You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by jx...@apache.org on 2018/02/09 01:53:30 UTC
[incubator-mxnet] branch master updated: Check padding size for
global pooling (#9730)
This is an automated email from the ASF dual-hosted git repository.
jxie pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/master by this push:
new 3e93e6e Check padding size for global pooling (#9730)
3e93e6e is described below
commit 3e93e6e7ec7872efc22095f7aad791c317916972
Author: Tao Lv <ta...@intel.com>
AuthorDate: Fri Feb 9 09:53:16 2018 +0800
Check padding size for global pooling (#9730)
* Check padding size for global pooling
* Check padding size for global pooing in cudnn impl
* Check padding size for global pooling in pooling_v1
* add test case for global pooling
* fix compiling issue in CI
* Check padding size for global pooling in mkl2017 impl
* Fix indent
---
src/operator/mkl/mkl_pooling-inl.h | 1 +
src/operator/nn/cudnn/cudnn_pooling-inl.h | 8 +--
src/operator/nn/pooling-inl.h | 16 +++++-
src/operator/pooling_v1-inl.h | 30 +++++++---
tests/python/gpu/test_operator_gpu.py | 92 +++++++++++++++++++++++++++++++
5 files changed, 133 insertions(+), 14 deletions(-)
diff --git a/src/operator/mkl/mkl_pooling-inl.h b/src/operator/mkl/mkl_pooling-inl.h
index 5662a61..e842aab 100644
--- a/src/operator/mkl/mkl_pooling-inl.h
+++ b/src/operator/mkl/mkl_pooling-inl.h
@@ -86,6 +86,7 @@ class MKLPoolingOp : public Operator {
pad_w_ = param_.pad[1];
if (global_pooling_) {
stride_h_ = stride_w_ = 1;
+ pad_h_ = pad_w_ = 0;
} else {
stride_h_ = param_.stride[0];
stride_w_ = param_.stride[1];
diff --git a/src/operator/nn/cudnn/cudnn_pooling-inl.h b/src/operator/nn/cudnn/cudnn_pooling-inl.h
index 104ed85..2d85d2f 100644
--- a/src/operator/nn/cudnn/cudnn_pooling-inl.h
+++ b/src/operator/nn/cudnn/cudnn_pooling-inl.h
@@ -213,8 +213,8 @@ class CuDNNPoolingOp : public Operator {
nan_prop_,
param_.global_pool ? dshape[2] : param_.kernel[0],
param_.global_pool ? dshape[3] : param_.kernel[1],
- param_.pad[0],
- param_.pad[1],
+ param_.global_pool ? 0 : param_.pad[0],
+ param_.global_pool ? 0 : param_.pad[1],
param_.global_pool ? 1 : param_.stride[0],
param_.global_pool ? 1 :param_.stride[1]));
#else
@@ -222,8 +222,8 @@ class CuDNNPoolingOp : public Operator {
mode_,
param_.global_pool ? dshape[2] : param_.kernel[0],
param_.global_pool ? dshape[3] : param_.kernel[1],
- param_.pad[0],
- param_.pad[1],
+ param_.global_pool ? 0 : param_.pad[0],
+ param_.global_ppol ? 0 : param_.pad[1],
param_.global_pool ? 1 : param_.stride[0],
param_.global_pool ? 1 : param_.stride[1]));
#endif
diff --git a/src/operator/nn/pooling-inl.h b/src/operator/nn/pooling-inl.h
index a32aaa2..a1db983 100644
--- a/src/operator/nn/pooling-inl.h
+++ b/src/operator/nn/pooling-inl.h
@@ -97,6 +97,12 @@ class PoolingOp : public Operator {
CHECK_EQ(out_data.size(), 1U);
Stream<xpu> *s = ctx.get_stream<xpu>();
const TShape& ishape = in_data[pool_enum::kData].shape_;
+ TShape padding = param_.pad;
+ if (param_.global_pool) {
+ for (index_t i = 0; i < padding.ndim(); i++) {
+ padding[i] = 0;
+ }
+ }
pool(s, in_data[pool_enum::kData].dptr<DType>(),
in_data[pool_enum::kData].shape_,
@@ -104,7 +110,7 @@ class PoolingOp : public Operator {
param_.global_pool?
