From 5a47c46b39475fc3581b9819f488b977ea1beca3 Mon Sep 17 00:00:00 2001
From: Edmond Yoo <hj3yoo@uwaterloo.ca>
Date: Sun, 16 Sep 2018 03:11:04 +0000
Subject: [PATCH] Moving files from MTGCardDetector
---
src/batchnorm_layer.c | 88 +++++++++++++++++++++++++++++++++++++-------
1 files changed, 74 insertions(+), 14 deletions(-)
diff --git a/src/batchnorm_layer.c b/src/batchnorm_layer.c
index 55bd3a8..3fa129d 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.normDstTensorDesc);
+ cudnnSetTensor4dDescriptor(layer.normDstTensorDesc, 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;
}
@@ -130,10 +136,10 @@
mean_cpu(l.output, l.batch, l.out_c, l.out_h*l.out_w, l.mean);
variance_cpu(l.output, l.mean, l.batch, l.out_c, l.out_h*l.out_w, l.variance);
- scal_cpu(l.out_c, .99, l.rolling_mean, 1);
- axpy_cpu(l.out_c, .01, l.mean, 1, l.rolling_mean, 1);
- scal_cpu(l.out_c, .99, l.rolling_variance, 1);
- axpy_cpu(l.out_c, .01, l.variance, 1, l.rolling_variance, 1);
+ scal_cpu(l.out_c, .9, l.rolling_mean, 1);
+ axpy_cpu(l.out_c, .1, l.mean, 1, l.rolling_mean, 1);
+ scal_cpu(l.out_c, .9, l.rolling_variance, 1);
+ axpy_cpu(l.out_c, .1, l.variance, 1, l.rolling_variance, 1);
copy_cpu(l.outputs*l.batch, l.output, 1, l.x, 1);
normalize_cpu(l.output, l.mean, l.variance, l.batch, l.out_c, l.out_h*l.out_w);
@@ -173,12 +179,30 @@
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);
- if(l.type == CONNECTED){
- l.out_c = l.outputs;
- l.out_h = l.out_w = 1;
- }
+ 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.normDstTensorDesc,
+ l.x_gpu, // input
+ l.normDstTensorDesc,
+ l.output_gpu, // output
+ l.normTensorDesc,
+ l.scales_gpu,
+ l.biases_gpu,
+ .01,
+ l.rolling_mean_gpu, // output (should be FP32)
+ l.rolling_variance_gpu, // output (should be FP32)
+ .00001,
+ l.mean_gpu, // output (should be FP32)
+ l.variance_gpu); // output (should be FP32)
+#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);
@@ -190,15 +214,50 @@
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);
- } else {
+
+ 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);
}
- scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
}
-void backward_batchnorm_layer_gpu(const layer l, network_state state)
+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.normDstTensorDesc,
+ l.x_gpu, // input
+ l.normDstTensorDesc,
+ l.delta_gpu, // input
+ l.normDstTensorDesc,
+ l.x_norm_gpu, // output
+ l.normTensorDesc,
+ l.scales_gpu, // output (should be FP32)
+ l.scale_updates_gpu, // output (should be FP32)
+ l.bias_updates_gpu, // output (should be FP32)
+ .00001,
+ l.mean_gpu, // input (should be FP32)
+ l.variance_gpu); // input (should be FP32)
+ 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);
@@ -206,6 +265,7 @@
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);
- if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, state.delta, 1);
-}
#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