From 741ada451cc7fee1b9a4c3deaec6af87a2af7497 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 17 Aug 2015 16:31:25 +0000
Subject: [PATCH] Added Darknet reference model
---
src/convolutional_layer.c | 435 ++++++++++++++++++++++++++++-------------------------
1 files changed, 231 insertions(+), 204 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 5aa76ee..7dcf5a4 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,255 +1,282 @@
#include "convolutional_layer.h"
#include "utils.h"
-#include "mini_blas.h"
+#include "im2col.h"
+#include "col2im.h"
+#include "blas.h"
+#include "gemm.h"
#include <stdio.h>
+#include <time.h>
-int convolutional_out_height(convolutional_layer layer)
+int convolutional_out_height(convolutional_layer l)
{
- return (layer.h-layer.size)/layer.stride + 1;
+ int h = l.h;
+ if (!l.pad) h -= l.size;
+ else h -= 1;
+ return h/l.stride + 1;
}
-int convolutional_out_width(convolutional_layer layer)
+int convolutional_out_width(convolutional_layer l)
{
- return (layer.w-layer.size)/layer.stride + 1;
+ int w = l.w;
+ if (!l.pad) w -= l.size;
+ else w -= 1;
+ return w/l.stride + 1;
}
-image get_convolutional_image(convolutional_layer layer)
+image get_convolutional_image(convolutional_layer l)
{
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);
+ 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 layer)
+image get_convolutional_delta(convolutional_layer l)
{
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);
+ 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, ACTIVATION activation)
+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;
- 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->size = size;
+ convolutional_layer l = {0};
+ l.type = CONVOLUTIONAL;
- 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));
+ l.h = h;
+ l.w = w;
+ l.c = c;
+ l.n = n;
+ l.batch = batch;
+ l.stride = stride;
+ l.size = size;
+ l.pad = pad;
- 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());
+ 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->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = .5;
+ l.biases[i] = scale;
}
- int out_h = convolutional_out_height(*layer);
- int out_w = convolutional_out_width(*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;
- 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));
+ 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
- layer->activation = activation;
+ 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);
- srand(0);
- return layer;
+ return l;
}
-void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c)
+void resize_convolutional_layer(convolutional_layer *l, int w, int h)
{
- layer->h = h;
- layer->w = w;
- layer->c = c;
- int out_h = convolutional_out_height(*layer);
- int out_w = convolutional_out_width(*layer);
+ l->w = w;
+ l->h = h;
+ int out_w = convolutional_out_width(*l);
+ int out_h = convolutional_out_height(*l);
- 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));
+ l->out_w = out_w;
+ l->out_h = out_h;
+
+ l->outputs = l->out_h * l->out_w * l->out_c;
+ l->inputs = l->w * l->h * l->c;
+
+ 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(0, out_h*out_w*l->size*l->size*l->c);
+ l->delta_gpu = cuda_make_array(0, l->batch*out_h*out_w*l->n);
+ l->output_gpu = cuda_make_array(0, l->batch*out_h*out_w*l->n);
+ #endif
}
-void bias_output(const convolutional_layer layer)
+void bias_output(float *output, float *biases, int batch, int n, int size)
{
- int i,j;
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- for(i = 0; i < layer.n; ++i){
- for(j = 0; j < out_h*out_w; ++j){
- layer.output[i*out_h*out_w + j] = layer.biases[i];
- }
- }
-}
-
-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){
- layer.bias_updates[i] += mean_array(layer.delta+size*(i+b*layer.n), size);
- }
- }
-}
-
-void backward_convolutional_layer(convolutional_layer layer, float *delta)
-{
- 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);
- }
- }
-}
-
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
-{
- 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));
- }
- }
- 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);
- }
- }
+ 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];
}
}
}
+}
+
+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 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;
+ 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);
+
+ 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);
+ }
+ }
+}
+
+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);
+ }
+ }
+}
+
+void rescale_filters(convolutional_layer l, float scale, float trans)
+{
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_filter(l, i);
+ if (im.c == 3) {
+ scale_image(im, scale);
+ float sum = sum_array(im.data, im.w*im.h*im.c);
+ l.biases[i] += sum*trans;
+ }
+ }
+}
+
+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));
+ normalize_image(filters[i]);
+ }
return filters;
}
-image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters)
+image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_filters)
{
- image *single_filters = weighted_sum_filters(layer, 0);
- show_images(single_filters, layer.n, window);
+ image *single_filters = get_filters(l);
+ show_images(single_filters, l.n, window);
- image delta = get_convolutional_image(layer);
+ 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);
+ //show_image(dc, buff);
+ //save_image(dc, buff);
free_image(dc);
return single_filters;
}
--
Gitblit v1.10.0