WIP: Convert models to formats supported by mace/ncnn via pnnx

.clang-format

@@ -1,132 +1,9 @@
-# find src/ tools/ tests/ examples/ benchmark/ -type f -name '*.c' -o -name '*.cpp' -o -name '*.h' | xargs -i clang-format -i {}
-
-# need clang-format >= 10.0
-
-AccessModifierOffset: -4
-AlignAfterOpenBracket: Align
-AlignConsecutiveAssignments: false
-# AlignConsecutiveBitFields: true
-AlignConsecutiveDeclarations: false
-AlignConsecutiveMacros: true
-AlignEscapedNewlines: Left
-# AlignOperands: AlignAfterOperator
-AlignTrailingComments: true
-AllowAllArgumentsOnNextLine: true
-AllowAllConstructorInitializersOnNextLine: true
-AllowAllParametersOfDeclarationOnNextLine: true
-AllowShortBlocksOnASingleLine: Always
-AllowShortCaseLabelsOnASingleLine: true
-# AllowShortEnumsOnASingleLine: true
-AllowShortFunctionsOnASingleLine: None
-AllowShortIfStatementsOnASingleLine: WithoutElse
-AllowShortLambdasOnASingleLine: All
-AllowShortLoopsOnASingleLine: true
-AlwaysBreakAfterReturnType: None
-AlwaysBreakBeforeMultilineStrings: false
-AlwaysBreakTemplateDeclarations: Yes
-BinPackArguments: true
-BinPackParameters: true
-BraceWrapping:
-  AfterCaseLabel: true
-  AfterClass: true
-  AfterControlStatement: Always
-  AfterEnum: true
-  AfterFunction: true
-  AfterNamespace: false
-  AfterObjCDeclaration: false
-  AfterStruct: true
-  AfterUnion: true
-  AfterExternBlock: false
-  BeforeCatch: true
-  BeforeElse: true
-#  BeforeLambdaBody: false
-#  BeforeWhile: false
-  IndentBraces: false
-  SplitEmptyFunction: true
-  SplitEmptyRecord: true
-  SplitEmptyNamespace: false
-BreakAfterJavaFieldAnnotations: true
-BreakBeforeBinaryOperators: All
-BreakBeforeBraces: Custom
-BreakBeforeTernaryOperators: true
-BreakConstructorInitializers: BeforeColon
-BreakInheritanceList: BeforeColon
-BreakStringLiterals: false
-ColumnLimit: 0
-# CommentPragmas:
-CompactNamespaces: false
-ConstructorInitializerAllOnOneLineOrOnePerLine: true
-ConstructorInitializerIndentWidth: 4
-ContinuationIndentWidth: 4
-Cpp11BracedListStyle: true
-DeriveLineEnding: false
+---
+BasedOnStyle: Google
+---
+Language:               Cpp
+Cpp11BracedListStyle:   true
+Standard:               Cpp11
 DerivePointerAlignment: false
-# DisableFormat:
-# ExperimentalAutoDetectBinPacking:
-FixNamespaceComments: true
-# ForEachMacros:
-IncludeBlocks: Regroup
-# IncludeCategories:
-# IncludeIsMainRegex:
-# IncludeIsMainSourceRegex:
-# IndentCaseBlocks: false
-IndentCaseLabels: false
-# IndentExternBlock: NoIndent
-IndentGotoLabels: false
-IndentPPDirectives: None
-IndentWidth: 4
-# IndentWrappedFunctionNames: 4
-# InsertTrailingCommas: None
-# JavaImportGroups:
-# JavaScriptQuotes
-# JavaScriptWrapImports:
-KeepEmptyLinesAtTheStartOfBlocks: false
-Language: Cpp
-# MacroBlockBegin:
-# MacroBlockEnd:
-MaxEmptyLinesToKeep: 1
-NamespaceIndentation: None
-# NamespaceMacros:
-# ObjCBinPackProtocolList:
-# ObjCBlockIndentWidth:
-# ObjCBreakBeforeNestedBlockParam:
-# ObjCSpaceAfterProperty:
-# ObjCSpaceBeforeProtocolList:
-# PenaltyBreakAssignment:
-# PenaltyBreakBeforeFirstCallParameter:
-# PenaltyBreakComment:
-# PenaltyBreakFirstLessLess:
-# PenaltyBreakString:
-# PenaltyBreakTemplateDeclaration:
-# PenaltyExcessCharacter:
-# PenaltyReturnTypeOnItsOwnLine:
-PointerAlignment: Left
-# RawStringFormats:
-ReflowComments: false
-SortIncludes: false
-SortUsingDeclarations: true
-SpaceAfterCStyleCast: false
-SpaceAfterLogicalNot: false
-SpaceAfterTemplateKeyword: false
-SpaceBeforeAssignmentOperators: true
-SpaceBeforeCpp11BracedList: false
-SpaceBeforeCtorInitializerColon: true
-SpaceBeforeInheritanceColon: true
-SpaceBeforeParens: ControlStatements
-SpaceBeforeRangeBasedForLoopColon: true
-SpaceBeforeSquareBrackets: false
-SpaceInEmptyBlock: false
-SpaceInEmptyParentheses: false
-SpacesBeforeTrailingComments: 1
-SpacesInAngles: false
-SpacesInCStyleCastParentheses: false
-SpacesInConditionalStatement: false
-SpacesInContainerLiterals: false
-SpacesInParentheses: false
-SpacesInSquareBrackets: false
-Standard: c++03
-#StatementMacros:
-TabWidth: 4
-# TypenameMacros:
-UseCRLF: false
-UseTab: Never
+PointerAlignment:       Right
+---

