m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #14701 from heavengate/adaptive_pool

Browse Source 
add adaptive pool2d and pool3d
for_weibo
Kaipeng Deng 6 years ago committed by GitHub
parent 0b1c7d838c f16aa394f6
commit dc76e4b0f1
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
14 changed files with 1014 additions and 317 deletions
Show Stats Download Patch File Download Diff File

2
paddle/fluid/API.spec
Unescape View File

@ -77,6 +77,8 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name']
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None)) paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.adaptive_pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
paddle.fluid.layers.adaptive_pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False)) paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))

181
paddle/fluid/operators/math/pooling.cc
View File

File diff suppressed because it is too large Load Diff

309
paddle/fluid/operators/math/pooling.cu
View File

File diff suppressed because it is too large Load Diff

32
paddle/fluid/operators/math/pooling.h
Unescape View File

@ -68,6 +68,18 @@ class AvgPoolGrad {
} }
}; };
/* used for adaptive pool to calculate start and end index of each divided grid
*/
HOSTDEVICE inline int AdaptStartIndex(int ph, int input_size, int output_size) {
return static_cast<int>(
floor(static_cast<double>(ph * input_size) / output_size));
}
HOSTDEVICE inline int AdaptEndIndex(int ph, int input_size, int output_size) {
return static_cast<int>(
ceil(static_cast<double>((ph + 1) * input_size) / output_size));
}
/* /*
* \brief Getting pooling results, and calculating gradient. * \brief Getting pooling results, and calculating gradient.
* *
@ -102,7 +114,7 @@ class Pool2dFunctor {
const std::vector<int>& ksize, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, PoolProcess pool_compute, const std::vector<int>& paddings, PoolProcess pool_compute,
bool exclusive, framework::Tensor* output); bool exclusive, bool adaptive, framework::Tensor* output);
}; };
template <typename DeviceContext, typename PoolProcess, typename T> template <typename DeviceContext, typename PoolProcess, typename T>
@ -114,7 +126,7 @@ class Pool2dGradFunctor {
const std::vector<int>& ksize, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, PoolProcess pool_compute, const std::vector<int>& paddings, PoolProcess pool_compute,
bool exclusive, framework::Tensor* input_grad); bool exclusive, bool adaptive, framework::Tensor* input_grad);
}; };
template <typename DeviceContext, class T> template <typename DeviceContext, class T>
@ -136,7 +148,7 @@ class Pool3dFunctor {
const std::vector<int>& ksize, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, PoolProcess pool_compute, const std::vector<int>& paddings, PoolProcess pool_compute,
bool exclusive, framework::Tensor* output); bool exclusive, bool adaptive, framework::Tensor* output);
}; };
template <typename DeviceContext, typename PoolProcess, typename T> template <typename DeviceContext, typename PoolProcess, typename T>
@ -148,7 +160,7 @@ class Pool3dGradFunctor {
const std::vector<int>& ksize, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, PoolProcess pool_compute, const std::vector<int>& paddings, PoolProcess pool_compute,
bool exclusive, framework::Tensor* input_grad); bool exclusive, bool adaptive, framework::Tensor* input_grad);
}; };
template <typename DeviceContext, class T> template <typename DeviceContext, class T>
@ -176,8 +188,8 @@ class MaxPool2dWithIndexFunctor {
void operator()(const DeviceContext& context, const framework::Tensor& input, void operator()(const DeviceContext& context, const framework::Tensor& input,
const std::vector<int>& ksize, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, framework::Tensor* output, const std::vector<int>& paddings, bool adaptive,
framework::Tensor* mask); framework::Tensor* output, framework::Tensor* mask);
}; };
template <typename DeviceContext, typename T1, typename T2> template <typename DeviceContext, typename T1, typename T2>
@ -187,7 +199,7 @@ class MaxPool2dWithIndexGradFunctor {
const framework::Tensor& output_grad, const framework::Tensor& output_grad,
const framework::Tensor& mask, const std::vector<int>& ksize, const framework::Tensor& mask, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, const std::vector<int>& paddings, bool adaptive,
framework::Tensor* input_grad); framework::Tensor* input_grad);
}; };
@ -197,8 +209,8 @@ class MaxPool3dWithIndexFunctor {
void operator()(const DeviceContext& context, const framework::Tensor& input, void operator()(const DeviceContext& context, const framework::Tensor& input,
const std::vector<int>& ksize, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, framework::Tensor* output, const std::vector<int>& paddings, bool adaptive,
framework::Tensor* mask); framework::Tensor* output, framework::Tensor* mask);
}; };
template <typename DeviceContext, typename T1, typename T2> template <typename DeviceContext, typename T1, typename T2>
@ -208,7 +220,7 @@ class MaxPool3dWithIndexGradFunctor {
const framework::Tensor& output_grad, const framework::Tensor& output_grad,
const framework::Tensor& mask, const std::vector<int>& ksize, const framework::Tensor& mask, const std::vector<int>& ksize,
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, const std::vector<int>& paddings, bool adaptive,
framework::Tensor* input_grad); framework::Tensor* input_grad);
}; };

