8.2 Release (#2433)

* 8.2 Release

* bump version to 8.2

* fix flake8 error https://gitlab.com/coremltools1/coremltools/-/jobs/8889872126

* fix CI flakiness: avoid duplicate elements in input tensor so output indices are unique

---------

Co-authored-by: yifan_shen3 <yifan_shen3@apple.com>

Author: YifanShenSZ

.gitlab-ci.yml

@@ -90,7 +90,7 @@ test_py39_pytorch_intel:
     PYTHON: "3.9"
     TEST_PACKAGE: coremltools.converters.mil.frontend.torch
     WHEEL_PATH: build/dist/*cp39*10_15*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_torch.pip
 
 test_py37_tf1_intel:
   <<: *test_macos_pkg
@@ -200,7 +200,7 @@ test_py310_pytorch_script:
     PYTHON: "3.10"
     TEST_PACKAGE: coremltools.converters.mil.frontend.torch
     WHEEL_PATH: build/dist/*cp310*11*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_torch.pip
     TORCH_FRONTENDS: TORCHSCRIPT
 
 test_py310_pytorch_export:
@@ -213,7 +213,7 @@ test_py310_pytorch_export:
     PYTHON: "3.10"
     TEST_PACKAGE: coremltools.converters.mil.frontend.torch
     WHEEL_PATH: build/dist/*cp310*11*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_torch.pip
     TORCH_FRONTENDS: TORCHEXPORT
 
 test_py310_pytorch_executorch:
@@ -226,7 +226,7 @@ test_py310_pytorch_executorch:
     PYTHON: "3.10"
     TEST_PACKAGE: coremltools.converters.mil.frontend.torch
     WHEEL_PATH: build/dist/*cp310*11*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_executorch.pip
     TORCH_FRONTENDS: EXECUTORCH
 
 test_py310_tf2-1:

coremlpython/CoreMLPython.h

@@ -17,6 +17,7 @@
 #import <Availability.h>
 
 #import <vector>
+#import <optional>
 
 #ifndef BUILT_WITH_MACOS15_SDK
 #define BUILT_WITH_MACOS15_SDK \
@@ -73,6 +74,9 @@ namespace CoreML {
         };
 
         struct CPUComputeDevice {
+            inline CPUComputeDevice():
+            m_impl(nil) {}
+
             // MLCPUComputeDevice must be wrapped in a C++ class for PyBind.
             inline CPUComputeDevice(id impl):
             m_impl(impl) {}
@@ -90,6 +94,9 @@ namespace CoreML {
         };
 
         struct GPUComputeDevice {
+            inline GPUComputeDevice():
+            m_impl(nil) {}
+
             // MLGPUComputeDevice must be wrapped in a C++ class for PyBind.
             inline GPUComputeDevice(id impl):
             m_impl(impl) {}
@@ -107,6 +114,9 @@ namespace CoreML {
         };
 
         struct NeuralEngineComputeDevice {
+            inline NeuralEngineComputeDevice():
+            m_impl(nil) {}
+
             // MLNeuralEngineComputeDevice must be wrapped in a C++ class for PyBind.
             inline NeuralEngineComputeDevice(id impl):
             m_impl(impl) {}
@@ -160,6 +170,9 @@ namespace CoreML {
         };
 
         struct ComputePlan {
+            inline ComputePlan():
+            m_impl(nil), m_modelStructure(py::none()) {}
+
             // MLComputePlan must be wrapped in a C++ class for PyBind.
             inline ComputePlan(id impl, py::object modelStructure):
             m_impl(impl),
@@ -191,10 +204,12 @@ namespace CoreML {
             m_impl(impl),
             m_datas(std::move(datas)) {}
 
+#if ML_MODEL_ASSET_IS_AVAILABLE
             API_AVAILABLE(macos(13.0))
             inline MLModelAsset *getImpl() const {
                 return (MLModelAsset *)m_impl;
             }
+#endif
 
             id m_impl = nil;
             std::vector<py::bytes> m_datas;
@@ -205,6 +220,8 @@ namespace CoreML {
             MLModel *m_model = nil;
             NSURL *compiledUrl = nil;
             bool m_deleteCompiledModelOnExit = false;
+            std::optional<uint64_t> m_loadDurationInNanoSeconds;
+            std::optional<uint64_t> m_lastPredictDurationInNanoSeconds;
 
         public:
             static py::bytes autoSetSpecificationVersion(const py::bytes& modelBytes);
@@ -229,11 +246,16 @@ namespace CoreML {
 
             explicit Model(MLModel* m_model, NSURL* compiledUrl, bool deleteCompiledModelOnExit);
 
