From a4485b8a6656c2a2fa0b78ca7c035523c8149b8c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 10 Jun 2015 15:49:25 +0000
Subject: [PATCH] Merge branch 'master' of pjreddie.com:jnet
---
src/convolutional_layer.c | 491 ++++++++++++++++++------------------------------------
1 files changed, 167 insertions(+), 324 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 5b4e0b5..e29d995 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,259 +1,257 @@
#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)
{
- int h = layer.h;
- if (!layer.pad) h -= layer.size;
+ int h = l.h;
+ if (!l.pad) h -= l.size;
else h -= 1;
- return h/layer.stride + 1;
+ return h/l.stride + 1;
}
-int convolutional_out_width(convolutional_layer layer)
+int convolutional_out_width(convolutional_layer l)
{
- int w = layer.w;
- if (!layer.pad) w -= layer.size;
+ int w = l.w;
+ if (!l.pad) w -= l.size;
else w -= 1;
- return w/layer.stride + 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, int pad, ACTIVATION activation, float learning_rate, float momentum, float decay)
+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));
+ convolutional_layer l = {0};
+ l.type = CONVOLUTIONAL;
- layer->learning_rate = learning_rate;
- layer->momentum = momentum;
- layer->decay = decay;
+ 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->h = h;
- layer->w = w;
- layer->c = c;
- layer->n = n;
- layer->batch = batch;
- layer->stride = stride;
- layer->size = size;
- layer->pad = pad;
+ l.filters = calloc(c*n*size*size, sizeof(float));
+ l.filter_updates = calloc(c*n*size*size, sizeof(float));
- layer->filters = calloc(c*n*size*size, sizeof(float));
- layer->filter_updates = calloc(c*n*size*size, sizeof(float));
-
- layer->biases = calloc(n, sizeof(float));
- layer->bias_updates = calloc(n, sizeof(float));
- float scale = 1./(size*size*c);
- scale = .01;
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal();
+ 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] = .01;
+ 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(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
- layer->filters_cl = cl_make_array(layer->filters, c*n*size*size);
- layer->filter_updates_cl = cl_make_array(layer->filter_updates, c*n*size*size);
+ 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);
- layer->biases_cl = cl_make_array(layer->biases, n);
- layer->bias_updates_cl = cl_make_array(layer->bias_updates, n);
+ l.biases_gpu = cuda_make_array(l.biases, n);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
- layer->col_image_cl = cl_make_array(layer->col_image, out_h*out_w*size*size*c);
- layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n);
- layer->output_cl = cl_make_array(layer->output, layer->batch*out_h*out_w*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);
- 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 h, int w)
{
- layer->h = h;
- layer->w = w;
- layer->c = c;
- int out_h = convolutional_out_height(*layer);
- int out_w = convolutional_out_width(*layer);
+ l->h = h;
+ l->w = w;
+ int out_h = convolutional_out_height(*l);
+ int out_w = convolutional_out_width(*l);
- layer->col_image = realloc(layer->col_image,
- 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->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 bias_output(const convolutional_layer layer)
+void bias_output(float *output, float *biases, int batch, int n, int size)
{
int i,j,b;
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- for(b = 0; b < layer.batch; ++b){
- for(i = 0; i < layer.n; ++i){
- for(j = 0; j < out_h*out_w; ++j){
- layer.output[(b*layer.n + i)*out_h*out_w + j] = layer.biases[i];
+ 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 forward_convolutional_layer(const convolutional_layer layer, float *in)
-{
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- int i;
-
- bias_output(layer);
-
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = out_h*out_w;
-
- float *a = layer.filters;
- float *b = layer.col_image;
- float *c = layer.output;
-
-
- for(i = 0; i < layer.batch; ++i){
- im2col_cpu(in, layer.c, layer.h, layer.w,
- layer.size, layer.stride, layer.pad, b);
- gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
- c += n*m;
- in += layer.c*layer.h*layer.w;
- }
- activate_array(layer.output, m*n*layer.batch, layer.activation);
-}
-
-void learn_bias_convolutional_layer(convolutional_layer layer)
+void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
{
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] += sum_array(layer.delta+size*(i+b*layer.n), size);
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ bias_updates[i] += sum_array(delta+size*(i+b*n), size);
}
}
}
-void backward_convolutional_layer(convolutional_layer layer, float *in, float *delta)
+
+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 = layer.n;
- int n = layer.size*layer.size*layer.c;
- int k = convolutional_out_height(layer)*
- convolutional_out_width(layer);
- gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
- learn_bias_convolutional_layer(layer);
+ int m = l.n;
+ int n = l.size*l.size*l.c;
+ int k = convolutional_out_height(l)*
+ convolutional_out_width(l);
- if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+ 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 < layer.batch; ++i){
- float *a = layer.delta + i*m*k;
- float *b = layer.col_image;
- float *c = layer.filter_updates;
+ if(state.delta) memset(state.delta, 0, l.batch*l.h*l.w*l.c*sizeof(float));
- float *im = in+i*layer.c*layer.h*layer.w;
+ for(i = 0; i < l.batch; ++i){
+ float *a = l.delta + i*m*k;
+ float *b = l.col_image;
+ float *c = l.filter_updates;
- im2col_cpu(im, layer.c, layer.h, layer.w,
- layer.size, layer.stride, layer.pad, b);
+ 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(delta){
- a = layer.filters;
- b = layer.delta + i*m*k;
- c = layer.col_image;
+ 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(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta+i*layer.c*layer.h*layer.w);
+ 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 layer)