63
paddle/fluid/operators/pool_op.cc
Unescape View File

@ -52,6 +52,7 @@ void PoolOp::InferShape(framework::InferShapeContext* ctx) const {
std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides"); std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings"); std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
bool ceil_mode = ctx->Attrs().Get<bool>("ceil_mode"); bool ceil_mode = ctx->Attrs().Get<bool>("ceil_mode");
bool adaptive = ctx->Attrs().Get<bool>("adaptive");
PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5, PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
"Pooling intput should be 4-D or 5-D tensor."); "Pooling intput should be 4-D or 5-D tensor.");
@ -72,9 +73,13 @@ void PoolOp::InferShape(framework::InferShapeContext* ctx) const {
"Paddings size and pooling size should be the same."); "Paddings size and pooling size should be the same.");
std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]}); std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
if (adaptive) {
output_shape.insert(output_shape.end(), ksize.begin(), ksize.end());
} else {
for (size_t i = 0; i < ksize.size(); ++i) { for (size_t i = 0; i < ksize.size(); ++i) {
output_shape.push_back(PoolOutputSize(in_x_dims[i + 2], ksize[i], output_shape.push_back(PoolOutputSize(
paddings[i], strides[i], ceil_mode)); in_x_dims[i + 2], ksize[i], paddings[i], strides[i], ceil_mode));
}
} }
ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
ctx->ShareLoD("X", "Out"); ctx->ShareLoD("X", "Out");
@ -186,6 +191,14 @@ void Pool2dOpMaker::Make() {
"averaging calculating, otherwise, include the zero-padding. Note, it " "averaging calculating, otherwise, include the zero-padding. Note, it "
"is only used when pooling_type is avg. The defalut is True.") "is only used when pooling_type is avg. The defalut is True.")
.SetDefault(true); .SetDefault(true);
AddAttr<bool>(
"adaptive",
"(bool, default False) When true, will perform adaptive pooling instead, "
"output shape in H and W dimensions will be same as ksize, input data "
"will be divided into grids specify by ksize averagely and perform "
"pooling in each grid area to get output pooling value.")
.SetDefault(false);
AddAttr<bool>( AddAttr<bool>(
"use_cudnn", "use_cudnn",
"(bool, default false) Only used in cudnn kernel, need install cudnn") "(bool, default false) Only used in cudnn kernel, need install cudnn")
@ -264,6 +277,14 @@ Example:
Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)}
$$ $$
For adaptive = true:
$$
hstart = floor(i * H_{in} / H_{out})
hend = ceil((i + 1) * H_{in} / H_{out})
wstart = floor(j * W_{in} / W_{out})
wend = ceil((j + 1) * W_{in} / W_{out})
Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)}
$$
)DOC"); )DOC");
} }
@ -325,6 +346,13 @@ void Pool3dOpMaker::Make() {
"averaging calculating, otherwise, include the zero-padding. Note, it " "averaging calculating, otherwise, include the zero-padding. Note, it "
"is only used when pooling_type is avg. The defalut is True.") "is only used when pooling_type is avg. The defalut is True.")
.SetDefault(true); .SetDefault(true);
AddAttr<bool>(
"adaptive",
"(bool, default False) When true, will perform adaptive pooling instead, "
"output shape in H and W dimensions will be same as ksize, input data "
"will be divided into grids specify by ksize averagely and perform "
"pooling in each grid area to get output pooling value.")
.SetDefault(false);
AddAttr<bool>( AddAttr<bool>(
"use_cudnn", "use_cudnn",
@ -376,6 +404,37 @@ Example:
H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 \\ H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 \\
W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1 W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1
$$ $$
For exclusive = true:
$$
dstart = i * strides[0] - paddings[0]
dend = dstart + ksize[0]
hstart = j * strides[1] - paddings[1]
hend = hstart + ksize[1]
wstart = k * strides[2] - paddings[2]
wend = wstart + ksize[2]
Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{ksize[0] * ksize[1] * ksize[2]}
$$
For exclusive = false:
$$
dstart = max(0, i * strides[0] - paddings[0])
dend = min(D, dstart + ksize[0])
hstart = max(0, j * strides[1] - paddings[1])
hend = min(H, hstart + ksize[1])
wstart = max(0, k * strides[2] - paddings[2])
wend = min(W, wstart + ksize[2])
Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)}
$$
For adaptive = true:
$$
dstart = floor(i * D_{in} / D_{out})
dend = ceil((i + 1) * D_{in} / D_{out})
hstart = floor(j * H_{in} / H_{out})
hend = ceil((j + 1) * H_{in} / H_{out})
wstart = floor(k * W_{in} / W_{out})
wend = ceil((k + 1) * W_{in} / W_{out})
Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)}
$$
)DOC"); )DOC");
} }

16
paddle/fluid/operators/pool_op.h
Unescape View File

@ -70,6 +70,7 @@ class PoolKernel : public framework::OpKernel<T> {
std::vector<int> strides = context.Attr<std::vector<int>>("strides"); std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings"); std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool exclusive = context.Attr<bool>("exclusive"); bool exclusive = context.Attr<bool>("exclusive");
bool adaptive = context.Attr<bool>("adaptive");
if (context.Attr<bool>("global_pooling")) { if (context.Attr<bool>("global_pooling")) {
for (size_t i = 0; i < ksize.size(); ++i) { for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0; paddings[i] = 0;
@ -85,7 +86,7 @@ class PoolKernel : public framework::OpKernel<T> {
pool2d_forward; pool2d_forward;
paddle::operators::math::MaxPool<T> pool_process; paddle::operators::math::MaxPool<T> pool_process;
pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
true, out); true, false, out);
} else if (pooling_type == "avg") { } else if (pooling_type == "avg") {
paddle::operators::math::Pool2dFunctor< paddle::operators::math::Pool2dFunctor<
@ -93,7 +94,7 @@ class PoolKernel : public framework::OpKernel<T> {
pool2d_forward; pool2d_forward;
paddle::operators::math::AvgPool<T> pool_process; paddle::operators::math::AvgPool<T> pool_process;
pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
exclusive, out); exclusive, adaptive, out);
} }
} break; } break;
case 3: { case 3: {
@ -103,14 +104,14 @@ class PoolKernel : public framework::OpKernel<T> {
pool3d_forward; pool3d_forward;
paddle::operators::math::MaxPool<T> pool_process; paddle::operators::math::MaxPool<T> pool_process;
pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
true, out); true, false, out);
} else if (pooling_type == "avg") { } else if (pooling_type == "avg") {
paddle::operators::math::Pool3dFunctor< paddle::operators::math::Pool3dFunctor<
DeviceContext, paddle::operators::math::AvgPool<T>, T> DeviceContext, paddle::operators::math::AvgPool<T>, T>
pool3d_forward; pool3d_forward;
paddle::operators::math::AvgPool<T> pool_process; paddle::operators::math::AvgPool<T> pool_process;
pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
exclusive, out); exclusive, adaptive, out);
} }
} break; } break;
default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
@ -133,6 +134,7 @@ class PoolGradKernel : public framework::OpKernel<T> {
std::vector<int> strides = context.Attr<std::vector<int>>("strides"); std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings"); std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool exclusive = context.Attr<bool>("exclusive"); bool exclusive = context.Attr<bool>("exclusive");
bool adaptive = context.Attr<bool>("adaptive");
if (context.Attr<bool>("global_pooling")) { if (context.Attr<bool>("global_pooling")) {
for (size_t i = 0; i < ksize.size(); ++i) { for (size_t i = 0; i < ksize.size(); ++i) {
@ -159,7 +161,8 @@ class PoolGradKernel : public framework::OpKernel<T> {
pool2d_backward; pool2d_backward;
paddle::operators::math::AvgPoolGrad<T> pool_process; paddle::operators::math::AvgPoolGrad<T> pool_process;
pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
paddings, pool_process, exclusive, in_x_grad); paddings, pool_process, exclusive, adaptive,
in_x_grad);
} }
} break; } break;
case 3: { case 3: {
@ -174,7 +177,8 @@ class PoolGradKernel : public framework::OpKernel<T> {
pool3d_backward; pool3d_backward;
paddle::operators::math::AvgPoolGrad<T> pool_process; paddle::operators::math::AvgPoolGrad<T> pool_process;
pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
paddings, pool_process, exclusive, in_x_grad); paddings, pool_process, exclusive, adaptive,
in_x_grad);
} }
} break; } break;
default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }

