feat(mgb/opr): add megbrain adaptive pooling opr

GitOrigin-RevId: 82833f41d9

dnn/scripts/opr_param_defs.py

@@ -179,6 +179,11 @@ pdef('Axis').add_fields('int32', 'axis', 0)
  add_enum_alias('Format', 'ConvolutionV0')
  )
 
+(pdef('AdaptivePooling').
+ add_enum_alias('Mode', 'Pooling').
+ add_enum_alias('Format', 'ConvolutionV0')
+ )
+
 (pdef('LRN',
       'see ImageNet Classification with Deep Convolutional Neural Networks for'
       ' meaning of the fields').

src/opr/impl/dnn/adaptive_pooling.cpp

@@ -0,0 +1,148 @@
+/**
+ * \file src/opr/impl/dnn/adaptive_pooling.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#include "megbrain/opr/dnn/adaptive_pooling.h"
+#include "../internal/megdnn_opr_wrapper.inl"
+#include "megbrain/graph/grad_impl.h"
+#include "megbrain/opr/utility.h"
+
+#include "megdnn/opr_param_defs.h"
+#include "megdnn/oprs/nn.h"
+
+using namespace mgb;
+using namespace opr;
+
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(AdaptivePoolingForward);
+AdaptivePoolingForward::AdaptivePoolingForward(VarNode* src, VarNode* out_shape,
+                                               const Param& param,
+                                               const OperatorNodeConfig& config)
+        : Super(OperatorNodeBaseCtorParam{src->owner_graph(),
+                                          config,
+                                          "adaptive_pooling",
+                                          {src, out_shape}}) {
+    init_megdnn_opr(*this, param);
+    add_input({src, out_shape});
+    outshape_by_symvar_enable(1, 1);
+}
+
+SymbolVar AdaptivePoolingForward::make(SymbolVar src, SymbolVar out_shape,
+                                       const Param& param,
+                                       const OperatorNodeConfig& config) {
+    return src.insert_single_output_opr<AdaptivePoolingForward>(
+            src.node(), out_shape.node(), param, config);
+}
+
+void AdaptivePoolingForward::scn_do_execute() {
+    megdnn_opr()->exec(input(0)->dev_tensor().as_megdnn(),
+                       output(0)->dev_tensor().as_megdnn(),
+                       intl::get_megdnn_workspace_from_var(output().back()));
+}
+
+void AdaptivePoolingForward::outshape_by_symvar_do_get_output_shape(
+        TensorShape& dest, const ShapeInferInfo& shpinfo) {
+    TensorShape oshp2d;
+    cg::copy_tensor_value_to_shape(oshp2d, *shpinfo.shpval_inp_val.at(0));
+    auto src = shpinfo.shape_inp_shp.at(0);
+    mgb_assert(src.ndim == 4 && oshp2d.ndim == 2,
+               "shape mismatch for AdaptivePooling: src=%s, out2d=%s",
+               src.to_string().c_str(), oshp2d.to_string().c_str());
+
+    mgb_assert(param().format == Param::Format::NCHW,
+               "AdaptivePooling only support NCHW");
+    dest.ndim = 4;
+    dest.shape[0] = src.shape[0];
+    dest.shape[1] = src.shape[1];
+    dest.shape[2] = oshp2d.shape[0];
+    dest.shape[3] = oshp2d.shape[1];
+}
+
+size_t AdaptivePoolingForward::get_workspace_size_bytes(
+        const TensorShapeArray& input_shapes,
+        const TensorShapeArray& output_shapes) const {
+    return megdnn_opr()->get_workspace_in_bytes(
