From aa5996d58e68edfbefe51061856aecd549dd09c4 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 13 Jan 2015 01:27:08 +0000
Subject: [PATCH] Faster
---
src/convolutional_layer.c | 371 ++++++++++++++++++++++++++++++++++++++--------------
1 files changed, 272 insertions(+), 99 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 31a4af6..fc5cb0e 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -2,15 +2,22 @@
#include "utils.h"
#include "mini_blas.h"
#include <stdio.h>
+#include <time.h>
int convolutional_out_height(convolutional_layer layer)
{
- return (layer.h-layer.size)/layer.stride + 1;
+ int h = layer.h;
+ if (!layer.pad) h -= layer.size;
+ else h -= 1;
+ return h/layer.stride + 1;
}
int convolutional_out_width(convolutional_layer layer)
{
- return (layer.w-layer.size)/layer.stride + 1;
+ int w = layer.w;
+ if (!layer.pad) w -= layer.size;
+ else w -= 1;
+ return w/layer.stride + 1;
}
image get_convolutional_image(convolutional_layer layer)
@@ -31,11 +38,16 @@
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)
+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)
{
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->learning_rate = learning_rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
+
layer->h = h;
layer->w = w;
layer->c = c;
@@ -43,30 +55,41 @@
layer->batch = batch;
layer->stride = stride;
layer->size = size;
+ layer->pad = pad;
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());
+ float scale = 1./sqrt(size*size*c);
+ //scale = .05;
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal();
for(i = 0; i < n; ++i){
//layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = 0;
+ layer->biases[i] = scale;
}
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->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));
+
+ #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);
+
+ layer->biases_cl = cl_make_array(layer->biases, n);
+ layer->bias_updates_cl = cl_make_array(layer->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);
+ #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;
}
@@ -80,123 +103,114 @@
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));
+ 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 forward_convolutional_layer(const convolutional_layer layer, float *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)*
- layer.batch;
-
- float *a = layer.filters;
- float *b = layer.col_image;
- float *c = layer.output;
- for(i = 0; i < layer.batch; ++i){
- im2col_gpu(in+i*(n/layer.batch), layer.c, layer.h, layer.w, layer.size, layer.stride, b+i*(n/layer.batch));
- }
- gemm(0,0,m,n,k,1,a,k,b,n,0,c,n);
- activate_array(layer.output, m*n, layer.activation);
-}
-
-void learn_bias_convolutional_layer(convolutional_layer layer)
+void bias_output(const convolutional_layer layer)
{
int i,j,b;
- int size = convolutional_out_height(layer)
- *convolutional_out_width(layer);
+ 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){
- float sum = 0;
- for(j = 0; j < size; ++j){
- sum += layer.delta[j+size*(i+b*layer.n)];
+ for(j = 0; j < out_h*out_w; ++j){
+ layer.output[(b*layer.n + i)*out_h*out_w + j] = layer.biases[i];
}
- layer.bias_updates[i] += sum/size;
}
}
}
-void learn_convolutional_layer(convolutional_layer layer)
+void forward_convolutional_layer(const convolutional_layer layer, float *in)
{
- 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);
-}
-
-void backward_convolutional_layer(convolutional_layer layer, float *delta)
-{
+ int out_h = convolutional_out_height(layer);
+ int out_w = convolutional_out_width(layer);
int i;
- int m = layer.size*layer.size*layer.c;
- int k = layer.n;
- int n = convolutional_out_height(layer)*
- convolutional_out_width(layer)*
- layer.batch;
+
+ 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.delta;
- float *c = layer.col_image;
+ float *b = layer.col_image;
+ float *c = layer.output;
- 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);
+ 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)
+{
+ 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);
+ }
}
}
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
+void backward_convolutional_layer(convolutional_layer layer, float *in, float *delta)
+{
+ 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);
+
+ if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+
+ for(i = 0; i < layer.batch; ++i){
+ float *a = layer.delta + i*m*k;
+ float *b = layer.col_image;
+ float *c = layer.filter_updates;
+
+ float *im = in+i*layer.c*layer.h*layer.w;
+
+ im2col_cpu(im, layer.c, layer.h, layer.w,
+ layer.size, layer.stride, layer.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;
+
+ 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);
+ }
+ }
+}
+
+void update_convolutional_layer(convolutional_layer layer)
{
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);
+ axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.n, layer.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);
+ 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);
}
-void test_convolutional_layer()
-{
- convolutional_layer l = *make_convolutional_layer(1,4,4,1,1,3,1,LINEAR);
- float input[] = {1,2,3,4,
- 5,6,7,8,
- 9,10,11,12,
- 13,14,15,16};
- float filter[] = {.5, 0, .3,
- 0 , 1, 0,
- .2 , 0, 1};
- float delta[] = {1, 2,
- 3, 4};
- float in_delta[] = {.5,1,.3,.6,
- 5,6,7,8,
- 9,10,11,12,
- 13,14,15,16};
- l.filters = filter;
- forward_convolutional_layer(l, input);
- l.delta = delta;
- learn_convolutional_layer(l);
- image filter_updates = float_to_image(3,3,1,l.filter_updates);
- print_image(filter_updates);
- printf("Delta:\n");
- backward_convolutional_layer(l, in_delta);
- pm(4,4,in_delta);
-}
image get_convolutional_filter(convolutional_layer layer, int i)
{
@@ -245,9 +259,168 @@
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;
}
+#ifdef GPU
+#define BLOCK 32
+
+#define STR_HELPER(x) #x
+#define STR(x) STR_HELPER(x)
+
+
+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", "-D BLOCK=" STR(BLOCK));
+ init = 1;
+ }
+ return kernel;
+}
+
+void learn_bias_convolutional_layer_ongpu(convolutional_layer layer)
+{
+ int size = convolutional_out_height(layer) * convolutional_out_width(layer);
+
+ 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*BLOCK};
+ const size_t local_size[] = {BLOCK};
+
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, local_size, 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", "-D BLOCK=" STR(BLOCK));
+ 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_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