32
paddle/fluid/operators/pool_with_index_op.cc
Unescape View File

@ -40,6 +40,7 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize"); std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize");
std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides"); std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings"); std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
bool adaptive = ctx->Attrs().Get<bool>("adaptive");
PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5, PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
"Pooling intput should be 4-D or 5-D tensor."); "Pooling intput should be 4-D or 5-D tensor.");
@ -60,10 +61,14 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
"Paddings size and pooling size should be the same."); "Paddings size and pooling size should be the same.");
std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]}); std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
if (adaptive) {
output_shape.insert(output_shape.end(), ksize.begin(), ksize.end());
} else {
for (size_t i = 0; i < ksize.size(); ++i) { for (size_t i = 0; i < ksize.size(); ++i) {
output_shape.push_back(MaxPoolOutputSize(in_x_dims[i + 2], ksize[i], output_shape.push_back(MaxPoolOutputSize(in_x_dims[i + 2], ksize[i],
paddings[i], strides[i])); paddings[i], strides[i]));
} }
}
ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
ctx->SetOutputDim("Mask", framework::make_ddim(output_shape)); ctx->SetOutputDim("Mask", framework::make_ddim(output_shape));
} }
@ -133,6 +138,14 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
"(bool, default:false) Whether to use the global pooling. " "(bool, default:false) Whether to use the global pooling. "
"If global_pooling = true, ksize and paddings will be ignored.") "If global_pooling = true, ksize and paddings will be ignored.")
.SetDefault(false); .SetDefault(false);
AddAttr<bool>(
"adaptive",
"(bool, default False) When true, will perform adaptive pooling "
"instead, "
"output shape in H and W dimensions will be same as ksize, input data "
"will be divided into grids specify by ksize averagely and perform "
"pooling in each grid area to get output pooling value.")
.SetDefault(false);
AddAttr<std::vector<int>>("strides", AddAttr<std::vector<int>>("strides",
"(vector<int>, default {1, 1}), strides(height, " "(vector<int>, default {1, 1}), strides(height, "
"width) of pooling operator.") "width) of pooling operator.")
@ -170,6 +183,12 @@ Example:
W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
$$ $$
For adaptive = true:
$$
H_{out} = ksize[0] W_{out} = ksize[1]
$$
)DOC"); )DOC");
} }
}; };
@ -209,6 +228,14 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
"(bool, default false) Whether to use the global pooling. " "(bool, default false) Whether to use the global pooling. "
"If global_pooling = true, ksize and paddings will be ignored.") "If global_pooling = true, ksize and paddings will be ignored.")
.SetDefault(false); .SetDefault(false);
AddAttr<bool>(
"adaptive",
"(bool, default False) When true, will perform adaptive pooling "
"instead, "
"output shape in H and W dimensions will be same as ksize, input data "
"will be divided into grids specify by ksize averagely and perform "
"pooling in each grid area to get output pooling value.")
.SetDefault(false);
AddAttr<std::vector<int>>("strides", AddAttr<std::vector<int>>("strides",
"(vector<int>, default {1,1,1}), strides(depth, " "(vector<int>, default {1,1,1}), strides(depth, "
"height, width) of pooling operator.") "height, width) of pooling operator.")
@ -247,6 +274,11 @@ Example:
W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1 W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1
$$ $$
For adaptive = true:
$$
D_{out} = ksize[0] H_{out} = ksize[1] W_{out} = ksize[2]
$$
)DOC"); )DOC");
} }
}; };

12
paddle/fluid/operators/pool_with_index_op.h
Unescape View File

@ -36,6 +36,7 @@ class MaxPoolWithIndexKernel : public framework::OpKernel<T1> {
std::vector<int> ksize = context.Attr<std::vector<int>>("ksize"); std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
std::vector<int> strides = context.Attr<std::vector<int>>("strides"); std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings"); std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool adaptive = context.Attr<bool>("adaptive");
auto& dev_ctx = context.template device_context<DeviceContext>(); auto& dev_ctx = context.template device_context<DeviceContext>();
if (context.Attr<bool>("global_pooling")) { if (context.Attr<bool>("global_pooling")) {
@ -50,13 +51,15 @@ class MaxPoolWithIndexKernel : public framework::OpKernel<T1> {
paddle::operators::math::MaxPool2dWithIndexFunctor<DeviceContext, T1, paddle::operators::math::MaxPool2dWithIndexFunctor<DeviceContext, T1,
T2> T2>
pool2d_forward; pool2d_forward;
pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, out, mask); pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, adaptive, out,
mask);
} break; } break;
case 3: { case 3: {
paddle::operators::math::MaxPool3dWithIndexFunctor<DeviceContext, T1, paddle::operators::math::MaxPool3dWithIndexFunctor<DeviceContext, T1,
T2> T2>
pool3d_forward; pool3d_forward;
pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, out, mask); pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, adaptive, out,
mask);
} break; } break;
default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
} }
@ -75,6 +78,7 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel<T1> {
std::vector<int> ksize = context.Attr<std::vector<int>>("ksize"); std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
std::vector<int> strides = context.Attr<std::vector<int>>("strides"); std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings"); std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool adaptive = context.Attr<bool>("adaptive");
if (context.Attr<bool>("global_pooling")) { if (context.Attr<bool>("global_pooling")) {
for (size_t i = 0; i < ksize.size(); ++i) { for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0; paddings[i] = 0;
@ -93,14 +97,14 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel<T1> {
T1, T2> T1, T2>
pool2d_backward; pool2d_backward;
pool2d_backward(device_ctx, *out_grad, *mask, ksize, strides, pool2d_backward(device_ctx, *out_grad, *mask, ksize, strides,
paddings, in_x_grad); paddings, adaptive, in_x_grad);
} break; } break;
case 3: { case 3: {
paddle::operators::math::MaxPool3dWithIndexGradFunctor<DeviceContext, paddle::operators::math::MaxPool3dWithIndexGradFunctor<DeviceContext,
T1, T2> T1, T2>
pool3d_backward; pool3d_backward;
pool3d_backward(device_ctx, *out_grad, *mask, ksize, strides, pool3d_backward(device_ctx, *out_grad, *mask, ksize, strides,
paddings, in_x_grad); paddings, adaptive, in_x_grad);
} break; } break;
default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
} }