+            {input_shapes[0], this->input(0)->dtype(),
+             this->input(0)->format()},
+            {output_shapes[0], this->output(0)->dtype(),
+             this->output(0)->format()});
+}
+
+void AdaptivePoolingForward::init_output_dtype() {
+    output(0)->dtype(input(0)->dtype());
+}
+
+void AdaptivePoolingForward::add_input_layout_constraint() {
+    mixin::megdnn_utils::add_input_layout_constraint_contig(*this);
+}
+
+void AdaptivePoolingForward::init_output_static_infer_desc() {
+    Super::init_output_static_infer_desc();
+    init_output_static_infer_desc_workspace(false);
+}
+
+void AdaptivePoolingForward::record_execute_deps(ExecDependencyArray& deps) {
+    record_megdnn_opr(deps);
+}
+
+#ifdef MGB_ENABLE_GRAD
+MGB_IMPL_OPR_GRAD(AdaptivePoolingForward) {
+    if (wrt_idx == 0) {
+        // wrt src
+        SymbolVar grad = AdaptivePoolingBackward::make(
+                opr.input(0), opr.input(1), opr.output(0), out_grad[0],
+                opr.param());
+        return grad.node();
+    } else {
+        mgb_assert(wrt_idx == 1);
+        return InvalidGrad::make(opr, wrt_idx);
+    }
+}
+#endif
+
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(AdaptivePoolingBackward);
+AdaptivePoolingBackward::AdaptivePoolingBackward(
+        VarNode* src, VarNode* out_shape, VarNode* dst, VarNode* diff,
+        const Param& param, const OperatorNodeConfig& config)
+        : Super(OperatorNodeBaseCtorParam{src->owner_graph(),
+                                          config,
+                                          "adaptive_pooling_bwd",
+                                          {src}},
+                0, true) {
+    init_megdnn_opr(*this, param);
+    add_input({src, out_shape, dst, diff});
+}
+
+SymbolVar AdaptivePoolingBackward::make(SymbolVar src, SymbolVar out_shape,
+                                        SymbolVar dst, SymbolVar diff,
+                                        const Param& param,
+                                        const OperatorNodeConfig& config) {
+    return src.insert_single_output_opr<AdaptivePoolingBackward>(
+            src.node(), out_shape.node(), dst.node(), diff.node(), param,
+            config);
+}
+
+void AdaptivePoolingBackward::scn_do_execute() {
+    megdnn_opr()->exec(input(0)->dev_tensor().as_megdnn(),
+                       input(2)->dev_tensor().as_megdnn(),
+                       input(3)->dev_tensor().as_megdnn(),
+                       output(0)->dev_tensor().as_megdnn(),
+                       intl::get_megdnn_workspace_from_var(output().back()));
+}
+size_t AdaptivePoolingBackward::get_workspace_size_bytes(
+        const TensorShapeArray& input_shapes,
+        const TensorShapeArray& output_shapes) const {
+    return megdnn_opr()->get_workspace_in_bytes(
+            {input_shapes[0], input(0)->dtype(), input(0)->format()},
+            {input_shapes[2], input(2)->dtype(), input(2)->format()},
+            {input_shapes[3], input(3)->dtype(), input(3)->format()},
+            {output_shapes[0], output(0)->dtype(), output(0)->format()});
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/dnn/dnn.oprdecl