-            py::list batchPredict(const py::list& batch) const;
+            py::list batchPredict(const py::list& batch);
 
             py::str getCompiledModelPath() const;
 
-            py::dict predict(const py::dict& input, State* state=NULL) const;
+            py::dict predict(const py::dict& input, State* state=NULL);
+
+            py::object getLoadDurationInNanoSeconds() const;
+
+            py::object getLastPredictDurationInNanoSeconds() const;
+
 
 #if BUILT_WITH_MACOS15_SDK
             static void setOptimizationHints(MLModelConfiguration *configuration, const py::dict& optimizationHints);

coremlpython/CoreMLPython.mm

@@ -4,6 +4,7 @@
 // found in the LICENSE.txt file or at https://opensource.org/licenses/BSD-3-Clause
 
 #import <CoreML/CoreML.h>
+#include <mach/mach_time.h>
 #import "CoreMLPythonArray.h"
 #import "CoreMLPython.h"
 #import "CoreMLPythonUtils.h"
@@ -289,20 +290,34 @@
     }
 #endif
 
+uint64_t convertMachTimeToNanoSeconds(uint64_t time) {
+    static dispatch_once_t once;
+    static mach_timebase_info_data_t timebase;
+    dispatch_once(&once, ^{
+        mach_timebase_info(&timebase);
+    });
+    uint64_t result = (time * timebase.numer) / timebase.denom;
+    return result;
+}
 
 #if ML_MODEL_ASSET_IS_AVAILABLE
     API_AVAILABLE(macos(13.0))
-    MLModel * createModelFromModelAsset(MLModelAsset *modelAsset,
-                                        MLModelConfiguration *configuration,
-                                        NSError * __autoreleasing *error) {
+    std::pair<MLModel *, uint64_t> createModelFromModelAsset(
+        MLModelAsset *modelAsset,
+        MLModelConfiguration *configuration,
+        NSError * __autoreleasing *error
+    ) {
         dispatch_semaphore_t sem = dispatch_semaphore_create(0);
         __block MLModel *result = nil;
         __block NSError *lError = nil;
+        uint64_t loadStartTime = mach_absolute_time();
+        __block uint64_t loadEndTime = loadStartTime;
         [MLModel loadModelAsset:modelAsset
                  configuration:configuration
              completionHandler:^(MLModel * _Nullable model, NSError * _Nullable loadError){
                 result = model;
                 lError = loadError;
+                loadEndTime = mach_absolute_time();
                 dispatch_semaphore_signal(sem);
         }];
 
@@ -312,9 +327,9 @@
             *error = lError;
         }
 
-        return result;
+        uint64_t loadDurationInNanoSeconds = convertMachTimeToNanoSeconds(loadEndTime - loadStartTime);
+        return {result, loadDurationInNanoSeconds};
     }
-
 #endif
 }
 
@@ -380,18 +395,25 @@ bool usingMacOS13OrHigher() {
             configuration.functionName = [NSString stringWithUTF8String:functionName.c_str()];
         }
 #endif
+        uint64_t loadDurationInNanoSeconds = 0;
         // Create MLModel
         if (asset.is_none()) {
+            uint64_t loadStartTime = mach_absolute_time();
             m_model = [MLModel modelWithContentsOfURL:compiledUrl configuration:configuration error:&error];
+            uint64_t loadEndTime = mach_absolute_time();
+            loadDurationInNanoSeconds = convertMachTimeToNanoSeconds(loadEndTime - loadStartTime);
         } else {
 #if ML_MODEL_ASSET_IS_AVAILABLE
-            m_model = createModelFromModelAsset(py::cast<ModelAsset>(asset).getImpl(), configuration, &error);
+            auto pair = createModelFromModelAsset(py::cast<ModelAsset>(asset).getImpl(), configuration, &error);
+            m_model = pair.first;
+            loadDurationInNanoSeconds = pair.second;
 #else
             throw std::runtime_error("MLModelAsset is only available on macOS >= 13.0");
 #endif
         }
 
         Utils::handleError(error);
+        m_loadDurationInNanoSeconds = loadDurationInNanoSeconds;
     }
 }
 
