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This repository has been archived by the owner on Nov 17, 2023.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

[MXNET-33] Enhance mkldnn pooling to support full convention #11047

Merged
merged 18 commits into from
Nov 17, 2018
Merged

[MXNET-33] Enhance mkldnn pooling to support full convention #11047

Show file tree
Hide file tree
Changes from 11 commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
5d4cb9f
fix mkldnn pooling to support full convention
TaoLv May 24, 2018
093214f
backward with full convention
TaoLv May 24, 2018
62d1685
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv May 25, 2018
76a7d55
fix
TaoLv May 25, 2018
75c2fc4
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv May 26, 2018
e64d9cd
add pooling test for full convention
TaoLv May 26, 2018
f476b0b
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv May 29, 2018
c272a24
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv May 30, 2018
c747a96
add function for computing padding size
TaoLv May 30, 2018
a1f27ba
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv May 31, 2018
71d565a
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv Jun 4, 2018
9ab696b
Merge branch 'master' of /~https://github.com/apache/incubator-mxnet in…
TaoLv Jul 17, 2018
cffdebd
fix unit test
TaoLv Jul 17, 2018
992ad86
only support max-pooling
TaoLv Jul 17, 2018
5d81c78
Merge remote-tracking branch 'origin/master' into fix-pooling-full
TaoLv Nov 8, 2018
9f96e23
fix pooling bwd
TaoLv Nov 8, 2018
62fe457
address review comment
TaoLv Nov 16, 2018
c588514
Merge remote-tracking branch 'origin/master' into fix-pooling-full
TaoLv Nov 16, 2018
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
17 changes: 1 addition & 16 deletions src/operator/nn/mkldnn/mkldnn_pooling-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,23 +87,8 @@ inline bool SupportMKLDNNPooling(const PoolingParam &param) {
inline bool SupportMKLDNNPooling(const PoolingParam &param,
const TShape &dshape) {
bool ret = SupportMKLDNNPooling(param);
if (!ret)
return false;

if (param.pooling_convention == pool_enum::kValid)
return true;
else
return false;

// need to support pooling convention full
// https://issues.apache.org/jira/browse/MXNET-33
#if 0
if (((dshape[2] + 2 * param.pad[0] - param.kernel[0]) % param.stride[0] == 0) &&
((dshape[3] + 2 * param.pad[1] - param.kernel[1]) % param.stride[1] == 0))
return true;
else
return false;
#endif
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we ignore the shape completely? even for the case of kValid?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, I think so. Previously, mkldnn pooling operator only supports pooling_convention=kValid and it's no need to check shape for kValid. But if we want to support kFull, we need adjust padding size to get correct output shape.

return ret;
}

inline bool MKLDNNRequireWorkspace(const PoolingParam &param) {
Expand Down
40 changes: 32 additions & 8 deletions src/operator/nn/mkldnn/mkldnn_pooling.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,6 +129,14 @@ mkldnn::algorithm GetMKLDNNPoolAlgo(const PoolingParam &param) {
}
}

static inline int GetPaddingSizeFull(int x, int padl, int padr, int k, int s) {
if ((x + padl + padr - k) % s != 0) {
return (padr + s - ((x + padl + padr - k) % s));
} else {
return padr;
}
}

mkldnn::pooling_forward::primitive_desc GetPoolingFwd(const PoolingParam &param,
const bool is_train,
const memory::desc &data_md,
Expand All @@ -150,11 +158,17 @@ mkldnn::pooling_forward::primitive_desc GetPoolingFwd(const PoolingParam &param,
int pad_l_ = param.pad[1], pad_r_ = param.pad[1];
int stride_h_ = param.stride[0], stride_w_ = param.stride[1];

if (param.pooling_convention == pool_enum::kFull) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is it possible to write up a macro/function for the same check of kFull in the different place in this file?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

pad_b_ = GetPaddingSizeFull(data_md.data.dims[2], pad_t_, pad_b_, kernel_h_, stride_h_);
pad_r_ = GetPaddingSizeFull(data_md.data.dims[3], pad_l_, pad_r_, kernel_w_, stride_w_);
}

const mkldnn::engine engine = CpuEngine::Get()->get_engine();
if (param.global_pool) {
pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
stride_h_ = stride_w_ = 1;
}