6
paddle/fluid/operators/spp_op.h
Unescape View File

@ -56,13 +56,13 @@ class SppKernel : public framework::OpKernel<T> {
math::Pool2dFunctor<DeviceContext, math::MaxPool<T>, T> pool_forward; math::Pool2dFunctor<DeviceContext, math::MaxPool<T>, T> pool_forward;
math::MaxPool<T> max_process; math::MaxPool<T> max_process;
pool_forward(context.template device_context<DeviceContext>(), *in_x, pool_forward(context.template device_context<DeviceContext>(), *in_x,
kernel_size, strides, paddings, max_process, true, kernel_size, strides, paddings, max_process, true, false,
&out_level); &out_level);
} else if (pooling_type == "avg") { } else if (pooling_type == "avg") {
math::Pool2dFunctor<DeviceContext, math::AvgPool<T>, T> pool_forward; math::Pool2dFunctor<DeviceContext, math::AvgPool<T>, T> pool_forward;
math::AvgPool<T> avg_process; math::AvgPool<T> avg_process;
pool_forward(context.template device_context<DeviceContext>(), *in_x, pool_forward(context.template device_context<DeviceContext>(), *in_x,
kernel_size, strides, paddings, avg_process, true, kernel_size, strides, paddings, avg_process, true, false,
&out_level); &out_level);
} }
// flatten pooling output shape // flatten pooling output shape
@ -156,7 +156,7 @@ class SppGradKernel : public framework::OpKernel<T> {
math::AvgPoolGrad<T> avg_process; math::AvgPoolGrad<T> avg_process;
pool_backward(context.template device_context<DeviceContext>(), *in_x, pool_backward(context.template device_context<DeviceContext>(), *in_x,
*&out_level, *&outgrad_level, kernel_size, strides, *&out_level, *&outgrad_level, kernel_size, strides,
paddings, avg_process, true, in_x_grad); paddings, avg_process, true, false, in_x_grad);
} }
} }
} }

200
python/paddle/fluid/layers/nn.py
Unescape View File

@ -52,6 +52,8 @@ __all__ = [
'softmax', 'softmax',
'pool2d', 'pool2d',
'pool3d', 'pool3d',
'adaptive_pool2d',
'adaptive_pool3d',
'batch_norm', 'batch_norm',
'beam_search_decode', 'beam_search_decode',
'conv2d_transpose', 'conv2d_transpose',
@ -2500,6 +2502,204 @@ def pool3d(input,
return pool_out return pool_out
@templatedoc(op_type="pool2d")
def adaptive_pool2d(input,
pool_size,
pool_type="max",
require_index=False,
name=None):
"""
${comment}
Args:
input (Variable): The input tensor of pooling operator. The format of
input tensor is NCHW, where N is batch size, C is
the number of channels, H is the height of the
feature, and W is the width of the feature.
pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
it must contain two integers, (pool_size_Height, pool_size_Width).
pool_type: ${pooling_type_comment}
require_index (bool): If true, the index of max pooling point along with outputs.
it cannot be set in average pooling type.
name (str|None): A name for this layer(optional). If set None, the
layer will be named automatically.
Returns:
Variable: The pooling result.
Raises:
ValueError: 'pool_type' is not 'max' nor 'avg'.
ValueError: invalid setting 'require_index' true when 'pool_type' is 'avg'.
ValueError: 'pool_size' should be a list or tuple with length as 2.
Examples:
.. code-block:: python
# suppose input data in shape of [N, C, H, W], `pool_size` is [m, n],
# output shape is [N, C, m, n], adaptive pool divide H and W dimentions
# of input data into m * n grids averagely and performs poolings in each
# grid to get output.
# adaptive average pool performs calculations as follow:
#
# for i in range(m):
# for j in range(n):
# hstart = floor(i * H / m)
# hend = ceil((i + 1) * H / m)
# wstart = floor(i * W / n)
# wend = ceil((i + 1) * W / n)
# output[:, :, i, j] = avg(input[:, :, hstart: hend, wstart: wend])
#
data = fluid.layers.data(
name='data', shape=[3, 32, 32], dtype='float32')
pool_out = fluid.layers.adaptive_pool2d(
input=data,
pool_size=[3, 3],
pool_type='avg')
"""
if pool_type not in ["max", "avg"]:
raise ValueError(
"Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
str(pool_type))
if pool_type == "avg" and require_index:
raise ValueError(
"invalid setting 'require_index' true when 'pool_type' is 'avg'.")
def _is_list_or_tuple_(data):
return (isinstance(data, list) or isinstance(data, tuple))
if not _is_list_or_tuple_(pool_size) or len(pool_size) != 2:
raise ValueError(
"'pool_size' should be a list or tuple with length as 2.")
if pool_type == "max":
l_type = 'max_pool2d_with_index'
else:
l_type = "pool2d"
helper = LayerHelper(l_type, **locals())
dtype = helper.input_dtype()
pool_out = helper.create_variable_for_type_inference(dtype)
outputs = {"Out": pool_out}
if pool_type == "max":
mask = helper.create_variable_for_type_inference(dtype)
outputs["Mask"] = mask
helper.append_op(
type=l_type,
inputs={"X": input},
outputs=outputs,
attrs={
"pooling_type": pool_type,
"ksize": pool_size,
"adaptive": True,
})
return (pool_out, mask) if require_index else pool_out
@templatedoc(op_type="pool3d")
def adaptive_pool3d(input,
pool_size,
pool_type="max",
require_index=False,
name=None):
"""
${comment}
Args:
input (Variable): The input tensor of pooling operator. The format of
input tensor is NCHW, where N is batch size, C is
the number of channels, H is the height of the
feature, and W is the width of the feature.
pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
it must contain two integers, (Depth, Height, Width).
pool_type: ${pooling_type_comment}
require_index (bool): If true, the index of max pooling point along with outputs.
it cannot be set in average pooling type.
name (str|None): A name for this layer(optional). If set None, the
layer will be named automatically.
Returns:
Variable: The pooling result.
Raises:
ValueError: 'pool_type' is not 'max' nor 'avg'.
ValueError: invalid setting 'require_index' true when 'pool_type' is 'avg'.
ValueError: 'pool_size' should be a list or tuple with length as 2.
Examples:
.. code-block:: python
# suppose input data in shape of [N, C, D, H, W], `pool_size` is [l, m, n],
# output shape is [N, C, l, m, n], adaptive pool divide D, H and W dimentions
# of input data into l * m * n grids averagely and performs poolings in each
# grid to get output.
# adaptive average pool performs calculations as follow:
#
# for i in range(l):
# for j in range(m):
# for k in range(n):
# dstart = floor(i * D / l)
# dend = ceil((i + 1) * D / l)
# hstart = floor(j * H / m)
# hend = ceil((j + 1) * H / m)
# wstart = floor(k * W / n)
# wend = ceil((k + 1) * W / n)
# output[:, :, i, j, k] =
# avg(input[:, :, dstart:dend, hstart: hend, wstart: wend])
#
data = fluid.layers.data(
name='data', shape=[3, 32, 32], dtype='float32')
pool_out, mask = fluid.layers.adaptive_pool3d(
input=data,
pool_size=[3, 3],
pool_type='avg')
"""
if pool_type not in ["max", "avg"]:
raise ValueError(
"Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
str(pool_type))
if pool_type == "avg" and require_index:
raise ValueError(
"invalid setting 'require_index' true when 'pool_type' is 'avg'.")
def _is_list_or_tuple_(data):
return (isinstance(data, list) or isinstance(data, tuple))
if not _is_list_or_tuple_(pool_size) or len(pool_size) != 3:
raise ValueError(
"'pool_size' should be a list or tuple with length as 3.")
if pool_type == "max":
l_type = 'max_pool3d_with_index'
else:
l_type = "pool3d"
helper = LayerHelper(l_type, **locals())
dtype = helper.input_dtype()
pool_out = helper.create_variable_for_type_inference(dtype)
outputs = {"Out": pool_out}
if pool_type == "max":
mask = helper.create_variable_for_type_inference(dtype)
outputs["Mask"] = mask
helper.append_op(
type=l_type,
inputs={"X": input},
outputs=outputs,
attrs={
"pooling_type": pool_type,
"ksize": pool_size,
"adaptive": True,
})
return (pool_out, mask) if require_index else pool_out
def batch_norm(input, def batch_norm(input,
act=None, act=None,
is_test=False, is_test=False,

