From c62b4f35aa2c59d7db0fd177affeed14b1ba4bcb Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 08 Sep 2016 07:04:39 +0000
Subject: [PATCH] adding coco models
---
src/convolutional_layer.c | 384 ++++++++++++++++++++++++++++--------------------------
1 files changed, 198 insertions(+), 186 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 5575aac..ad2d8a5 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -8,6 +8,15 @@
#include <stdio.h>
#include <time.h>
+#ifdef AI2
+#include "xnor_layer.h"
+#endif
+
+#ifndef AI2
+#define AI2 0
+void forward_xnor_layer(layer l, network_state state);
+#endif
+
void swap_binary(convolutional_layer *l)
{
float *swap = l->filters;
@@ -21,24 +30,6 @@
#endif
}
-void binarize_filters2(float *filters, int n, int size, char *binary, float *scales)
-{
- int i, k, f;
- for(f = 0; f < n; ++f){
- float mean = 0;
- for(i = 0; i < size; ++i){
- mean += fabs(filters[f*size + i]);
- }
- mean = mean / size;
- scales[f] = mean;
- for(i = 0; i < size/8; ++i){
- binary[f*size + i] = (filters[f*size + i] > 0) ? 1 : 0;
- for(k = 0; k < 8; ++k){
- }
- }
- }
-}
-
void binarize_filters(float *filters, int n, int size, float *binary)
{
int i, f;
@@ -54,20 +45,37 @@
}
}
+void binarize_cpu(float *input, int n, float *binary)
+{
+ int i;
+ for(i = 0; i < n; ++i){
+ binary[i] = (input[i] > 0) ? 1 : -1;
+ }
+}
+
+void binarize_input(float *input, int n, int size, float *binary)
+{
+ int i, s;
+ for(s = 0; s < size; ++s){
+ float mean = 0;
+ for(i = 0; i < n; ++i){
+ mean += fabs(input[i*size + s]);
+ }
+ mean = mean / n;
+ for(i = 0; i < n; ++i){
+ binary[i*size + s] = (input[i*size + s] > 0) ? mean : -mean;
+ }
+ }
+}
+
int convolutional_out_height(convolutional_layer l)
{
- int h = l.h;
- if (!l.pad) h -= l.size;
- else h -= 1;
- return h/l.stride + 1;
+ return (l.h + 2*l.pad - l.size) / l.stride + 1;
}
int convolutional_out_width(convolutional_layer l)
{
- int w = l.w;
- if (!l.pad) w -= l.size;
- else w -= 1;
- return w/l.stride + 1;
+ return (l.w + 2*l.pad - l.size) / l.stride + 1;
}
image get_convolutional_image(convolutional_layer l)
@@ -89,40 +97,81 @@
}
size_t get_workspace_size(layer l){
- #ifdef CUDNN
- size_t most = 0;
- size_t s = 0;
- cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
- l.srcTensorDesc,
- l.filterDesc,
- l.convDesc,
- l.dstTensorDesc,
- l.fw_algo,
- &s);
- if (s > most) most = s;
- cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
- l.srcTensorDesc,
- l.ddstTensorDesc,
- l.convDesc,
- l.dfilterDesc,
- l.bf_algo,
- &s);
- if (s > most) most = s;
- cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
- l.filterDesc,
- l.ddstTensorDesc,
- l.convDesc,
- l.dsrcTensorDesc,
- l.bd_algo,
- &s);
- if (s > most) most = s;
- return most;
- #else
- return (size_t)l.out_h*l.out_w*l.size*l.size*l.c*sizeof(float);
+#ifdef CUDNN
+ if(gpu_index >= 0){
+ size_t most = 0;
+ size_t s = 0;
+ cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.filterDesc,
+ l.convDesc,
+ l.dstTensorDesc,
+ l.fw_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dfilterDesc,
+ l.bf_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
+ l.filterDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dsrcTensorDesc,
+ l.bd_algo,
+ &s);
+ if (s > most) most = s;
+ return most;
+ }
#endif
+ return (size_t)l.out_h*l.out_w*l.size*l.size*l.c*sizeof(float);
}
-convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize, int binary, int xnor)
+#ifdef GPU
+#ifdef CUDNN
+void cudnn_convolutional_setup(layer *l)
+{
+ cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->dfilterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->filterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
+ cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->filterDesc,
+ l->convDesc,
+ l->dstTensorDesc,
+ CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
+ 0,
+ &l->fw_algo);
+ cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
+ l->filterDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dsrcTensorDesc,
+ CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
+ 0,
+ &l->bd_algo);
+ cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dfilterDesc,
+ CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
+ 0,
+ &l->bf_algo);
+}
+#endif
+#endif
+
+convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int padding, ACTIVATION activation, int batch_normalize, int binary, int xnor)
{
int i;
convolutional_layer l = {0};
@@ -133,10 +182,11 @@
l.c = c;
l.n = n;
l.binary = binary;
+ l.xnor = xnor;
l.batch = batch;
l.stride = stride;
l.size = size;
- l.pad = pad;
+ l.pad = padding;
l.batch_normalize = batch_normalize;
l.filters = calloc(c*n*size*size, sizeof(float));
@@ -164,6 +214,10 @@
l.cfilters = calloc(c*n*size*size, sizeof(char));
l.scales = calloc(n, sizeof(float));
}
+ if(xnor){
+ l.binary_filters = calloc(c*n*size*size, sizeof(float));
+ l.binary_input = calloc(l.inputs*l.batch, sizeof(float));
+ }
if(batch_normalize){
l.scales = calloc(n, sizeof(float));
@@ -180,82 +234,51 @@
}
#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);
+ if(gpu_index >= 0){
+ 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.biases_gpu = cuda_make_array(l.biases, n);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
- l.scales_gpu = cuda_make_array(l.scales, n);
- l.scale_updates_gpu = cuda_make_array(l.scale_updates, n);
+ l.scales_gpu = cuda_make_array(l.scales, n);
+ l.scale_updates_gpu = cuda_make_array(l.scale_updates, n);
- 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);
+ 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);
- if(binary){
- l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
- }
- if(xnor){
- l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
- l.binary_input_gpu = cuda_make_array(0, l.inputs*l.batch);
- }
- l.xnor = xnor;
+ if(binary){
+ l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
+ }
+ if(xnor){
+ l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
+ l.binary_input_gpu = cuda_make_array(0, l.inputs*l.batch);
+ }
- if(batch_normalize){
- l.mean_gpu = cuda_make_array(l.mean, n);
- l.variance_gpu = cuda_make_array(l.variance, n);
+ if(batch_normalize){
+ l.mean_gpu = cuda_make_array(l.mean, n);
+ l.variance_gpu = cuda_make_array(l.variance, n);
- l.rolling_mean_gpu = cuda_make_array(l.mean, n);
- l.rolling_variance_gpu = cuda_make_array(l.variance, n);
+ l.rolling_mean_gpu = cuda_make_array(l.mean, n);
+ l.rolling_variance_gpu = cuda_make_array(l.variance, n);
- l.mean_delta_gpu = cuda_make_array(l.mean, n);
- l.variance_delta_gpu = cuda_make_array(l.variance, n);
+ l.mean_delta_gpu = cuda_make_array(l.mean, n);
+ l.variance_delta_gpu = cuda_make_array(l.variance, n);
- l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
- l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
- }
+ l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ }
#ifdef CUDNN
- cudnnCreateTensorDescriptor(&l.srcTensorDesc);
- cudnnCreateTensorDescriptor(&l.dstTensorDesc);
- cudnnCreateFilterDescriptor(&l.filterDesc);
- cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
- cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
- cudnnCreateFilterDescriptor(&l.dfilterDesc);
- cudnnCreateConvolutionDescriptor(&l.convDesc);
- cudnnSetTensor4dDescriptor(l.dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.c, l.h, l.w);
- cudnnSetTensor4dDescriptor(l.ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
- cudnnSetFilter4dDescriptor(l.dfilterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l.n, l.c, l.size, l.size);
-
- cudnnSetTensor4dDescriptor(l.srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.c, l.h, l.w);
- cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
- cudnnSetFilter4dDescriptor(l.filterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l.n, l.c, l.size, l.size);
- int padding = l.pad ? l.size/2 : 0;
- cudnnSetConvolution2dDescriptor(l.convDesc, padding, padding, l.stride, l.stride, 1, 1, CUDNN_CROSS_CORRELATION);
- cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
- l.srcTensorDesc,
- l.filterDesc,
- l.convDesc,
- l.dstTensorDesc,
- CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
- 0,
- &l.fw_algo);
- cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
- l.filterDesc,
- l.ddstTensorDesc,
- l.convDesc,
- l.dsrcTensorDesc,
- CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
- 0,
