From 08b757a0bf76efe8c76b453063a1bb19315bcaa6 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 14 Jan 2015 20:18:57 +0000
Subject: [PATCH] Stable, needs to be way faster
---
src/convolutional_layer.c | 478 +++++++++++++++++++++++++++++++++++++++++++++++------------
1 files changed, 377 insertions(+), 101 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index d4aff73..fc5cb0e 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,150 +1,426 @@
#include "convolutional_layer.h"
+#include "utils.h"
+#include "mini_blas.h"
#include <stdio.h>
+#include <time.h>
+
+int convolutional_out_height(convolutional_layer layer)
+{
+ 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)
+{
+ 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)
{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.n;
- return double_to_image(h,w,c,layer.output);
+ 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);
}
image get_convolutional_delta(convolutional_layer layer)
{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.n;
- return double_to_image(h,w,c,layer.delta);
+ 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);
}
-convolutional_layer *make_convolutional_layer(int h, int w, int c, int n, int size, int stride, ACTIVATION activator)
+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)
{
- printf("Convolutional Layer: %d x %d x %d image, %d filters\n", h,w,c,n);
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;
layer->n = n;
+ layer->batch = batch;
layer->stride = stride;
- layer->kernels = calloc(n, sizeof(image));
- layer->kernel_updates = calloc(n, sizeof(image));
- layer->biases = calloc(n, sizeof(double));
- layer->bias_updates = calloc(n, sizeof(double));
- for(i = 0; i < n; ++i){
- layer->biases[i] = .005;
- layer->kernels[i] = make_random_kernel(size, c);
- layer->kernel_updates[i] = make_random_kernel(size, c);
- }
- layer->output = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * n, sizeof(double));
- layer->delta = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * n, sizeof(double));
- layer->upsampled = make_image(h,w,n);
+ layer->size = size;
+ layer->pad = pad;
- if(activator == SIGMOID){
- layer->activation = sigmoid_activation;
- layer->gradient = sigmoid_gradient;
- }else if(activator == RELU){
- layer->activation = relu_activation;
- layer->gradient = relu_gradient;
- }else if(activator == IDENTITY){
- layer->activation = identity_activation;
- layer->gradient = identity_gradient;
+ 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./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] = scale;
}
+ int out_h = convolutional_out_height(*layer);
+ int out_w = convolutional_out_width(*layer);
+
+ 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);
+
return layer;
}
-void forward_convolutional_layer(const convolutional_layer layer, double *in)
+void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c)
{
- image input = double_to_image(layer.h, layer.w, layer.c, in);
- image output = get_convolutional_image(layer);
- int i,j;
- for(i = 0; i < layer.n; ++i){
- convolve(input, layer.kernels[i], layer.stride, i, output);
- }
- for(i = 0; i < output.c; ++i){
- for(j = 0; j < output.h*output.w; ++j){
- int index = i*output.h*output.w + j;
- output.data[index] += layer.biases[i];
- output.data[index] = layer.activation(output.data[index]);
- }
- }
+ layer->h = h;
+ layer->w = w;
+ layer->c = c;
+ int out_h = convolutional_out_height(*layer);
+ int out_w = convolutional_out_width(*layer);
+
+ 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));
}
-void backward_convolutional_layer(convolutional_layer layer, double *input, double *delta)
+void bias_output(const convolutional_layer layer)
{
- int i;
-
- image in_image = double_to_image(layer.h, layer.w, layer.c, input);
- image in_delta = double_to_image(layer.h, layer.w, layer.c, delta);
- image out_delta = get_convolutional_delta(layer);
- zero_image(in_delta);
-
- for(i = 0; i < layer.n; ++i){
- back_convolve(in_delta, layer.kernels[i], layer.stride, i, out_delta);
- }
- for(i = 0; i < layer.h*layer.w*layer.c; ++i){
- in_delta.data[i] *= layer.gradient(in_image.data[i]);
- }
-}
-
-/*
-void backpropagate_convolutional_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){
+ 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){
- two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, input, j);
+ for(j = 0; j < out_h*out_w; ++j){
+ layer.output[(b*layer.n + i)*out_h*out_w + j] = layer.biases[i];
+ }
}
}
-
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
}
-*/
-void learn_convolutional_layer(convolutional_layer layer, double *input)
+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;
- image in_image = double_to_image(layer.h, layer.w, layer.c, input);
- image out_delta = get_convolutional_delta(layer);
- for(i = 0; i < layer.n; ++i){
- kernel_update(in_image, layer.kernel_updates[i], layer.stride, i, out_delta);
- layer.bias_updates[i] += avg_image_layer(out_delta, 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 update_convolutional_layer(convolutional_layer layer, double step)
+void learn_bias_convolutional_layer(convolutional_layer layer)
{
- return;
- int i,j;
- for(i = 0; i < layer.n; ++i){
- layer.biases[i] += step*layer.bias_updates[i];
- layer.bias_updates[i] = 0;
- int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c;
- for(j = 0; j < pixels; ++j){
- layer.kernels[i].data[j] += step*layer.kernel_updates[i].data[j];
+ 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);
}
- zero_image(layer.kernel_updates[i]);
}
}
-void visualize_convolutional_layer(convolutional_layer layer)
+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, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.n, layer.momentum, layer.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);
+}
+
+
+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);
+ }
+ }
+ }
+ }
+ }
+ return filters;
+}
+
+image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters)
+{
+ image *single_filters = weighted_sum_filters(layer, 0);
+ show_images(single_filters, layer.n, window);
+
+ image delta = get_convolutional_image(layer);
+ image dc = collapse_image_layers(delta, 1);
char buff[256];
- //image vis = make_image(layer.n*layer.size, layer.size*layer.kernels[0].c, 3);
- for(i = 0; i < layer.n; ++i){
- image k = layer.kernels[i];
- sprintf(buff, "Kernel %d", i);
- if(k.c <= 3) show_image(k, buff);
- else show_image_layers(k, buff);
+ sprintf(buff, "%s: Output", window);
+ //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