From 6e1d5b45de988bb795c4c505f22f2170a78b7746 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 20 Jan 2015 06:06:18 +0000
Subject: [PATCH] fast sort of working
---
src/convolutional_layer.c | 163 ++++++++++++++++++++++++++---------------------------
1 files changed, 80 insertions(+), 83 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 42f4f21..4e8c44b 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -59,34 +59,31 @@
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);
- scale = .05;
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*2*(rand_uniform()-.5);
+ 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] = .5;
+ 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->filter_momentum_cl = cl_make_array(layer->filter_momentum, 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->bias_momentum_cl = cl_make_array(layer->bias_momentum, n);
- layer->col_image_cl = cl_make_array(layer->col_image, layer->batch*out_h*out_w*size*size*c);
+ 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
@@ -106,7 +103,7 @@
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,
@@ -143,13 +140,13 @@
float *b = layer.col_image;
float *c = layer.output;
- im2col_cpu(in, layer.batch, layer.c, layer.h, layer.w,
- layer.size, layer.stride, layer.pad, b);
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);
- b += k*n;
c += n*m;
+ in += layer.c*layer.h*layer.w;
}
activate_array(layer.output, m*n*layer.batch, layer.activation);
}
@@ -166,45 +163,40 @@
}
}
-void backward_convolutional_layer(convolutional_layer layer, float *delta)
+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);
- float *a = layer.delta;
- float *b = layer.col_image;
- float *c = layer.filter_updates;
+ 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);
- a += m*k;
- b += k*n;
- }
- if(delta){
- m = layer.size*layer.size*layer.c;
- k = layer.n;
- n = convolutional_out_height(layer)*
- convolutional_out_width(layer);
+ if(delta){
+ a = layer.filters;
+ b = layer.delta + i*m*k;
+ c = layer.col_image;
- a = layer.filters;
- b = layer.delta;
- c = layer.col_image;
+ gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
- for(i = 0; i < layer.batch; ++i){
- gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
- b += k*n;
- c += m*n;
+ 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);
}
-
- memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
-
- col2im_cpu(layer.col_image, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta);
}
}
@@ -214,7 +206,7 @@
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.-layer.learning_rate*layer.decay, layer.filters, 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);
}
@@ -274,13 +266,18 @@
}
#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", 0);
+ kernel = get_kernel("src/convolutional_layer.cl", "learn_bias", "-D BLOCK=" STR(BLOCK));
init = 1;
}
return kernel;
@@ -290,7 +287,6 @@
{
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;
@@ -302,18 +298,40 @@
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};
+ const size_t global_size[] = {layer.n*BLOCK};
+ const size_t local_size[] = {BLOCK};
- clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, local_size, 0, 0, 0);
check_error(cl);
}
+void test_learn_bias(convolutional_layer l)
+{
+ int i;
+ int size = convolutional_out_height(l) * convolutional_out_width(l);
+ for(i = 0; i < size*l.batch*l.n; ++i){
+ l.delta[i] = rand_uniform();
+ }
+ for(i = 0; i < l.n; ++i){
+ l.bias_updates[i] = rand_uniform();
+ }
+ cl_write_array(l.delta_cl, l.delta, size*l.batch*l.n);
+ cl_write_array(l.bias_updates_cl, l.bias_updates, l.n);
+ float *gpu = calloc(l.n, sizeof(float));
+ cl_read_array(l.bias_updates_cl, gpu, l.n);
+ for(i = 0; i < l.n; ++i) printf("%.9g %.9g\n", l.bias_updates[i], gpu[i]);
+ learn_bias_convolutional_layer_ongpu(l);
+ learn_bias_convolutional_layer(l);
+ cl_read_array(l.bias_updates_cl, gpu, l.n);
+ for(i = 0; i < l.n; ++i) printf("%.9g %.9g\n", l.bias_updates[i], gpu[i]);
+}
+
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);
+ kernel = get_kernel("src/convolutional_layer.cl", "bias", "-D BLOCK=" STR(BLOCK));
init = 1;
}
return kernel;
@@ -325,7 +343,6 @@
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;
@@ -336,9 +353,9 @@
cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
check_error(cl);
- const size_t global_size[] = {layer.batch, layer.n*size};
+ const size_t global_size[] = {layer.n*size, layer.batch};
- clEnqueueNDRangeKernel(queue, kernel, 2, 0, global_size, 0, 0, 0, 0);
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 2, 0, global_size, 0, 0, 0, 0);
check_error(cl);
}
@@ -354,36 +371,17 @@
bias_output_gpu(layer);
- #ifdef TIMEIT
- clock_t time = clock();
- printf("Forward\n");
- #endif
-
- im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
-
- #ifdef TIMEIT
- clFinish(cl.queue);
- printf("Im2col %f\n", sec(clock()-time));
- time = clock();
- #endif
-
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,i*k*n,n,1.,c,i*m*n,n);
+ gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,0,n,1.,c,i*m*n,n);
}
- #ifdef TIMEIT
- clFinish(cl.queue);
- printf("Gemm %f\n", sec(clock()-time));
- #endif
activate_array_ongpu(layer.output_cl, m*n*layer.batch, layer.activation);
- #ifdef TIMEIT
- cl_read_array(layer.output_cl, layer.output, m*n*layer.batch);
- #endif
}
-void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem delta_cl)
+void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in, cl_mem delta_cl)
{
int i;
int m = layer.n;
@@ -393,31 +391,26 @@
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;
- gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,i*k*n,k,1,c,0,n);
- }
- //cl_read_array(layer.delta_cl, layer.delta, m*k*layer.batch);
+ 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){
- m = layer.size*layer.size*layer.c;
- k = layer.n;
- n = convolutional_out_height(layer)*
- convolutional_out_width(layer);
+ if(delta_cl){
- for(i = 0; i < layer.batch; ++i){
cl_mem a = layer.filters_cl;
cl_mem b = layer.delta_cl;
cl_mem c = layer.col_image_cl;
- gemm_ongpu_offset(1,0,m,n,k,1,a,0,m,b,i*k*n,n,0,c,i*m*n,n);
- }
+ gemm_ongpu_offset(1,0,n,k,m,1,a,0,n,b,i*k*m,k,0,c,0,k);
- scal_ongpu(layer.batch*layer.h*layer.w*layer.c,0,delta_cl, 1);
- col2im_ongpu(layer.col_image_cl, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl);
+ 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);
+ }
}
}
@@ -425,12 +418,16 @@
{
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)
@@ -439,10 +436,10 @@
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);
- scal_ongpu(size, 1.-layer.learning_rate*layer.decay, layer.filters_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);
+ //pull_convolutional_layer(layer);
}
--
Gitblit v1.10.0