@@ -410,13 +432,14 @@ bool usingMacOS13OrHigher() {
 }
 
 
-py::dict Model::predict(const py::dict& input, State* state) const {
+py::dict Model::predict(const py::dict& input, State* state) {
     @autoreleasepool {
         NSError *error = nil;
         MLDictionaryFeatureProvider *inFeatures = Utils::dictToFeatures(input, &error);
         Utils::handleError(error);
 
         id<MLFeatureProvider> outFeatures;
+        uint64_t predictStartTime = mach_absolute_time();
 #if BUILT_WITH_MACOS15_SDK
         if (state == NULL) {
           outFeatures = [m_model predictionFromFeatures:static_cast<MLDictionaryFeatureProvider * _Nonnull>(inFeatures)
@@ -430,8 +453,10 @@ bool usingMacOS13OrHigher() {
         outFeatures = [m_model predictionFromFeatures:static_cast<MLDictionaryFeatureProvider * _Nonnull>(inFeatures)
                                                 error:&error];
 #endif
-
+        uint64_t predictEndTime = mach_absolute_time();
         Utils::handleError(error);
+
+        m_lastPredictDurationInNanoSeconds = convertMachTimeToNanoSeconds(predictEndTime - predictStartTime);
         return Utils::featuresToDict(outFeatures);
     }
 }
@@ -485,7 +510,7 @@ bool usingMacOS13OrHigher() {
 }
 #endif
 
-py::list Model::batchPredict(const py::list& batch) const {
+py::list Model::batchPredict(const py::list& batch) {
   @autoreleasepool {
       NSError* error = nil;
 
@@ -498,11 +523,14 @@ bool usingMacOS13OrHigher() {
       }
       MLArrayBatchProvider* batchProvider = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray: array];
 
+      uint64_t predictStartTime = mach_absolute_time();
       // Get predictions
       MLArrayBatchProvider* predictions = (MLArrayBatchProvider*)[m_model predictionsFromBatch:batchProvider
                                                                                          error:&error];
+      uint64_t predictEndTime = mach_absolute_time();
       Utils::handleError(error);
 
+      m_lastPredictDurationInNanoSeconds = convertMachTimeToNanoSeconds(predictEndTime - predictStartTime);
       // Convert predictions to output
       py::list ret;
       for (int i = 0; i < predictions.array.count; i++) {
@@ -773,6 +801,22 @@ bool usingMacOS13OrHigher() {
     return CoreML::MLMODEL_SPECIFICATION_VERSION_NEWEST;
 }
 
+py::object Model::getLoadDurationInNanoSeconds() const {
+    if (m_loadDurationInNanoSeconds) {
+        return py::cast(m_loadDurationInNanoSeconds.value());
+    }
+
+    return py::none();
+}
+
+py::object Model::getLastPredictDurationInNanoSeconds() const {
+    if (m_lastPredictDurationInNanoSeconds) {
+        return py::cast(m_lastPredictDurationInNanoSeconds.value());
+    }
+
+    return py::none();
+}
+
 /*
  *
  * bindings
@@ -788,6 +832,8 @@ bool usingMacOS13OrHigher() {
         .def("predict", &Model::predict)
         .def("batchPredict", &Model::batchPredict)
         .def("get_compiled_model_path", &Model::getCompiledModelPath)
+        .def("get_load_duration_in_nano_seconds", &Model::getLoadDurationInNanoSeconds)
+        .def("get_last_predict_duration_in_nano_seconds", &Model::getLastPredictDurationInNanoSeconds)
         .def_static("auto_set_specification_version", &Model::autoSetSpecificationVersion)
         .def_static("maximum_supported_specification_version", &Model::maximumSupportedSpecificationVersion)
 #if BUILT_WITH_MACOS15_SDK
@@ -804,14 +850,18 @@ bool usingMacOS13OrHigher() {
     py::class_<State>(m, "_State", py::module_local());
 
