From 3459c47bfa02d23d300df9a44ba3ec9c958850fe Mon Sep 17 00:00:00 2001
From: Philip Kahn <tigerhawkvok@gmail.com>
Date: Fri, 04 May 2018 21:00:11 +0000
Subject: [PATCH] Also replace root version to Python3
---
src/batchnorm_layer.c | 83 ++++++++++-------------------------------
1 files changed, 21 insertions(+), 62 deletions(-)
diff --git a/src/batchnorm_layer.c b/src/batchnorm_layer.c
index 0151582..d35d9d2 100644
--- a/src/batchnorm_layer.c
+++ b/src/batchnorm_layer.c
@@ -54,8 +54,8 @@
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);
+ 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
@@ -176,47 +176,6 @@
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);
- if(l.type == CONNECTED){
- l.out_c = l.outputs;
- l.out_h = l.out_w = 1;
- }
- if (state.train) {
- 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);
- } 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);
-}
-
-void backward_batchnorm_layer_gpu(const layer l, network_state state)
-{
- 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);
- 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)
{
@@ -230,19 +189,19 @@
CUDNN_BATCHNORM_SPATIAL,
&one,
&zero,
- l.dstTensorDesc,
- l.x_gpu,
- l.dstTensorDesc,
- l.output_gpu,
+ 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,
- l.rolling_variance_gpu,
+ l.rolling_mean_gpu, // output (should be FP32)
+ l.rolling_variance_gpu, // output (should be FP32)
.00001,
- l.mean_gpu,
- l.variance_gpu);
+ 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);
@@ -283,19 +242,19 @@
&zero,
&one,
&one,
- l.dstTensorDesc,
- l.x_gpu,
- l.dstTensorDesc,
- l.delta_gpu,
- l.dstTensorDesc,
- l.x_norm_gpu,
+ l.normDstTensorDesc,
+ l.x_gpu, // input
+ l.normDstTensorDesc,
+ l.delta_gpu, // input
+ l.normDstTensorDesc,
+ l.x_norm_gpu, // output
l.normTensorDesc,
- l.scales_gpu,
- l.scale_updates_gpu,
- l.bias_updates_gpu,
+ 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,
- l.variance_gpu);
+ 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);
--
Gitblit v1.10.0