23
python/paddle/fluid/tests/unittests/test_layers.py
Unescape View File

@ -233,6 +233,29 @@ class TestBook(unittest.TestCase):
pool_stride=[1, 2], pool_stride=[1, 2],
pool_padding=(2, 1))) pool_padding=(2, 1)))
def test_adaptive_pool2d(self):
program = Program()
with program_guard(program):
x = layers.data(name='x', shape=[3, 224, 224], dtype='float32')
self.assertIsNotNone(
layers.adaptive_pool2d(
x, [3, 3], pool_type='avg'))
pool, mask = layers.adaptive_pool2d(x, [3, 3], require_index=True)
self.assertIsNotNone(pool)
self.assertIsNotNone(mask)
def test_adaptive_pool3d(self):
program = Program()
with program_guard(program):
x = layers.data(name='x', shape=[3, 244, 224, 224], dtype='float32')
self.assertIsNotNone(
layers.adaptive_pool3d(
x, [3, 3, 3], pool_type='avg'))
pool, mask = layers.adaptive_pool3d(
x, [3, 3, 3], require_index=True)
self.assertIsNotNone(pool)
self.assertIsNotNone(mask)
def test_lstm_unit(self): def test_lstm_unit(self):
program = Program() program = Program()
with program_guard(program): with program_guard(program):