@@ -140,6 +140,13 @@ decl_opr('Pooling',
          inputs=['src'],
          params='Pooling')
 
+decl_opr('AdaptivePooling',
+        inputs=[Doc('src', 'input image, shape (n, c, ih, iw)'),
+                Doc('out_shape', 'output image shape, containing two elements specifying output height and width.')],
+        params='AdaptivePooling',
+        desc='Adaptive Pooling.'
+        'The output shape is (n, c, oh, ow), where (oh, ow) is given by *out_shape*.')
+
 decl_opr('ROIPooling', outputs=[0],
          inputs=[Doc('src', 'input image, shape (n, c, ih, iw)'),
                  Doc('rois', 'regions of interest, shape (m, 5). '
@@ -258,7 +265,7 @@ decl_opr('ROIAlign', outputs=[0],
                      'store it as a float, but it should be an integral value.'
                      ' The rois[:, 1:5] are (x0, y0, x1, y1) for each ROI, '
                      'which would be multiplied by the scale value given in '
-                     'param.')], 
+                     'param.')],
          params='ROIAlign',
          desc='ROI Align, see '
          'Mask-RCNN: https://arxiv.org/pdf/1703.06870.pdf, '
@@ -295,7 +302,7 @@ decl_opr('BatchConvBiasForward',
                  ('execution_policy', 'ExecutionPolicy')],
          desc=Doc(None,
 r"""
-    Apply a convolution of input tensor and filter tensor whose weights are not shared in batch dimensions. Outputs with batch index use the same weight. 
+    Apply a convolution of input tensor and filter tensor whose weights are not shared in batch dimensions. Outputs with batch index use the same weight.
     Assume input shape is :math:`(N, IC, IH, IW)` and filter shape is :math:`(batch, OC, IC, FH, FW)`, the output shape will be :math:`(N, OC, OH, OW)` where :math:`(OH, OW)` would be computed from padding, stride, :math:`(FH, FW)` and :math:`(IH, IW)`, as in convolution.
     for each output location, we have;
 

src/opr/impl/dnn/dnn.sereg.h

@@ -13,6 +13,7 @@
 #include "megbrain/opr/dnn/convolution.h"
 #include "megbrain/opr/dnn/images2neibs.h"
 #include "megbrain/opr/dnn/pooling.h"
+#include "megbrain/opr/dnn/adaptive_pooling.h"
 #include "megbrain/opr/dnn/roi_pooling.h"
 #include "megbrain/opr/dnn/roi_align.h"
 #include "megbrain/opr/dnn/local.h"
@@ -388,6 +389,9 @@ namespace opr {
     MGB_SEREG_OPR(Pooling, 1);
     MGB_SEREG_OPR(PoolingBackward, 3);
 
+    MGB_SEREG_OPR(AdaptivePooling, 2);
+    MGB_SEREG_OPR(AdaptivePoolingBackward, 4);
+
     MGB_SEREG_OPR(ROIPooling, 3);
     MGB_SEREG_OPR(ROIPoolingBackward, 4);
 

src/opr/include/megbrain/opr/dnn/adaptive_pooling.h

@@ -0,0 +1,76 @@
+/**
+ * \file src/opr/include/megbrain/opr/dnn/adaptive_pooling.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#pragma once
+
+#include "megbrain/opr/internal/megdnn_opr_wrapper.h"
+#include "megbrain/opr/internal/out_shape_by_sym_var.h"
+#include "megdnn/opr_param_defs.h"
+#include "megdnn/oprs/nn.h"
+
+namespace mgb {
+namespace opr {
+
+MGB_DEFINE_OPR_CLASS(
+        AdaptivePoolingForward,
+        intl::WorkspaceSizeInfer<intl::OutshapeBySymvarSCNOpr<
+                mixin::MegDNNOprHolderImpl<megdnn::AdaptivePoolingForward>>>) // {
+public:
+    AdaptivePoolingForward(VarNode * src, VarNode * out_shape,
+                           const Param& param,
+                           const OperatorNodeConfig& config);
+    static SymbolVar make(SymbolVar src, SymbolVar out_shape,
+                          const Param& param,
+                          const OperatorNodeConfig& config = {});
+    static SymbolVar make(SymbolVar src, const TensorShape& out_shape,
+                          const Param& param,
+                          const OperatorNodeConfig& config = {}) {
+        return make(src, cg::var_from_tensor_shape(src, out_shape), param,
+                    config);
+    }
+
+private:
+    void scn_do_execute() override;
+    void outshape_by_symvar_do_get_output_shape(
+            TensorShape & dest, const ShapeInferInfo& shpinfo) override;
+    size_t get_workspace_size_bytes(const TensorShapeArray& input_shapes,
+                                    const TensorShapeArray& output_shapes)
+            const override;
+    void init_output_dtype() override;
+    void add_input_layout_constraint() override;
+    void init_output_static_infer_desc() override;
+    void record_execute_deps(ExecDependencyArray& deps) override;
+};
+using AdaptivePooling = AdaptivePoolingForward;
+
+MGB_DEFINE_OPR_CLASS(
+        AdaptivePoolingBackward,
+        intl::MegDNNOprWrapperBwd<megdnn::AdaptivePoolingBackward>) // {
+public:
+    AdaptivePoolingBackward(VarNode * src, VarNode * out_shape, VarNode * dst,
+                            VarNode * diff, const Param& param,
+                            const OperatorNodeConfig& config);
+    static SymbolVar make(SymbolVar src, SymbolVar out_shape, SymbolVar dst,
+                          SymbolVar diff, const Param& param,
+                          const OperatorNodeConfig& config = {});
+
+private:
+    void scn_do_execute() override;
+    size_t get_workspace_size_bytes(const TensorShapeArray& input_shapes,
+                                    const TensorShapeArray& output_shapes)
+            const override;
+};
+
+}  // namespace opr
+}  // namespace mgb
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/test/dnn/adaptive_pooling.cpp

