From aebe937710ced03d03f73ab23f410f29685655c1 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 11 Aug 2016 18:54:24 +0000
Subject: [PATCH] what do you even write here?
---
src/convolutional_layer.c | 252 ++++++++++++++++++++++++--------------------------
1 files changed, 121 insertions(+), 131 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index af867e5..006dc4c 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -14,6 +14,7 @@
#ifndef AI2
#define AI2 0
+void forward_xnor_layer(layer l, network_state state);
#endif
void swap_binary(convolutional_layer *l)
@@ -44,6 +45,14 @@
}
}
+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;
@@ -95,38 +104,80 @@
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
+ 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);
-#endif
}
+#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);
+ 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
+#endif
+
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)
{
int i;
@@ -190,81 +241,51 @@
}
#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);
- }
+ 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;
@@ -283,6 +304,9 @@
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;
}
}
@@ -335,39 +359,7 @@
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);
+ cudnn_convolutional_setup(l);
#endif
#endif
l->workspace_size = get_workspace_size(*l);
@@ -448,12 +440,10 @@
}
*/
- if(l.xnor && (l.c%32 != 0 || !AI2)){
+ if(l.xnor){
binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
swap_binary(&l);
- for(i = 0; i < l.batch; ++i){
- binarize_input(state.input + i*l.inputs, l.c, l.h*l.w, l.binary_input + i*l.inputs);
- }
+ binarize_cpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input);
state.input = l.binary_input;
}
--
Gitblit v1.10.0