TShape(ishape.data()+ishape.ndim()-param_.kernel.ndim(), ishape.data()+ishape.ndim())
: param_.kernel,
- param_.pad,
+ padding,
param_.global_pool? TShape(param_.kernel.ndim()) : param_.stride,
param_.pool_type,
req[pool_enum::kOut],
@@ -126,6 +132,12 @@ class PoolingOp : public Operator {
CHECK_EQ(in_grad.size(), 1U);
Stream<xpu> *s = ctx.get_stream<xpu>();
const TShape& ishape = in_data[pool_enum::kData].shape_;
+ TShape padding = param_.pad;
+ if (param_.global_pool) {
+ for (index_t i = 0; i < padding.ndim(); i++) {
+ padding[i] = 0;
+ }
+ }
unpool(s, out_grad[pool_enum::kOut].dptr<DType>(),
in_data[pool_enum::kData].dptr<DType>(),
@@ -135,7 +147,7 @@ class PoolingOp : public Operator {
param_.global_pool?
TShape(ishape.data()+ishape.ndim()-param_.kernel.ndim(), ishape.data()+ishape.ndim())
: param_.kernel,
- param_.pad,
+ padding,
param_.global_pool? TShape(param_.kernel.ndim()) : param_.stride,
param_.pool_type,
req[pool_enum::kData],
diff --git a/src/operator/pooling_v1-inl.h b/src/operator/pooling_v1-inl.h
index 7070c0d..4934dbe 100644
--- a/src/operator/pooling_v1-inl.h
+++ b/src/operator/pooling_v1-inl.h
@@ -102,6 +102,13 @@ class PoolingV1Op : public Operator {
if (param_.kernel.ndim() == 3) {
LOG(FATAL) << "3D kernel not implemented";
}
+
+ // reset padding size for global pooling
+ TShape padding = param_.pad;
+ if (param_.global_pool) {
+ padding[0] = padding[1] = 0;
+ }
+
Tensor<xpu, 4, DType> data = in_data[pool_v1_enum::kData].get<xpu, 4, DType>(s);
Tensor<xpu, 4, DType> out = out_data[pool_v1_enum::kOut].get<xpu, 4, DType>(s);
mshadow::Shape<2> out_shape = Shape2(out.shape_[2], out.shape_[3]);
@@ -109,7 +116,7 @@ class PoolingV1Op : public Operator {
|| param_.pool_type == pool_v1_enum::kSumPooling) {
Assign(out,
req[pool_v1_enum::kOut],
- pool<Reducer>(pad(data, param_.pad[0], param_.pad[1]),
+ pool<Reducer>(pad(data, padding[0], padding[1]),
out_shape,
param_.global_pool ? data.shape_[2] : param_.kernel[0],
param_.global_pool ? data.shape_[3] : param_.kernel[1],
@@ -121,7 +128,7 @@ class PoolingV1Op : public Operator {
scalar<DType>(1.0f / (param_.global_pool ?
data.shape_[2] * data.shape_[3] :
param_.kernel[0] * param_.kernel[1])) * \
- pool<Reducer>(pad(data, param_.pad[0], param_.pad[1]),
+ pool<Reducer>(pad(data, padding[0], padding[1]),
out_shape,
param_.global_pool ? data.shape_[2] : param_.kernel[0],
param_.global_pool ? data.shape_[3] : param_.kernel[1],
@@ -148,6 +155,13 @@ class PoolingV1Op : public Operator {
if (param_.kernel.ndim() == 3) {
LOG(FATAL) << "3D kernel not implemented";
}
+
+ // reset padding size for global pooling
+ TShape padding = param_.pad;
+ if (param_.global_pool) {
+ padding[0] = padding[1] = 0;
+ }
+
Stream<xpu> *s = ctx.get_stream<xpu>();
Tensor<xpu, 4, DType> grad = out_grad[pool_v1_enum::kOut].get<xpu, 4, DType>(s);
Tensor<xpu, 4, DType> data = in_data[pool_v1_enum::kData].get<xpu, 4, DType>(s);
@@ -159,7 +173,7 @@ class PoolingV1Op : public Operator {
if (param_.pool_type == pool_v1_enum::kMaxPooling
|| param_.pool_type == pool_v1_enum::kSumPooling) {
Assign(input_grad, req[pool_v1_enum::kData],
- crop(unpool<Reducer>(pad(data, param_.pad[0], param_.pad[1]),
+ crop(unpool<Reducer>(pad(data, padding[0], padding[1]),
pad(output_data, 0, 0),
pad(grad, 0, 0),
param_.global_pool ? in_shape[0] : param_.kernel[0],
@@ -167,14 +181,14 @@ class PoolingV1Op : public Operator {
param_.global_pool ? 1 : param_.stride[0],
param_.global_pool ? 1 : param_.stride[1]),
in_shape,
- param_.pad[0],
- param_.pad[1]));
+ padding[0],
+ padding[1]));
} else if (param_.pool_type == pool_v1_enum::kAvgPooling) {
Assign(input_grad, req[pool_v1_enum::kData],
scalar<DType>(1.0f / (param_.global_pool ?