@@ -0,0 +1,87 @@
+/**
+ * \file src/opr/test/dnn/adaptive_pooling.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megbrain/opr/dnn/adaptive_pooling.h"
+#include "megbrain/comp_node_env.h"
+#include "megbrain/opr/dnn/pooling.h"
+#include "megbrain/opr/tensor_manip.h"
+#include "megbrain/test/autocheck.h"
+#include "megbrain/test/megdnn_helper.h"
+#include "megdnn/dtype.h"
+#include "megdnn/opr_param_defs.h"
+
+using namespace std;
+using namespace mgb;
+
+namespace {
+
+using Param = opr::AdaptivePoolingForward::Param;
+void run(Param::Mode mode) {
+    using Checker = AutoOprChecker<2, 1>;
+    Param param{mode};
+
+    auto make_graph =
+            [&](const Checker::SymInpArray& inputs) -> Checker::SymOutArray {
+        auto o0 = opr::GetVarShape::make(inputs[1]);
+        auto o1 = opr::AdaptivePoolingForward::make(inputs[0], o0, param);
+        return {o1};
+    };
+
+    auto fwd = [&](Checker::NumOutArray& dest, Checker::NumInpArray inp) {
+        auto opr = MegDNNHandle::get(
+                           CompNodeEnv::from_comp_node(CompNode::default_cpu()))
+                           ->create_operator<megdnn::AdaptivePoolingForward>();
+        opr->param() = param;
+        size_t N = inp[0].get()->shape(0), C = inp[0].get()->shape(1);
+        size_t OH = inp[1].get()->shape(0), OW = inp[1].get()->shape(1);
+        dest[0].resize(TensorShape{N, C, OH, OW});
+        opr->exec(inp[0]->as_megdnn(), dest[0].as_megdnn(), {});
+    };
+
+    auto gen = [&](HostTensorND& src) {
+        if (mode == Param::Mode::MAX) {
+            HostTensorGenerator<dtype::Float32, RandomDistribution::CONSECUTIVE>
+                    src_gen(1.0f, 0.1f);
+            src = *src_gen(src.shape(), src.comp_node());
+        } else {
+            HostTensorGenerator<dtype::Float32, RandomDistribution::GAUSSIAN>
+                    src_gen(10.f);
+            src = *src_gen(src.shape(), src.comp_node());
+        }
+    };
+
+    Checker::RunOptions opt;
+    opt.numdiff_max_err = 1e-2;
+
+    Checker checker{make_graph, fwd};
+    checker.set_input_allow_grad(1, false)
+           .set_input_generator(0, gen);
+    checker.run({TensorShape{1, 1, 10, 7}, TensorShape{5, 4}}, opt);
+    checker.run({TensorShape{1, 1, 9, 7}, TensorShape{5, 4}}, opt);
+    checker.run({TensorShape{1, 2, 8, 9}, TensorShape{3, 4}}, opt);
+}
+
+}  // anonymous namespace
+
+TEST(TestOprDNN, AdaptivePoolingMax) {
+    run(Param::Mode::MAX);
+}
+
+TEST(TestOprDNN, AdaptivePoolingAverage) {
+    run(Param::Mode::AVERAGE);
+}
+
+TEST(TestOprDNN, AdaptivePoolingAverageCountExcludePadding) {
+    run(Param::Mode::AVERAGE_COUNT_EXCLUDE_PADDING);
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/serialization/impl/schema.fbs