 #if ML_COMPUTE_DEVICE_IS_AVAILABLE
-    py::class_<CPUComputeDevice>(m, "_MLCPUComputeDeviceProxy", py::module_local());
-    py::class_<GPUComputeDevice>(m, "_MLGPUComputeDeviceProxy", py::module_local());
+    py::class_<CPUComputeDevice>(m, "_MLCPUComputeDeviceProxy", py::module_local())
+       .def(py::init());
+    py::class_<GPUComputeDevice>(m, "_MLGPUComputeDeviceProxy", py::module_local())
+       .def(py::init());
     py::class_<NeuralEngineComputeDevice>(m, "_MLNeuralEngineComputeDeviceProxy", py::module_local())
+        .def(py::init())
         .def("get_total_core_count", &NeuralEngineComputeDevice::getTotalCoreCount);
 #endif
 
 #if ML_COMPUTE_PLAN_IS_AVAILABLE
     py::class_<ComputePlan>(m, "_MLComputePlanProxy", py::module_local())
+        .def(py::init())
         .def_property_readonly("model_structure", &ComputePlan::getModelStructure)
         .def("get_compute_device_usage_for_mlprogram_operation", &ComputePlan::getComputeDeviceUsageForMLProgramOperation)
         .def("get_compute_device_usage_for_neuralnetwork_layer", &ComputePlan::getComputeDeviceUsageForNeuralNetworkLayer)

coremltools/__init__.py

@@ -116,9 +116,9 @@ class SpecializationStrategy(_Enum):
 _LOWEST_ALLOWED_SPECIFICATION_VERSION_FOR_NEURALNETWORK = _SPECIFICATION_VERSION_IOS_13
 _LOWEST_ALLOWED_SPECIFICATION_VERSION_FOR_MILPROGRAM = _SPECIFICATION_VERSION_IOS_15
 
-
 # expose sub packages as directories
 from . import converters, models, optimize, proto
+
 # expose unified converter in coremltools package level
 from .converters import ClassifierConfig
 from .converters import ColorLayout as colorlayout

coremltools/_deps/__init__.py

@@ -153,7 +153,7 @@ def __get_sklearn_version(version):
 
 # ---------------------------------------------------------------------------------------
 _HAS_TORCH = True
-_TORCH_MAX_VERSION = "2.4.0"
+_TORCH_MAX_VERSION = "2.5.0"
 _HAS_TORCH_EXPORT_API = False
 _CT_OPTIMIZE_TORCH_MIN_VERSION = "2.1.0"
 _IMPORT_CT_OPTIMIZE_TORCH = False

coremltools/converters/libsvm/_libsvm_converter.py

@@ -3,8 +3,9 @@
 # Use of this source code is governed by a BSD-3-clause license that can be
 # found in the LICENSE.txt file or at https://opensource.org/licenses/BSD-3-Clause
 
+import coremltools as ct
 from coremltools import __version__ as ct_version
-from coremltools.models import _METADATA_SOURCE, _METADATA_VERSION
+from coremltools import proto
 
 from ... import SPECIFICATION_VERSION
 from ..._deps import _HAS_LIBSVM
@@ -58,12 +59,11 @@ def convert(libsvm_model, feature_names, target, input_length, probability):
     from libsvm import svm as _svm
 
     from ...models import MLModel
-    from ...proto import Model_pb2
 
     svm_type_enum = libsvm_model.param.svm_type
 
     # Create the spec
-    export_spec = Model_pb2.Model()
+    export_spec = proto.Model_pb2.Model()
     export_spec.specificationVersion = SPECIFICATION_VERSION
 
     if svm_type_enum == _svm.EPSILON_SVR or svm_type_enum == _svm.NU_SVR:
@@ -90,7 +90,7 @@ def convert(libsvm_model, feature_names, target, input_length, probability):
         input = export_spec.description.input.add()
         input.name = feature_names
         input.type.multiArrayType.shape.append(input_length)
-        input.type.multiArrayType.dataType = Model_pb2.ArrayFeatureType.DOUBLE
+        input.type.multiArrayType.dataType = proto.Model_pb2.ArrayFeatureType.DOUBLE
 
     else:
         # input will be a series of doubles
@@ -193,7 +193,7 @@ def convert(libsvm_model, feature_names, target, input_length, probability):
     from libsvm import __version__ as libsvm_version
 
     libsvm_version = "libsvm=={0}".format(libsvm_version)
-    model.user_defined_metadata[_METADATA_VERSION] = ct_version
-    model.user_defined_metadata[_METADATA_SOURCE] = libsvm_version
+    model.user_defined_metadata[ct.models._METADATA_VERSION] = ct_version
+    model.user_defined_metadata[ct.models._METADATA_SOURCE] = libsvm_version
 
     return model