+void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay)
{
- int size = layer.size*layer.size*layer.c*layer.n;
- axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
- scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1);
+ 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, -layer.decay, layer.filters, 1, layer.filter_updates, 1);
- axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
- scal_cpu(size, layer.momentum, layer.filter_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 layer, int i)
+image get_convolutional_filter(convolutional_layer l, 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);
+ 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);
}
-image *weighted_sum_filters(convolutional_layer layer, image *prev_filters)
+void rgbgr_filters(convolutional_layer l)
{
- 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));
- }
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_filter(l, i);
+ if (im.c == 3) rgbgr_image(im);
}
- 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);
- }
- }
- }
- }
+}
+
+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));
}
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);
@@ -263,158 +261,3 @@
return single_filters;
}
-#ifdef GPU
-
-cl_kernel get_convolutional_learn_bias_kernel()
-{
- static int init = 0;
- static cl_kernel kernel;
- if(!init){
- kernel = get_kernel("src/convolutional_layer.cl", "learn_bias", 0);
- init = 1;
- }
- return kernel;
-}
-
-void learn_bias_convolutional_layer_ongpu(convolutional_layer layer)
-{
- int size = convolutional_out_height(layer) * convolutional_out_width(layer);
-
- cl_setup();
- cl_kernel kernel = get_convolutional_learn_bias_kernel();
- cl_command_queue queue = cl.queue;
-
- cl_uint i = 0;
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.batch), (void*) &layer.batch);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n);
- cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.delta_cl), (void*) &layer.delta_cl);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.bias_updates_cl), (void*) &layer.bias_updates_cl);
- check_error(cl);
-
- const size_t global_size[] = {layer.n};
-
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
- check_error(cl);
-}
-
-cl_kernel get_convolutional_bias_kernel()
-{
- static int init = 0;
- static cl_kernel kernel;
- if(!init){
- kernel = get_kernel("src/convolutional_layer.cl", "bias", 0);
- init = 1;
- }
- return kernel;
-}
-
-void bias_output_gpu(const convolutional_layer layer)
-{
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- int size = out_h*out_w;
-
- cl_setup();
- cl_kernel kernel = get_convolutional_bias_kernel();
- cl_command_queue queue = cl.queue;
-
- cl_uint i = 0;
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n);
- cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.biases_cl), (void*) &layer.biases_cl);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
- check_error(cl);
-
- const size_t global_size[] = {layer.n*size, layer.batch};
-
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 2, 0, global_size, 0, 0, 0, 0);
- check_error(cl);
-}
-
-//#define TIMEIT
-
-void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
-{
- int i;
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = convolutional_out_height(layer)*
- convolutional_out_width(layer);
-
- bias_output_gpu(layer);
-
- for(i = 0; i < layer.batch; ++i){
- im2col_ongpu(in, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
- cl_mem a = layer.filters_cl;
- cl_mem b = layer.col_image_cl;
- cl_mem c = layer.output_cl;
- gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,0,n,1.,c,i*m*n,n);
- }
- activate_array_ongpu(layer.output_cl, m*n*layer.batch, layer.activation);
-}
-
-void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in, cl_mem delta_cl)
-{
- int i;
- int m = layer.n;
- int n = layer.size*layer.size*layer.c;
- int k = convolutional_out_height(layer)*
- convolutional_out_width(layer);
- gradient_array_ongpu(layer.output_cl, m*k*layer.batch, layer.activation, layer.delta_cl);
- learn_bias_convolutional_layer_ongpu(layer);
-
- if(delta_cl) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, delta_cl, 1);
-
- for(i = 0; i < layer.batch; ++i){
- cl_mem a = layer.delta_cl;
- cl_mem b = layer.col_image_cl;
- cl_mem c = layer.filter_updates_cl;
-
- im2col_ongpu(in, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
- gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,0,k,1,c,0,n);
-
- if(delta_cl){
-
- cl_mem a = layer.filters_cl;
- cl_mem b = layer.delta_cl;
- cl_mem c = layer.col_image_cl;
-
- gemm_ongpu_offset(1,0,n,k,m,1,a,0,n,b,i*k*m,k,0,c,0,k);
-
- col2im_ongpu(layer.col_image_cl, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl);
- }
- }
-}
-
-void pull_convolutional_layer(convolutional_layer layer)
-{
- cl_read_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
- cl_read_array(layer.biases_cl, layer.biases, layer.n);
- cl_read_array(layer.filter_updates_cl, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
- cl_read_array(layer.bias_updates_cl, layer.bias_updates, layer.n);
-}
-
-void push_convolutional_layer(convolutional_layer layer)
-{
- cl_write_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
- cl_write_array(layer.biases_cl, layer.biases, layer.n);
- cl_write_array(layer.filter_updates_cl, layer.filter_updates, layer.c*layer.n*layer.size*layer.size);
- cl_write_array(layer.bias_updates_cl, layer.bias_updates, layer.n);
-}
-
-void update_convolutional_layer_gpu(convolutional_layer layer)
-{
- int size = layer.size*layer.size*layer.c*layer.n;
- axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
- scal_ongpu(layer.n,layer.momentum, layer.bias_updates_cl, 1);
-
- axpy_ongpu(size, -layer.decay, layer.filters_cl, 1, layer.filter_updates_cl, 1);
- axpy_ongpu(size, layer.learning_rate, layer.filter_updates_cl, 1, layer.filters_cl, 1);
- scal_ongpu(size, layer.momentum, layer.filter_updates_cl, 1);
- pull_convolutional_layer(layer);
-}
-
-
-#endif
-
--
Gitblit v1.10.0