_convolutional_layer(convolutional_layer *l, int h, int w)
{
- int i;
- for(i = 0; i < layer.n; ++i){
- convolve(input, layer.kernels[i], layer.stride, i, layer.output);
- }
- for(i = 0; i < input.h*input.w*input.c; ++i){
- input.data[i] = convolution_activation(input.data[i]);
- }
+ l->h = h;
+ l->w = w;
+ int out_h = convolutional_out_height(*l);
+ int out_w = convolutional_out_width(*l);
+
+ l->col_image = realloc(l->col_image,
+ out_h*out_w*l->size*l->size*l->c*sizeof(float));
+ l->output = realloc(l->output,
+ l->batch*out_h * out_w * l->n*sizeof(float));
+ l->delta = realloc(l->delta,
+ l->batch*out_h * out_w * l->n*sizeof(float));
+
+ #ifdef GPU
+ cuda_free(l->col_image_gpu);
+ cuda_free(l->delta_gpu);
+ cuda_free(l->output_gpu);
+
+ l->col_image_gpu = cuda_make_array(l->col_image, out_h*out_w*l->size*l->size*l->c);
+ l->delta_gpu = cuda_make_array(l->delta, l->batch*out_h*out_w*l->n);
+ l->output_gpu = cuda_make_array(l->output, l->batch*out_h*out_w*l->n);
+ #endif
}
-void backpropagate_layer(image input, convolutional_layer layer)
+void bias_output(float *output, float *biases, int batch, int n, int size)
{
- 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_layer_convolve(image input, convolutional_layer layer)
-{
- int i,j;
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
-
- zero_image(input);
- upsample_image(layer.output, layer.stride, layer.upsampled);
- for(j = 0; j < input.c; ++j){
- for(i = 0; i < layer.n; ++i){
- two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, input, j);
+ 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];
+ }
}
}
+}
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
+void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
+{
+ int i,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ bias_updates[i] += sum_array(delta+size*(i+b*n), size);
+ }
}
}
-void error_convolutional_layer(image input, convolutional_layer layer)
+
+void forward_convolutional_layer(const convolutional_layer l, network_state state)
+{
+ int out_h = convolutional_out_height(l);
+ int out_w = convolutional_out_width(l);
+ int i;
+
+ bias_output(l.output, l.biases, l.batch, l.n, out_h*out_w);
+
+ int m = l.n;
+ int k = l.size*l.size*l.c;
+ int n = out_h*out_w;
+
+ float *a = l.filters;
+ float *b = l.col_image;
+ float *c = l.output;
+
+ for(i = 0; i < l.batch; ++i){
+ im2col_cpu(state.input, l.c, l.h, l.w,
+ l.size, l.stride, l.pad, b);
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ c += n*m;
+ state.input += l.c*l.h*l.w;
+ }
+ activate_array(l.output, m*n*l.batch, l.activation);
+}
+
+void backward_convolutional_layer(convolutional_layer l, network_state state)
{
int i;
- for(i = 0; i < layer.n; ++i){
- kernel_update(input, layer.kernel_updates[i], layer.stride, i, layer.output);
+ int m = l.n;
+ int n = l.size*l.size*l.c;
+ int k = convolutional_out_height(l)*
+ convolutional_out_width(l);
+
+ gradient_array(l.output, m*k*l.batch, l.activation, l.delta);
+ backward_bias(l.bias_updates, l.delta, l.batch, l.n, k);
+
+ if(state.delta) memset(state.delta, 0, l.batch*l.h*l.w*l.c*sizeof(float));
+
+ for(i = 0; i < l.batch; ++i){
+ float *a = l.delta + i*m*k;
+ float *b = l.col_image;
+ float *c = l.filter_updates;
+
+ float *im = state.input+i*l.c*l.h*l.w;
+
+ im2col_cpu(im, l.c, l.h, l.w,
+ l.size, l.stride, l.pad, b);
+ gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+
+ if(state.delta){
+ a = l.filters;
+ b = l.delta + i*m*k;
+ c = l.col_image;
+
+ gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
+
+ col2im_cpu(l.col_image, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w);
+ }
}
- image old_input = copy_image(input);
- zero_image(input);
- for(i = 0; i < layer.n; ++i){
- back_convolve(input, layer.kernels[i], layer.stride, i, layer.output);
+}
+
+void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay)
+{
+ int size = l.size*l.size*l.c*l.n;
+ axpy_cpu(l.n, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
+ scal_cpu(l.n, momentum, l.bias_updates, 1);
+
+ axpy_cpu(size, -decay*batch, l.filters, 1, l.filter_updates, 1);
+ axpy_cpu(size, learning_rate/batch, l.filter_updates, 1, l.filters, 1);
+ scal_cpu(size, momentum, l.filter_updates, 1);
+}
+
+
+image get_convolutional_filter(convolutional_layer l, int i)
+{
+ int h = l.size;
+ int w = l.size;
+ int c = l.c;
+ return float_to_image(w,h,c,l.filters+i*h*w*c);
+}
+
+void rgbgr_filters(convolutional_layer l)
+{
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_filter(l, i);
+ if (im.c == 3) rgbgr_image(im);
}
- for(i = 0; i < input.h*input.w*input.c; ++i){
- input.data[i] = input.data[i]*convolution_gradient(input.data[i]);
+}
+
+image *get_filters(convolutional_layer l)
+{
+ image *filters = calloc(l.n, sizeof(image));
+ int i;
+ for(i = 0; i < l.n; ++i){
+ filters[i] = copy_image(get_convolutional_filter(l, i));
}
- free_image(old_input);
+ return filters;
+}
+
+image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_filters)
+{
+ image *single_filters = get_filters(l);
+ show_images(single_filters, l.n, window);
+
+ image delta = get_convolutional_image(l);
+ 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