.github/workflows/release-python.yml

@@ -53,7 +53,7 @@ jobs:
         brew uninstall --ignore-dependencies libomp
 
     - name: Build wheels
-      uses: pypa/cibuildwheel@v2.8.1
+      uses: pypa/cibuildwheel@v2.9.0
       env:
         CIBW_ARCHS_MACOS: x86_64 universal2 arm64
         CIBW_ARCHS_LINUX: x86_64 i686
@@ -103,7 +103,7 @@ jobs:
         platforms: all
 
     - name: Build wheels
-      uses: pypa/cibuildwheel@v2.8.1
+      uses: pypa/cibuildwheel@v2.9.0
       env:
         CIBW_ARCHS_LINUX: ${{ matrix.arch }}
         CIBW_BUILD: ${{ matrix.build }}

benchmark/README.md

@@ -4322,3 +4322,51 @@ cooling_down = 0
       yolo-fastestv2  min =    3.33  max =    3.73  avg =    3.44
   vision_transformer  min =  214.91  max =  229.91  avg =  220.82
 ```
+
+### Intel Celeron M 420 (Yonah 1.60 GHz x 1)
+
+Tested on `Debian GNU/Linux 11 (bullseye) i686` with `cmake -DNCNN_RUNTIME_CPU=OFF -DNCNN_AVX=OFF -DNCNN_AVX2=OFF -DNCNN_AVX512=OFF-DNCNN_BUILD_TESTS=ON ..`.
+
+```
+mouri@Mouri-Laptop-2:~/ncnn/benchmark$ ./../build/benchmark/benchncnn 
+loop_count = 4
+num_threads = 1
+powersave = 0
+gpu_device = -1
+cooling_down = 1
+          squeezenet  min =  289.23  max =  301.83  avg =  292.90
+     squeezenet_int8  min =  442.82  max =  457.21  avg =  446.89
+           mobilenet  min =  549.62  max =  561.20  avg =  554.78
+      mobilenet_int8  min =  823.92  max =  837.70  avg =  830.52
+        mobilenet_v2  min =  341.72  max =  353.77  avg =  345.34
+        mobilenet_v3  min =  267.68  max =  282.08  avg =  273.10
+          shufflenet  min =  151.66  max =  153.02  avg =  152.24
+       shufflenet_v2  min =  161.54  max =  163.38  avg =  162.13
+             mnasnet  min =  322.66  max =  336.91  avg =  326.86
+     proxylessnasnet  min =  356.63  max =  368.79  avg =  360.66
+     efficientnet_b0  min =  489.92  max =  505.11  avg =  497.32
+   efficientnetv2_b0  min =  618.16  max =  632.02  avg =  622.82
+        regnety_400m  min =  414.83  max =  428.42  avg =  419.28
+           blazeface  min =   38.56  max =   40.05  avg =   39.05
+           googlenet  min = 1022.54  max = 1037.53  avg = 1029.48
+      googlenet_int8  min = 1493.35  max = 1495.46  avg = 1494.31
+            resnet18  min =  803.32  max =  818.27  avg =  812.49
+       resnet18_int8  min = 1188.26  max = 1200.88  avg = 1192.56
+             alexnet  min =  613.78  max =  623.88  avg =  619.99
+               vgg16  min = 4465.44  max = 4478.12  avg = 4474.16
+          vgg16_int8  min = 6042.40  max = 6114.37  avg = 6077.07
+            resnet50  min = 2517.75  max = 2528.42  avg = 2522.83
+       resnet50_int8  min = 3746.28  max = 3771.09  avg = 3756.88
+      squeezenet_ssd  min =  585.56  max =  636.01  avg =  602.62
+ squeezenet_ssd_int8  min =  822.43  max =  968.77  avg =  862.33
+       mobilenet_ssd  min = 1116.98  max = 1139.17  avg = 1127.65
+  mobilenet_ssd_int8  min = 1665.03  max = 1670.55  avg = 1668.37
+      mobilenet_yolo  min = 2638.61  max = 2666.54  avg = 2652.26
+  mobilenetv2_yolov3  min = 1248.56  max = 1255.98  avg = 1251.22
+         yolov4-tiny  min = 1507.31  max = 1525.56  avg = 1514.66
+           nanodet_m  min =  386.41  max =  400.63  avg =  391.21
+    yolo-fastest-1.1  min =  159.97  max =  164.53  avg =  161.41
+      yolo-fastestv2  min =  134.29  max =  135.47  avg =  134.70
+  vision_transformer  min = 22201.32  max = 22510.75  avg = 22315.09
+          FastestDet  min =  146.94  max =  148.50  avg =  147.44
+```

notes.md

@@ -0,0 +1,6 @@
+# How to install
+
+```bash
+python3 setup.py bdist_wheel
+pip install ./dist/*.whl
+```

pyproject.toml

@@ -1,8 +1,8 @@
 [build-system]
 requires = [
     "setuptools>=42",
-    "wheel",
-    "ninja; sys_platform != 'win32'",
-    "cmake>=3.12",
+    # "wheel",
+    # "ninja; sys_platform != 'win32'",
+    # "cmake>=3.12",
 ]
 build-backend = "setuptools.build_meta"

python/src/main.cpp