51
python/paddle/fluid/tests/unittests/test_pool2d_op.py
Unescape View File

@ -13,6 +13,7 @@
# limitations under the License. # limitations under the License.
from __future__ import print_function from __future__ import print_function
from __future__ import division
import unittest import unittest
import numpy as np import numpy as np
@ -21,16 +22,28 @@ import paddle.fluid.core as core
from op_test import OpTest from op_test import OpTest
def adaptive_start_index(index, input_size, output_size):
return int(np.floor(index * input_size / output_size))
def adaptive_end_index(index, input_size, output_size):
return int(np.ceil((index + 1) * input_size / output_size))
def max_pool2D_forward_naive(x, def max_pool2D_forward_naive(x,
ksize, ksize,
strides, strides,
paddings, paddings,
global_pool=0, global_pool=0,
ceil_mode=False, ceil_mode=False,
exclusive=True): exclusive=True,
adaptive=False):
N, C, H, W = x.shape N, C, H, W = x.shape
if global_pool == 1: if global_pool == 1:
ksize = [H, W] ksize = [H, W]
if adaptive:
H_out, W_out = ksize
else:
H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1 H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1
) // strides[0] + 1 if ceil_mode else ( ) // strides[0] + 1 if ceil_mode else (
H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H - ksize[0] + 2 * paddings[0]) // strides[0] + 1
@ -40,6 +53,12 @@ def max_pool2D_forward_naive(x,
out = np.zeros((N, C, H_out, W_out)) out = np.zeros((N, C, H_out, W_out))
for i in range(H_out): for i in range(H_out):
for j in range(W_out): for j in range(W_out):
if adaptive:
r_start = adaptive_start_index(i, H, ksize[0])
r_end = adaptive_end_index(i, H, ksize[0])
c_start = adaptive_start_index(j, W, ksize[1])
c_end = adaptive_end_index(j, W, ksize[1])
else:
r_start = np.max((i * strides[0] - paddings[0], 0)) r_start = np.max((i * strides[0] - paddings[0], 0))
r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H))
c_start = np.max((j * strides[1] - paddings[1], 0)) c_start = np.max((j * strides[1] - paddings[1], 0))
@ -56,10 +75,14 @@ def avg_pool2D_forward_naive(x,
paddings, paddings,
global_pool=0, global_pool=0,
ceil_mode=False, ceil_mode=False,
exclusive=True): exclusive=True,
adaptive=False):
N, C, H, W = x.shape N, C, H, W = x.shape
if global_pool == 1: if global_pool == 1:
ksize = [H, W] ksize = [H, W]
if adaptive:
H_out, W_out = ksize
else:
H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1 H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1
) // strides[0] + 1 if ceil_mode else ( ) // strides[0] + 1 if ceil_mode else (
H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H - ksize[0] + 2 * paddings[0]) // strides[0] + 1
@ -69,14 +92,20 @@ def avg_pool2D_forward_naive(x,
out = np.zeros((N, C, H_out, W_out)) out = np.zeros((N, C, H_out, W_out))
for i in range(H_out): for i in range(H_out):
for j in range(W_out): for j in range(W_out):
if adaptive:
r_start = adaptive_start_index(i, H, ksize[0])
r_end = adaptive_end_index(i, H, ksize[0])
c_start = adaptive_start_index(j, W, ksize[1])
c_end = adaptive_end_index(j, W, ksize[1])
else:
r_start = np.max((i * strides[0] - paddings[0], 0)) r_start = np.max((i * strides[0] - paddings[0], 0))
r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H))
c_start = np.max((j * strides[1] - paddings[1], 0)) c_start = np.max((j * strides[1] - paddings[1], 0))
c_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) c_end = np.min((j * strides[1] + ksize[1] - paddings[1], W))
x_masked = x[:, :, r_start:r_end, c_start:c_end] x_masked = x[:, :, r_start:r_end, c_start:c_end]
field_size = ((r_end - r_start) * (c_end - c_start)) if exclusive \ field_size = ((r_end - r_start) * (c_end - c_start)) \
else (ksize[0] * ksize[1]) if (exclusive or adaptive) else (ksize[0] * ksize[1])
out[:, :, i, j] = np.sum(x_masked, axis=(2, 3)) / field_size out[:, :, i, j] = np.sum(x_masked, axis=(2, 3)) / field_size
return out return out
@ -93,12 +122,13 @@ class TestPool2D_Op(OpTest):
self.init_pool_type() self.init_pool_type()
self.init_ceil_mode() self.init_ceil_mode()
self.init_exclusive() self.init_exclusive()
self.init_adaptive()
if self.global_pool: if self.global_pool:
self.paddings = [0 for _ in range(len(self.paddings))] self.paddings = [0 for _ in range(len(self.paddings))]
input = np.random.random(self.shape).astype(self.dtype) input = np.random.random(self.shape).astype(self.dtype)
output = self.pool2D_forward_naive( output = self.pool2D_forward_naive(
input, self.ksize, self.strides, self.paddings, self.global_pool, input, self.ksize, self.strides, self.paddings, self.global_pool,
self.ceil_mode, self.exclusive).astype(self.dtype) self.ceil_mode, self.exclusive, self.adaptive).astype(self.dtype)
self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)} self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)}
self.attrs = { self.attrs = {
@ -112,7 +142,8 @@ class TestPool2D_Op(OpTest):
'ceil_mode': self.ceil_mode, 'ceil_mode': self.ceil_mode,
'data_format': 'data_format':
'AnyLayout', # TODO(dzhwinter) : should be fix latter 'AnyLayout', # TODO(dzhwinter) : should be fix latter
'exclusive': self.exclusive 'exclusive': self.exclusive,
'adaptive': self.adaptive
} }
self.outputs = {'Out': output} self.outputs = {'Out': output}
@ -159,6 +190,9 @@ class TestPool2D_Op(OpTest):
def init_exclusive(self): def init_exclusive(self):
self.exclusive = True self.exclusive = True
def init_adaptive(self):
self.adaptive = False
class TestCase1(TestPool2D_Op): class TestCase1(TestPool2D_Op):
def init_test_case(self): def init_test_case(self):
@ -315,5 +349,10 @@ class TestCUDNNAvgInclude(TestCase2):
self.exclusive = False self.exclusive = False
class TestAvgPoolAdaptive(TestCase1):
def init_adaptive(self):
self.adaptive = True
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()

