From 537d135feba179636e9bbfe296e078d51f59914c Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Mon, 19 Mar 2018 23:16:51 +0000
Subject: [PATCH] Improve training performance - batch-norm using cuDNN.
---
src/batchnorm_layer.c | 103 +++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 102 insertions(+), 1 deletions(-)
diff --git a/src/batchnorm_layer.c b/src/batchnorm_layer.c
index b53548b..0151582 100644
--- a/src/batchnorm_layer.c
+++ b/src/batchnorm_layer.c
@@ -52,6 +52,12 @@
layer.x_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
layer.x_norm_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
+#ifdef CUDNN
+ cudnnCreateTensorDescriptor(&layer.normTensorDesc);
+ cudnnCreateTensorDescriptor(&layer.dstTensorDesc);
+ cudnnSetTensor4dDescriptor(layer.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, layer.batch, layer.out_c, layer.out_h, layer.out_w);
+ cudnnSetTensor4dDescriptor(layer.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, layer.out_c, 1, 1);
+#endif
#endif
return layer;
}
@@ -170,7 +176,7 @@
cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.c);
cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.c);
}
-
+/*
void forward_batchnorm_layer_gpu(layer l, network_state state)
{
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
@@ -209,3 +215,98 @@
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, state.delta, 1);
}
#endif
+*/
+
+
+void forward_batchnorm_layer_gpu(layer l, network_state state)
+{
+ if (l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
+ copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
+ if (state.train) {
+#ifdef CUDNN
+ float one = 1;
+ float zero = 0;
+ cudnnBatchNormalizationForwardTraining(cudnn_handle(),
+ CUDNN_BATCHNORM_SPATIAL,
+ &one,
+ &zero,
+ l.dstTensorDesc,
+ l.x_gpu,
+ l.dstTensorDesc,
+ l.output_gpu,
+ l.normTensorDesc,
+ l.scales_gpu,
+ l.biases_gpu,
+ .01,
+ l.rolling_mean_gpu,
+ l.rolling_variance_gpu,
+ .00001,
+ l.mean_gpu,
+ l.variance_gpu);
+#else
+ fast_mean_gpu(l.output_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.mean_gpu);
+ fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.variance_gpu);
+
+ scal_ongpu(l.out_c, .99, l.rolling_mean_gpu, 1);
+ axpy_ongpu(l.out_c, .01, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
+ scal_ongpu(l.out_c, .99, l.rolling_variance_gpu, 1);
+ axpy_ongpu(l.out_c, .01, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
+
+ copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
+ normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
+ copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
+
+ scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
+ add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.out_c, l.out_w*l.out_h);
+#endif
+ }
+ else {
+ normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
+ scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
+ add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.out_c, l.out_w*l.out_h);
+ }
+
+}
+
+void backward_batchnorm_layer_gpu(layer l, network_state state)
+{
+ if (!state.train) {
+ l.mean_gpu = l.rolling_mean_gpu;
+ l.variance_gpu = l.rolling_variance_gpu;
+ }
+#ifdef CUDNN
+ float one = 1;
+ float zero = 0;
+ cudnnBatchNormalizationBackward(cudnn_handle(),
+ CUDNN_BATCHNORM_SPATIAL,
+ &one,
+ &zero,
+ &one,
+ &one,
+ l.dstTensorDesc,
+ l.x_gpu,
+ l.dstTensorDesc,
+ l.delta_gpu,
+ l.dstTensorDesc,
+ l.x_norm_gpu,
+ l.normTensorDesc,
+ l.scales_gpu,
+ l.scale_updates_gpu,
+ l.bias_updates_gpu,
+ .00001,
+ l.mean_gpu,
+ l.variance_gpu);
+ copy_ongpu(l.outputs*l.batch, l.x_norm_gpu, 1, l.delta_gpu, 1);
+#else
+ backward_bias_gpu(l.bias_updates_gpu, l.delta_gpu, l.batch, l.out_c, l.out_w*l.out_h);
+ backward_scale_gpu(l.x_norm_gpu, l.delta_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.scale_updates_gpu);
+
+ scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
+
+ fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.mean_delta_gpu);
+ fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.variance_delta_gpu);
+ normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.delta_gpu);
+#endif
+ if (l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, state.delta, 1);
+}
+#endif
\ No newline at end of file
--
Gitblit v1.10.0