data.shape_[2] * data.shape_[3] :
param_.kernel[0] * param_.kernel[1])) * \
- crop(unpool<Reducer>(pad(data, param_.pad[0], param_.pad[1]),
+ crop(unpool<Reducer>(pad(data, padding[0], padding[1]),
pad(output_data, 0, 0),
pad(grad, 0, 0),
param_.global_pool ? in_shape[0] : param_.kernel[0],
@@ -182,8 +196,8 @@ class PoolingV1Op : public Operator {
param_.global_pool ? 1 : param_.stride[0],
param_.global_pool ? 1 : param_.stride[1]),
in_shape,
- param_.pad[0],
- param_.pad[1]));
+ padding[0],
+ padding[1]));
}
}
diff --git a/tests/python/gpu/test_operator_gpu.py b/tests/python/gpu/test_operator_gpu.py
index 55bb30c..f9f00c4 100644
--- a/tests/python/gpu/test_operator_gpu.py
+++ b/tests/python/gpu/test_operator_gpu.py
@@ -875,6 +875,98 @@ def test_pooling_versions():
test_3d_pooling('sum')
+def test_global_pooling():
+ def test_1d_pooling(pool_type):
+ data = (2, 3, 20)
+ kernel = (4,)
+ pad = (2,)
+ stride = (2,)
+
+ ctx_list = []
+ sym_list = []
+
+ pooling_convention = 'valid'
+
+ 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=pooling_convention, global_pool=True, name='pool'))
+
+ ctx_list.append({'ctx': mx.cpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
+ sym_list.append(mx.sym.Pooling(kernel=kernel, pool_type=pool_type,
+ pooling_convention=pooling_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=pooling_convention, global_pool=True, cudnn_off=False, 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, pool_type=pool_type,
+ pooling_convention=pooling_convention, global_pool=True, cudnn_off=False, 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=pooling_convention, global_pool=True, cudnn_off=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, pool_type=pool_type,
+ pooling_convention=pooling_convention, global_pool=True, cudnn_off=True, name='pool'))
+
+ check_consistency(sym_list, ctx_list)
+
+ def test_2d_pooling(pool_type):
+ data = (2, 3, 20, 20)
+ kernel = (4, 4)
+ pad = (2, 2)
+ stride = (2, 2)
+
+ ctx_list = []
+ sym_list = []
+
+ pooling_convention = 'valid'
+
+ ctx_list.append({'ctx': mx.cpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
+ sym_list.append(mx.sym.Pooling_v1(kernel=kernel, pad=pad, stride=stride, pool_type=pool_type,
+ pooling_convention=pooling_convention, global_pool=True, name='pool'))
+
+ ctx_list.append({'ctx': mx.cpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
+ sym_list.append(mx.sym.Pooling_v1(kernel=kernel, pool_type=pool_type,
+ pooling_convention=pooling_convention, global_pool=True, name='pool'))
+
+ 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=pooling_convention, global_pool=True, name='pool'))
+
+ ctx_list.append({'ctx': mx.cpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
+ sym_list.append(mx.sym.Pooling(kernel=kernel, pool_type=pool_type,
+ pooling_convention=pooling_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=pooling_convention, global_pool=True, cudnn_off=False, 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, pool_type=pool_type,
+ pooling_convention=pooling_convention, global_pool=True, cudnn_off=False, 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=pooling_convention, global_pool=True, cudnn_off=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, pool_type=pool_type,
+ pooling_convention=pooling_convention, global_pool=True, cudnn_off=True, name='pool'))
+
+ check_consistency(sym_list, ctx_list)
+
+ test_1d_pooling('max')
+ test_1d_pooling('avg')
+ test_1d_pooling('sum')
+
+ test_2d_pooling('max')
+ test_2d_pooling('avg')
+ test_2d_pooling('sum')
+
+
def test_upsampling_with_type():
sym = mx.sym.UpSampling(scale=2, num_filter=2, name='up', sample_type='nearest', num_args=1)
ctx_list = [{'ctx': mx.gpu(0), 'up_arg0': (2, 2, 2, 10), 'type_dict': {'up_arg0': np.float64}},
--
To stop receiving notification emails like this one, please contact
jxie@apache.org.