77
python/paddle/fluid/tests/unittests/test_pool3d_op.py
Unescape View File

@ -13,6 +13,7 @@
# limitations under the License. # limitations under the License.
from __future__ import print_function from __future__ import print_function
from __future__ import division
import unittest import unittest
import numpy as np import numpy as np
@ -21,16 +22,28 @@ import paddle.fluid.core as core
from op_test import OpTest from op_test import OpTest
def adaptive_start_index(index, input_size, output_size):
return int(np.floor(index * input_size / output_size))
def adaptive_end_index(index, input_size, output_size):
return int(np.ceil((index + 1) * input_size / output_size))
def max_pool3D_forward_naive(x, def max_pool3D_forward_naive(x,
ksize, ksize,
strides, strides,
paddings, paddings,
global_pool=0, global_pool=0,
ceil_mode=False, ceil_mode=False,
exclusive=True): exclusive=True,
adaptive=False):
N, C, D, H, W = x.shape N, C, D, H, W = x.shape
if global_pool == 1: if global_pool == 1:
ksize = [D, H, W] ksize = [D, H, W]
if adaptive:
D_out, H_out, W_out = ksize
else:
D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1 D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1
) // strides[0] + 1 if ceil_mode else ( ) // strides[0] + 1 if ceil_mode else (
H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H - ksize[0] + 2 * paddings[0]) // strides[0] + 1
@ -42,14 +55,26 @@ def max_pool3D_forward_naive(x,
W - ksize[2] + 2 * paddings[2]) // strides[2] + 1 W - ksize[2] + 2 * paddings[2]) // strides[2] + 1
out = np.zeros((N, C, D_out, H_out, W_out)) out = np.zeros((N, C, D_out, H_out, W_out))
for k in range(D_out): for k in range(D_out):
if adaptive:
d_start = adaptive_start_index(k, D, ksize[0])
d_end = adaptive_end_index(k, D, ksize[0])
else:
d_start = np.max((k * strides[0] - paddings[0], 0)) d_start = np.max((k * strides[0] - paddings[0], 0))
d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D)) d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D))
for i in range(H_out): for i in range(H_out):
h_start = np.max((i * strides[0] - paddings[0], 0)) if adaptive:
h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) h_start = adaptive_start_index(i, H, ksize[1])
h_end = adaptive_end_index(i, H, ksize[1])
else:
h_start = np.max((i * strides[1] - paddings[1], 0))
h_end = np.min((i * strides[1] + ksize[1] - paddings[1], H))
for j in range(W_out): for j in range(W_out):
w_start = np.max((j * strides[1] - paddings[1], 0)) if adaptive:
w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) w_start = adaptive_start_index(j, W, ksize[2])
w_end = adaptive_end_index(j, W, ksize[2])
else:
w_start = np.max((j * strides[2] - paddings[2], 0))
w_end = np.min((j * strides[2] + ksize[2] - paddings[2], W))
x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end]
out[:, :, k, i, j] = np.max(x_masked, axis=(2, 3, 4)) out[:, :, k, i, j] = np.max(x_masked, axis=(2, 3, 4))
@ -62,10 +87,14 @@ def avg_pool3D_forward_naive(x,
paddings, paddings,
global_pool=0, global_pool=0,
ceil_mode=False, ceil_mode=False,
exclusive=True): exclusive=True,
adaptive=False):
N, C, D, H, W = x.shape N, C, D, H, W = x.shape
if global_pool == 1: if global_pool == 1:
ksize = [D, H, W] ksize = [D, H, W]
if adaptive:
D_out, H_out, W_out = ksize
else:
D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1 D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1
) // strides[0] + 1 if ceil_mode else ( ) // strides[0] + 1 if ceil_mode else (
H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H - ksize[0] + 2 * paddings[0]) // strides[0] + 1
@ -77,18 +106,30 @@ def avg_pool3D_forward_naive(x,
W - ksize[2] + 2 * paddings[2]) // strides[2] + 1 W - ksize[2] + 2 * paddings[2]) // strides[2] + 1
out = np.zeros((N, C, D_out, H_out, W_out)) out = np.zeros((N, C, D_out, H_out, W_out))
for k in range(D_out): for k in range(D_out):
if adaptive:
d_start = adaptive_start_index(k, D, ksize[0])
d_end = adaptive_end_index(k, D, ksize[0])
else:
d_start = np.max((k * strides[0] - paddings[0], 0)) d_start = np.max((k * strides[0] - paddings[0], 0))
d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D)) d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D))
for i in range(H_out): for i in range(H_out):
h_start = np.max((i * strides[0] - paddings[0], 0)) if adaptive:
h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) h_start = adaptive_start_index(i, H, ksize[1])
h_end = adaptive_end_index(i, H, ksize[1])
else:
h_start = np.max((i * strides[1] - paddings[1], 0))
h_end = np.min((i * strides[1] + ksize[1] - paddings[1], H))
for j in range(W_out): for j in range(W_out):
w_start = np.max((j * strides[1] - paddings[1], 0)) if adaptive:
w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) w_start = adaptive_start_index(j, W, ksize[2])
w_end = adaptive_end_index(j, W, ksize[2])
else:
w_start = np.max((j * strides[2] - paddings[2], 0))
w_end = np.min((j * strides[2] + ksize[2] - paddings[2], W))
x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end]
field_size = (d_end - d_start) * (h_end - h_start) * (w_end - w_start) \ field_size = (d_end - d_start) * (h_end - h_start) * (w_end - w_start) \
if exclusive else ksize[0] * ksize[1] * ksize[2] if (exclusive or adaptive) else ksize[0] * ksize[1] * ksize[2]
out[:, :, k, i, j] = np.sum(x_masked, axis=(2, 3, out[:, :, k, i, j] = np.sum(x_masked, axis=(2, 3,
4)) / field_size 4)) / field_size
return out return out
@ -105,13 +146,14 @@ class TestPool3d_Op(OpTest):
self.init_pool_type() self.init_pool_type()
self.init_ceil_mode() self.init_ceil_mode()
self.init_exclusive() self.init_exclusive()
self.init_adaptive()
if self.global_pool: if self.global_pool:
self.paddings = [0 for _ in range(len(self.paddings))] self.paddings = [0 for _ in range(len(self.paddings))]
input = np.random.random(self.shape).astype(self.dtype) input = np.random.random(self.shape).astype(self.dtype)
output = self.pool3D_forward_naive( output = self.pool3D_forward_naive(
input, self.ksize, self.strides, self.paddings, self.global_pool, input, self.ksize, self.strides, self.paddings, self.global_pool,
self.ceil_mode, self.exclusive).astype(self.dtype) self.ceil_mode, self.exclusive, self.adaptive).astype(self.dtype)
self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)} self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)}
self.attrs = { self.attrs = {
@ -124,7 +166,8 @@ class TestPool3d_Op(OpTest):
'ceil_mode': self.ceil_mode, 'ceil_mode': self.ceil_mode,
'data_format': 'data_format':
'AnyLayout', # TODO(dzhwinter) : should be fix latter 'AnyLayout', # TODO(dzhwinter) : should be fix latter
'exclusive': self.exclusive 'exclusive': self.exclusive,
'adaptive': self.adaptive
} }
self.outputs = {'Out': output} self.outputs = {'Out': output}
@ -171,6 +214,9 @@ class TestPool3d_Op(OpTest):
def init_exclusive(self): def init_exclusive(self):
self.exclusive = True self.exclusive = True
def init_adaptive(self):
self.adaptive = False
class TestCase1(TestPool3d_Op): class TestCase1(TestPool3d_Op):
def init_test_case(self): def init_test_case(self):
@ -353,5 +399,10 @@ class TestCUDNNAvgInclude(TestCUDNNCase3):
self.exclusive = False self.exclusive = False
class TestAvgPoolAdaptive(TestCase1):
def init_adaptive(self):
self.adaptive = True
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()