@@ -244,30 +244,15 @@ PYBIND11_MODULE(ncnn, m)
 
     .def(py::init([](py::buffer const b) {
         py::buffer_info info = b.request();
+
         if (info.ndim > 4)
         {
             std::stringstream ss;
             ss << "convert numpy.ndarray to ncnn.Mat only dims <=4 support now, but given " << info.ndim;
             pybind11::pybind11_fail(ss.str());
         }
 
-        size_t elemsize = 4u;
-        if (info.format == py::format_descriptor<double>::format())
-        {
-            elemsize = 8u;
-        }
-        if (info.format == py::format_descriptor<float>::format() || info.format == py::format_descriptor<int>::format())
-        {
-            elemsize = 4u;
-        }
-        else if (info.format == "e")
-        {
-            elemsize = 2u;
-        }
-        else if (info.format == py::format_descriptor<int8_t>::format() || info.format == py::format_descriptor<uint8_t>::format())
-        {
-            elemsize = 1u;
-        }
+        size_t elemsize = info.itemsize;
 
         Mat* v = nullptr;
         if (info.ndim == 1)
@@ -296,66 +281,16 @@ PYBIND11_MODULE(ncnn, m)
             // so we set the cstep as numpy's cstep
             v->cstep = (int)info.shape[3] * (int)info.shape[2] * (int)info.shape[1];
         }
-        return v;
+        return std::unique_ptr<Mat>(v);
     }),
     py::arg("array"))
     .def_buffer([](Mat& m) -> py::buffer_info {
-        if (m.elemsize != 1 && m.elemsize != 2 && m.elemsize != 4)
-        {
-            std::stringstream ss;
-            ss << "convert ncnn.Mat to numpy.ndarray only elemsize 1, 2, 4 support now, but given " << m.elemsize;
-            pybind11::pybind11_fail(ss.str());
-        }
-        if (m.elempack != 1)
-        {
-            std::stringstream ss;
-            ss << "convert ncnn.Mat to numpy.ndarray only elempack 1 support now, but given " << m.elempack;
-            pybind11::pybind11_fail(ss.str());
-        }
-        std::string format = get_mat_format(m);
-        std::vector<py::ssize_t> shape;
-        std::vector<py::ssize_t> strides;
-        if (m.dims == 1)
-        {
-            shape.push_back(m.w);
-            strides.push_back(m.elemsize);
-        }
-        else if (m.dims == 2)
-        {
-            shape.push_back(m.h);
-            shape.push_back(m.w);
-            strides.push_back(m.w * m.elemsize);
-            strides.push_back(m.elemsize);
-        }
-        else if (m.dims == 3)
-        {
-            shape.push_back(m.c);
-            shape.push_back(m.h);
-            shape.push_back(m.w);
-            strides.push_back(m.cstep * m.elemsize);
-            strides.push_back(m.w * m.elemsize);
-            strides.push_back(m.elemsize);
-        }
-        else if (m.dims == 4)
-        {
-            shape.push_back(m.c);
-            shape.push_back(m.d);
-            shape.push_back(m.h);
-            shape.push_back(m.w);
-            strides.push_back(m.cstep * m.elemsize);
-            strides.push_back(m.w * m.h * m.elemsize);
-            strides.push_back(m.w * m.elemsize);
-            strides.push_back(m.elemsize);
-        }
-        return py::buffer_info(
-            m.data,     /* Pointer to buffer */
-            m.elemsize, /* Size of one scalar */
-            format,     /* Python struct-style format descriptor */
-            m.dims,     /* Number of dimensions */
-            shape,      /* Buffer dimensions */
-            strides     /* Strides (in bytes) for each index */
-        );
+        return to_buffer_info(m);
     })
+    .def("numpy", [](py::object obj, const std::string& format="") -> py::array {
+        auto* m = obj.cast<Mat*>();
+        return py::array(to_buffer_info(*m, format), obj);
+    }, py::arg("format")="")
     //.def("fill", (void (Mat::*)(int))(&Mat::fill), py::arg("v"))
     .def("fill", (void (Mat::*)(float))(&Mat::fill), py::arg("v"))
     .def("clone", &Mat::clone, py::arg("allocator") = nullptr)