Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

Support SyncBatchNorm5D #14542

Merged
merged 27 commits into from
Apr 2, 2019
Merged

Support SyncBatchNorm5D #14542

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
27 commits
Select commit Hold shift + click to select a range
df2f264
support SyncBatchNorm5D
wkcn Mar 27, 2019
c314241
fix
wkcn Mar 27, 2019
46bc426
update testcase and reformat code
wkcn Mar 28, 2019
4729860
retrigger CI
wkcn Mar 28, 2019
e7bd3bb
update test case
wkcn Mar 28, 2019
02415f9
test
wkcn Mar 28, 2019
380e750
Retrigger CI
wkcn Mar 28, 2019
4653390
disable cudnn for batchnorm
wkcn Mar 28, 2019
e11d076
Merge branch 'support_5d_sync_batchnorm' of github.com:wkcn/incubator…
wkcn Mar 28, 2019
c8ad1a8
fix BatchNorm(cudnn)
wkcn Mar 28, 2019
c22b400
fix build
wkcn Mar 28, 2019
f316e91
Remove a testcase
wkcn Mar 28, 2019
ea470be
Update sync_batch_norm-inl.h
wkcn Mar 28, 2019
d4a118d
update unittest
wkcn Mar 29, 2019
f126fb1
merge master
wkcn Mar 29, 2019
d1a7787
update unittest
wkcn Mar 29, 2019
2a3ba52
update test
wkcn Mar 29, 2019
6d6142f
fix test
wkcn Mar 29, 2019
659f1db
change atol and rtol
wkcn Mar 29, 2019
36f930d
BN(cudnn) 5d
wkcn Mar 29, 2019
9ec1f51
update test
wkcn Mar 29, 2019
904f5bd
test
wkcn Mar 29, 2019
e238132
Testing
wkcn Mar 29, 2019
1ba2c84
Update batch_norm.cu
wkcn Mar 30, 2019
784b77d
test cudnnoff
wkcn Mar 30, 2019
767efeb
Update test_operator.py
wkcn Mar 30, 2019
993e8e3
update BN! : )
wkcn Mar 30, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 19 additions & 11 deletions src/operator/contrib/sync_batch_norm-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -275,14 +275,18 @@ class SyncBatchNorm : public Operator {
static_cast<real_t>(in_data[syncbatchnorm::kData].shape_.Size());
Tensor<xpu, 4> data;
Tensor<xpu, 4> out;
if (in_data[syncbatchnorm::kData].ndim() == 2) {
if (in_data[syncbatchnorm::kData].ndim() == 4) {
zhreshold marked this conversation as resolved.
Show resolved Hide resolved
data = in_data[syncbatchnorm::kData].get<xpu, 4, real_t>(s);
out = out_data[syncbatchnorm::kOut].get<xpu, 4, real_t>(s);
} else {
index_t num_channels = in_data[syncbatchnorm::kData].ndim() > 1 ?
in_data[syncbatchnorm::kData].shape_[1] : 1;
index_t spatial_size = in_data[syncbatchnorm::kData].Size() / (
in_data[syncbatchnorm::kData].shape_[0] * num_channels);
Shape<4> dshape = Shape4(in_data[syncbatchnorm::kData].shape_[0],
in_data[syncbatchnorm::kData].shape_[1], 1, 1);
num_channels, 1, spatial_size);
data = in_data[syncbatchnorm::kData].get_with_shape<xpu, 4, real_t>(dshape, s);
out = out_data[syncbatchnorm::kOut].get_with_shape<xpu, 4, real_t>(dshape, s);
} else {
data = in_data[syncbatchnorm::kData].get<xpu, 4, real_t>(s);
out = out_data[syncbatchnorm::kOut].get<xpu, 4, real_t>(s);
}
Tensor<xpu, 1> slope = in_data[syncbatchnorm::kGamma].get<xpu, 1, real_t>(s);
Tensor<xpu, 1> bias = in_data[syncbatchnorm::kBeta].get<xpu, 1, real_t>(s);
Expand Down Expand Up @@ -354,16 +358,20 @@ class SyncBatchNorm : public Operator {
Tensor<xpu, 4> data, grad, grad_in;
const real_t scale = static_cast<real_t>(out_grad[syncbatchnorm::kOut].shape_[1]) /
static_cast<real_t>(out_grad[syncbatchnorm::kOut].shape_.Size());
if (in_data[syncbatchnorm::kData].ndim() == 2) {
if (in_data[syncbatchnorm::kData].ndim() == 4) {
data = in_data[syncbatchnorm::kData].get<xpu, 4, real_t>(s);
grad = out_grad[syncbatchnorm::kOut].get<xpu, 4, real_t>(s);
grad_in = in_grad[syncbatchnorm::kData].get<xpu, 4, real_t>(s);
} else {
index_t num_channels = out_grad[syncbatchnorm::kOut].ndim() > 1 ?
out_grad[syncbatchnorm::kOut].shape_[1] : 1;
index_t spatial_size = out_grad[syncbatchnorm::kOut].Size() / (
out_grad[syncbatchnorm::kOut].shape_[0] * num_channels);
Shape<4> dshape = Shape4(out_grad[syncbatchnorm::kOut].shape_[0],
out_grad[syncbatchnorm::kOut].shape_[1], 1, 1);
num_channels, 1, spatial_size);
data = in_data[syncbatchnorm::kData].get_with_shape<xpu, 4, real_t>(dshape, s);
grad = out_grad[syncbatchnorm::kOut].get_with_shape<xpu, 4, real_t>(dshape, s);
grad_in = in_grad[syncbatchnorm::kData].get_with_shape<xpu, 4, real_t>(dshape, s);
} else {
data = in_data[syncbatchnorm::kData].get<xpu, 4, real_t>(s);
grad = out_grad[syncbatchnorm::kOut].get<xpu, 4, real_t>(s);
grad_in = in_grad[syncbatchnorm::kData].get<xpu, 4, real_t>(s);
}

Tensor<xpu, 1> mean = out_data[syncbatchnorm::kMean].get<xpu, 1, real_t>(s);
Expand Down
Loading