if (pad_t_ != 0 || pad_l_ != 0) {
CHECK(param.pool_type == pool_enum::kAvgPooling ||
param.pool_type == pool_enum::kMaxPooling)
Expand All @@ -163,7 +177,6 @@ mkldnn::pooling_forward::primitive_desc GetPoolingFwd(const PoolingParam &param,
CHECK_LT(pad_t_, kernel_h_);
}


const mkldnn::algorithm alg = GetMKLDNNPoolAlgo(param);
mkldnn::prop_kind kind = mkldnn::prop_kind::forward_scoring;
if (is_train && alg != algorithm::pooling_avg) {
Expand Down Expand Up @@ -223,17 +236,22 @@ MKLDNNPoolingFwd &GetPoolingFwd(const PoolingParam &param,
int pad_l_ = param.pad[1], pad_r_ = param.pad[1];
int stride_h_ = param.stride[0], stride_w_ = param.stride[1];

if (param.pooling_convention == pool_enum::kFull) {
pad_b_ = GetPaddingSizeFull(data_md.data.dims[2], pad_t_, pad_b_, kernel_h_, stride_h_);
pad_r_ = GetPaddingSizeFull(data_md.data.dims[3], pad_l_, pad_r_, kernel_w_, stride_w_);
}

if (param.global_pool) {
pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
stride_h_ = stride_w_ = 1;
pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
stride_h_ = stride_w_ = 1;
}

if (pad_t_ != 0 || pad_l_ != 0) {
CHECK(param.pool_type == pool_enum::kAvgPooling ||
param.pool_type == pool_enum::kMaxPooling)
<< "Padding implemented only for average and max pooling.";
CHECK_LT(pad_l_, kernel_w_);
CHECK_LT(pad_t_, kernel_h_);
CHECK(param.pool_type == pool_enum::kAvgPooling ||
param.pool_type == pool_enum::kMaxPooling)
<< "Padding implemented only for average and max pooling.";
CHECK_LT(pad_l_, kernel_w_);
CHECK_LT(pad_t_, kernel_h_);
}

const mkldnn::algorithm alg = GetMKLDNNPoolAlgo(param);
Expand Down Expand Up @@ -299,6 +317,12 @@ void MKLDNNPoolingGradCompute(const OpContext &ctx, const PoolingParam &param,
int pad_t_ = param.pad[0], pad_b_ = param.pad[0];
int pad_l_ = param.pad[1], pad_r_ = param.pad[1];
int stride_h_ = param.stride[0], stride_w_ = param.stride[1];

if (param.pooling_convention == pool_enum::kFull) {
pad_b_ = GetPaddingSizeFull(data_md.data.dims[2], pad_t_, pad_b_, kernel_h_, stride_h_);
pad_r_ = GetPaddingSizeFull(data_md.data.dims[3], pad_l_, pad_r_, kernel_w_, stride_w_);
}

if (param.global_pool) {
pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
stride_h_ = stride_w_ = 1;
Expand Down
29 changes: 29 additions & 0 deletions tests/python/gpu/test_operator_gpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -920,6 +920,35 @@ def test_3d_pooling(pool_type, p_value=2):
test_3d_pooling('lp', p_value=3)


@with_seed()
def test_pooling_full_2d():
def test_pooling_full_2d_type(pool_type):
data = (2, 2, 10, 10)
kernel = (4, 5)
pad = (1, 2)
stride = (3, 4)

convention = 'full'
ctx_list = []
sym_list = []

# o_h = ceil((10 + 1 + 1 - 4) / 3) + 1 = 4
# o_w = ceil((10 + 2 + 2 - 5) / 4) + 1 = 4
ctx_list.append({'ctx': mx.cpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
sym_list.append(mx.sym.Pooling(kernel=kernel, pad=pad, stride=stride, pool_type=pool_type,
pooling_convention=convention, global_pool=True, name='pool'))

ctx_list.append({'ctx': mx.gpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
sym_list.append(mx.sym.Pooling(kernel=kernel, pad=pad, stride=stride, pool_type=pool_type,
pooling_convention=convention, global_pool=True, name='pool'))
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

it doesn't seem you test your code. Once global_pool is true, all paddings are set to 0.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The symbol defined on line 938 and 942 is exactly the same.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please check your test code with and without your fix to make sure that your test can trigger the bug.


check_consistency(sym_list, ctx_list)

test_pooling_full_2d_type('max')
test_pooling_full_2d_type('avg')
test_pooling_full_2d_type('sum')


@with_seed()
def test_global_pooling():
def test_1d_pooling(pool_type, p_value=2):
Expand Down