- &l.bd_algo);
- cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
- l.srcTensorDesc,
- l.ddstTensorDesc,
- l.convDesc,
- l.dfilterDesc,
- CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
- 0,
- &l.bf_algo);
+ cudnnCreateTensorDescriptor(&l.srcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.filterDesc);
+ cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.dfilterDesc);
+ cudnnCreateConvolutionDescriptor(&l.convDesc);
+ cudnn_convolutional_setup(&l);
#endif
+ }
#endif
l.workspace_size = get_workspace_size(l);
l.activation = activation;
@@ -274,6 +297,9 @@
l.filters[i*l.c*l.size*l.size + j] *= scale;
}
l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
}
}
@@ -325,41 +351,9 @@
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);
- #ifdef CUDNN
- cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
- cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
- cudnnSetFilter4dDescriptor(l->dfilterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
-
- cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
- cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
- cudnnSetFilter4dDescriptor(l->filterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
- int padding = l->pad ? l->size/2 : 0;
- cudnnSetConvolution2dDescriptor(l->convDesc, padding, padding, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
- cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
- l->srcTensorDesc,
- l->filterDesc,
- l->convDesc,
- l->dstTensorDesc,
- CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
- 0,
- &l->fw_algo);
- cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
- l->filterDesc,
- l->ddstTensorDesc,
- l->convDesc,
- l->dsrcTensorDesc,
- CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
- 0,
- &l->bd_algo);
- cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
- l->srcTensorDesc,
- l->ddstTensorDesc,
- l->convDesc,
- l->dfilterDesc,
- CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
- 0,
- &l->bf_algo);
- #endif
+#ifdef CUDNN
+ cudnn_convolutional_setup(l);
+#endif
#endif
l->workspace_size = get_workspace_size(*l);
}
@@ -404,7 +398,9 @@
int out_w = convolutional_out_width(l);
int i;
+
fill_cpu(l.outputs*l.batch, 0, l.output, 1);
+
/*
if(l.binary){
binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
@@ -413,42 +409,57 @@
}
*/
- if(l.binary){
- int m = l.n;
- int k = l.size*l.size*l.c;
- int n = out_h*out_w;
+ /*
+ if(l.binary){
+ int m = l.n;
+ int k = l.size*l.size*l.c;
+ int n = out_h*out_w;
- char *a = l.cfilters;
- float *b = state.workspace;
- float *c = l.output;
+ char *a = l.cfilters;
+ float *b = state.workspace;
+ 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_bin(m,n,k,1,a,k,b,n,c,n);
- c += n*m;
- state.input += l.c*l.h*l.w;
- }
- scale_bias(l.output, l.scales, l.batch, l.n, out_h*out_w);
- add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
- activate_array(l.output, m*n*l.batch, l.activation);
- return;
+ 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_bin(m,n,k,1,a,k,b,n,c,n);
+ c += n*m;
+ state.input += l.c*l.h*l.w;
+ }
+ scale_bias(l.output, l.scales, l.batch, l.n, out_h*out_w);
+ add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
+ activate_array(l.output, m*n*l.batch, l.activation);
+ return;
+ }
+ */
+
+ if(l.xnor){
+ binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
+ swap_binary(&l);
+ binarize_cpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input);
+ state.input = l.binary_input;
}
int m = l.n;
int k = l.size*l.size*l.c;
int n = out_h*out_w;
- float *a = l.filters;
- float *b = state.workspace;
- float *c = l.output;
+ if (l.xnor && l.c%32 == 0 && AI2) {
+ forward_xnor_layer(l, state);
+ printf("xnor\n");
+ } else {
- 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;
+ float *a = l.filters;
+ float *b = state.workspace;
+ 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;
+ }
}
if(l.batch_normalize){
@@ -457,6 +468,7 @@
add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
activate_array(l.output, m*n*l.batch, l.activation);
+ if(l.binary || l.xnor) swap_binary(&l);
}
void backward_convolutional_layer(convolutional_layer l, network_state state)
--
Gitblit v1.10.0