@@ -99,6 +99,7 @@ union OperatorParam {
     DType = 67,
     param.Remap = 68,
     param.NMSKeep = 69,
+    param.AdaptivePooling = 70,
 }
 
 table Operator {

test/src/helper.cpp

@@ -113,6 +113,20 @@ dtype, RandomDistribution::CONSTANT>::operator ()(
     return ret;
 }
 
+template<typename dtype>
+std::shared_ptr<HostTensorND> HostTensorGenerator<
+dtype, RandomDistribution::CONSECUTIVE>::operator ()(
+        const TensorShape &shape, CompNode cn) {
+    if (!cn.valid())
+        cn = CompNode::load("xpu0");
+    std::shared_ptr<HostTensorND> ret =
+        std::make_shared<HostTensorND>(cn, shape, dtype());
+    auto ptr = ret->ptr<ctype>();
+    for (size_t i = 0, it = shape.total_nr_elems(); i < it; ++ i) {
+        ptr[i] = m_val + i * m_delta;
+    }
+    return ret;
+}
 
 // explicit instantialization of HostTensorGenerator
 namespace mgb {
@@ -122,12 +136,16 @@ namespace mgb {
         dtype::Float32, RandomDistribution::UNIFORM>;
     template class HostTensorGenerator<
         dtype::Float32, RandomDistribution::CONSTANT>;
+    template class HostTensorGenerator<
+        dtype::Float32, RandomDistribution::CONSECUTIVE>;
     template class HostTensorGenerator<
         dtype::Float16, RandomDistribution::GAUSSIAN>;
     template class HostTensorGenerator<
         dtype::Int8, RandomDistribution::UNIFORM>;
     template class HostTensorGenerator<
         dtype::Int8, RandomDistribution::CONSTANT>;
+    template class HostTensorGenerator<
+        dtype::Int8, RandomDistribution::CONSECUTIVE>;
     template class HostTensorGenerator<
         dtype::Uint8, RandomDistribution::UNIFORM>;
     template class HostTensorGenerator<

test/src/include/megbrain/test/helper.h

@@ -168,7 +168,7 @@ class RNGxorshf {
 };
 
 enum class RandomDistribution {
-    GAUSSIAN, UNIFORM, CONSTANT
+    GAUSSIAN, UNIFORM, CONSTANT, CONSECUTIVE
 };
 
 template<class dtype>
@@ -342,6 +342,29 @@ class HostTensorGenerator<dtype, RandomDistribution::CONSTANT> final:
     private:
         ctype m_default_val;
 };
+
+//! consecutive value
+template<class dtype>
+class HostTensorGenerator<dtype, RandomDistribution::CONSECUTIVE> final:
+        public HostTensorGeneratorBase {
+
+    public:
+        using ctype = typename DTypeTrait<dtype>::ctype;
+
+        HostTensorGenerator(ctype val, ctype delta)
+                : HostTensorGeneratorBase{next_rand_seed()},
+                  m_val{val}, m_delta{delta} {}
+
+        std::shared_ptr<HostTensorND> operator ()(
+                const TensorShape &shape, CompNode cn = {}) override;
+        using HostTensorGeneratorBase::operator();
+
+    private:
+        ctype m_val;
+        ctype m_delta;
+};
+
+
 template <>
 class HostTensorGenerator<dtype::Bool, RandomDistribution::UNIFORM> final
         : public HostTensorGeneratorBase {