73
python/paddle/fluid/tests/unittests/test_pool_max_op.py
Unescape View File

@ -13,33 +13,62 @@
# limitations under the License. # limitations under the License.
from __future__ import print_function from __future__ import print_function
from __future__ import division
import unittest import unittest
import numpy as np import numpy as np
from op_test import OpTest from op_test import OpTest
def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=False): def adaptive_start_index(index, input_size, output_size):
return int(np.floor(index * input_size / output_size))
def adaptive_end_index(index, input_size, output_size):
return int(np.ceil((index + 1) * input_size / output_size))
def max_pool3D_forward_naive(x,
ksize,
strides,
paddings,
global_pool=False,
adaptive=False):
N, C, D, H, W = x.shape N, C, D, H, W = x.shape
if global_pool: if global_pool:
ksize = [D, H, W] ksize = [D, H, W]
paddings = [0, 0, 0] paddings = [0, 0, 0]
if adaptive:
D_out, H_out, W_out = ksize
else:
D_out = (D - ksize[0] + 2 * paddings[0]) // strides[0] + 1 D_out = (D - ksize[0] + 2 * paddings[0]) // strides[0] + 1
H_out = (H - ksize[1] + 2 * paddings[1]) // strides[1] + 1 H_out = (H - ksize[1] + 2 * paddings[1]) // strides[1] + 1
W_out = (W - ksize[2] + 2 * paddings[2]) // strides[2] + 1 W_out = (W - ksize[2] + 2 * paddings[2]) // strides[2] + 1
out = np.zeros((N, C, D_out, H_out, W_out)) out = np.zeros((N, C, D_out, H_out, W_out))
mask = np.zeros((N, C, D_out, H_out, W_out)) mask = np.zeros((N, C, D_out, H_out, W_out))
for k in range(D_out): for k in range(D_out):
if adaptive:
d_start = adaptive_start_index(k, D, ksize[0])
d_end = adaptive_end_index(k, D, ksize[0])
else:
d_start = np.max((k * strides[0] - paddings[0], 0)) d_start = np.max((k * strides[0] - paddings[0], 0))
d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D)) d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D))
for i in range(H_out): for i in range(H_out):
h_start = np.max((i * strides[0] - paddings[0], 0)) if adaptive:
h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) h_start = adaptive_start_index(i, H, ksize[1])
h_end = adaptive_end_index(i, H, ksize[1])
else:
h_start = np.max((i * strides[1] - paddings[1], 0))
h_end = np.min((i * strides[1] + ksize[1] - paddings[1], H))
for j in range(W_out): for j in range(W_out):
w_start = np.max((j * strides[1] - paddings[1], 0)) if adaptive:
w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) w_start = adaptive_start_index(j, W, ksize[2])
w_end = adaptive_end_index(j, W, ksize[2])
else:
w_start = np.max((j * strides[2] - paddings[2], 0))
w_end = np.min((j * strides[2] + ksize[2] - paddings[2], W))
x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end]
out[:, :, k, i, j] = np.max(x_masked, axis=(2, 3, 4)) out[:, :, k, i, j] = np.max(x_masked, axis=(2, 3, 4))
@ -58,19 +87,33 @@ def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=False):
return out, mask return out, mask
def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=False): def max_pool2D_forward_naive(x,
ksize,
strides,
paddings,
global_pool=False,
adaptive=False):
N, C, H, W = x.shape N, C, H, W = x.shape
if global_pool: if global_pool:
ksize = [H, W] ksize = [H, W]
paddings = [0, 0] paddings = [0, 0]
if adaptive:
H_out, W_out = ksize
else:
H_out = (H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H_out = (H - ksize[0] + 2 * paddings[0]) // strides[0] + 1
W_out = (W - ksize[1] + 2 * paddings[1]) // strides[1] + 1 W_out = (W - ksize[1] + 2 * paddings[1]) // strides[1] + 1
out = np.zeros((N, C, H_out, W_out)) out = np.zeros((N, C, H_out, W_out))
mask = np.zeros((N, C, H_out, W_out)) mask = np.zeros((N, C, H_out, W_out))
for i in range(H_out): for i in range(H_out):
for j in range(W_out): for j in range(W_out):
if adaptive:
r_start = adaptive_start_index(i, H, ksize[0])
r_end = adaptive_end_index(i, H, ksize[0])
c_start = adaptive_start_index(j, W, ksize[1])
c_end = adaptive_end_index(j, W, ksize[1])
else:
r_start = np.max((i * strides[0] - paddings[0], 0)) r_start = np.max((i * strides[0] - paddings[0], 0))
r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H))
c_start = np.max((j * strides[1] - paddings[1], 0)) c_start = np.max((j * strides[1] - paddings[1], 0))
@ -95,10 +138,12 @@ class TestMaxPoolWithIndex_Op(OpTest):
def setUp(self): def setUp(self):
self.init_test_case() self.init_test_case()
self.init_global() self.init_global()
self.init_adaptive()
input = np.random.random(self.shape).astype("float32") input = np.random.random(self.shape).astype("float32")
output, mask = self.pool_forward_naive(input, self.ksize, self.strides, output, mask = self.pool_forward_naive(input, self.ksize, self.strides,
self.paddings, self.global_pool) self.paddings, self.global_pool,
self.adaptive)
output = output.astype("float32") output = output.astype("float32")
mask = mask.astype("int32") mask = mask.astype("int32")
@ -107,6 +152,7 @@ class TestMaxPoolWithIndex_Op(OpTest):
'paddings': self.paddings, 'paddings': self.paddings,
'ksize': self.ksize, 'ksize': self.ksize,
'global_pooling': self.global_pool, 'global_pooling': self.global_pool,
'adaptive': self.adaptive,
} }
self.inputs = {'X': input} self.inputs = {'X': input}
@ -129,6 +175,9 @@ class TestMaxPoolWithIndex_Op(OpTest):
def init_global(self): def init_global(self):
self.global_pool = False self.global_pool = False
def init_adaptive(self):
self.adaptive = False
class TestCase1(TestMaxPoolWithIndex_Op): class TestCase1(TestMaxPoolWithIndex_Op):
def init_global(self): def init_global(self):
@ -190,5 +239,15 @@ class TestCase7(TestCase6):
self.global_pool = False self.global_pool = False
class TestCastAdaptive2d(TestCase6):
def init_adaptive(self):
self.adaptive = True
class TestCastAdaptive3d(TestMaxPoolWithIndex_Op):
def init_adaptive(self):
self.adaptive = True
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()

Write Preview
Loading…
Cancel
Save