[ExecuTorch] Model Coverage, Partitioner, Quantizer, Debug Handle

.gitlab-ci.yml

@@ -114,7 +114,7 @@ test_py39_tf2_intel-1:
     PYTHON: "3.9"
     TEST_PACKAGE: coremltools.converters.mil.frontend.tensorflow
     WHEEL_PATH: build/dist/*cp39*10_15*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_tf2.pip
 
 test_py39_tf2_intel-2:
   <<: *test_macos_pkg
@@ -126,7 +126,7 @@ test_py39_tf2_intel-2:
     PYTHON: "3.9"
     TEST_PACKAGE: coremltools.converters.mil.frontend.tensorflow2
     WHEEL_PATH: build/dist/*cp39*10_15*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_tf2.pip
 
 test_py39_mil_intel:
   <<: *test_macos_pkg
@@ -174,7 +174,7 @@ test_py39_milproto_intel:
     WHEEL_PATH: build/dist/*cp39*10_15*
     TEST_PACKAGE: coremltools.converters.mil.frontend.milproto
     PYTHON: "3.9"
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_tf2.pip
 
 
 
@@ -212,7 +212,7 @@ test_py310_tf2-1:
     PYTHON: "3.10"
     TEST_PACKAGE: coremltools.converters.mil.frontend.tensorflow
     WHEEL_PATH: build/dist/*cp310*11*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_tf2.pip
 
 test_py310_tf2-2:
   <<: *test_macos_pkg
@@ -224,7 +224,7 @@ test_py310_tf2-2:
     PYTHON: "3.10"
     TEST_PACKAGE: coremltools.converters.mil.frontend.tensorflow2
     WHEEL_PATH: build/dist/*cp310*11*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_tf2.pip
 
 test_py310_mil:
   <<: *test_macos_pkg
@@ -272,7 +272,7 @@ test_py310_milproto:
     PYTHON: "3.10"
     TEST_PACKAGE: coremltools.converters.mil.frontend.milproto
     WHEEL_PATH: build/dist/*cp310*11*
-    REQUIREMENTS: reqs/test.pip
+    REQUIREMENTS: reqs/test_tf2.pip
 
 
 

coremlpython/CoreMLPython.h

@@ -3,6 +3,8 @@
 // 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
 
+
+// Disable a few warnings and include pybind first, then re-enable warnings
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wexit-time-destructors"
 #pragma clang diagnostic ignored "-Wdocumentation"
@@ -28,11 +30,13 @@ namespace CoreML {
             static py::bytes autoSetSpecificationVersion(const py::bytes& modelBytes);
             static py::str compileModel(const std::string& urlStr);
             static int32_t maximumSupportedSpecificationVersion();
+            static void setComputeUnit(MLModelConfiguration *configuration, const std::string& computeUnits);
 
             Model(const Model&) = delete;
             Model& operator=(const Model&) = delete;
             ~Model();
             explicit Model(const std::string& urlStr, const std::string& computeUnits);
+            explicit Model(MLModel* m_model, NSURL* compiledUrl, bool deleteCompiledModelOnExit);
 
             py::dict predict(const py::dict& input) const;
             py::list batchPredict(const py::list& batch) const;

coremlpython/CoreMLPython.mm

@@ -33,18 +33,6 @@ bool usingMacOS13OrHigher() {
     return (NSProtocolFromString(@"MLProgram") != nil);
 }
 
-bool isCompiledModelPath(const std::string& path) {
-    const std::string fileExtension = ".mlmodelc";
-
-    size_t start = path.length() - fileExtension.length();
-    if (path.back() == '/') {
-        start--;
-    }
-    const std::string match = path.substr(start, fileExtension.length());
-
-    return (match == fileExtension);
-}
-
 Model::~Model() {
     @autoreleasepool {
         NSFileManager *fileManager = [NSFileManager defaultManager];
@@ -58,7 +46,7 @@ bool isCompiledModelPath(const std::string& path) {
     @autoreleasepool {
         NSError *error = nil;
 
-        if (! isCompiledModelPath(urlStr)) {
+        if (! Utils::isCompiledModelPath(urlStr)) {
             // Compile the model
             NSURL *specUrl = Utils::stringToNSURL(urlStr);
 
@@ -89,31 +77,23 @@ bool isCompiledModelPath(const std::string& path) {
             compiledUrl = Utils::stringToNSURL(urlStr);
         }
 
-        // Set compute unit
         MLModelConfiguration *configuration = [MLModelConfiguration new];
-        if (computeUnits == "CPU_ONLY") {
-            configuration.computeUnits = MLComputeUnitsCPUOnly;
-        } else if (computeUnits == "CPU_AND_GPU") {
-            configuration.computeUnits = MLComputeUnitsCPUAndGPU;
-        } else if (computeUnits == "CPU_AND_NE") {
-            if (usingMacOS13OrHigher()) {
-#if BUILT_WITH_MACOS13_SDK
-                configuration.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
-#endif // BUILT_WITH_MACOS13_SDK
-            } else {
-                throw std::runtime_error("CPU_AND_NE is only available on macOS >= 13.0");
-            }
-        } else {
-            assert(computeUnits == "ALL");
-            configuration.computeUnits = MLComputeUnitsAll;
-        }
+        setComputeUnit(configuration, computeUnits);
 
         // Create MLModel
         m_model = [MLModel modelWithContentsOfURL:compiledUrl configuration:configuration error:&error];
         Utils::handleError(error);
     }
 }
 
+
+Model::Model(MLModel* mlModel, NSURL* compiledUrl, bool deleteCompiledModelOnExit)
+    : m_model(mlModel),
+      compiledUrl(compiledUrl),
+      m_deleteCompiledModelOnExit(deleteCompiledModelOnExit)
+{
+}
+
 py::dict Model::predict(const py::dict& input) const {
     @autoreleasepool {
         NSError *error = nil;
@@ -127,6 +107,26 @@ bool isCompiledModelPath(const std::string& path) {
 }
 
 
+void Model::setComputeUnit(MLModelConfiguration *configuration, const std::string& computeUnits) {
+    if (computeUnits == "CPU_ONLY") {
+        configuration.computeUnits = MLComputeUnitsCPUOnly;
+    } else if (computeUnits == "CPU_AND_GPU") {
+        configuration.computeUnits = MLComputeUnitsCPUAndGPU;
+    } else if (computeUnits == "CPU_AND_NE") {
+        if (usingMacOS13OrHigher()) {
+#if BUILT_WITH_MACOS13_SDK
+            configuration.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
+#endif // BUILT_WITH_MACOS13_SDK
+        } else {
+            throw std::runtime_error("CPU_AND_NE is only available on macOS >= 13.0");
+        }
+    } else {
+        assert(computeUnits == "ALL");
+        configuration.computeUnits = MLComputeUnitsAll;
+    }
+}
+
+
 py::list Model::batchPredict(const py::list& batch) const {
   @autoreleasepool {
       NSError* error = nil;
@@ -156,6 +156,9 @@ bool isCompiledModelPath(const std::string& path) {
 
 
 py::str Model::getCompiledModelPath() const {
+    if (this->compiledUrl == nil) {
+        return nil;
+    }
     return [this->compiledUrl.path UTF8String];
 }
 

coremlpython/CoreMLPythonUtils.h

@@ -23,6 +23,7 @@ namespace CoreML {
     namespace Python {
         namespace Utils {
 
+            bool isCompiledModelPath(const std::string& path);
             NSURL * stringToNSURL(const std::string& str);
             void handleError(NSError *error);
 

coremlpython/CoreMLPythonUtils.mm

@@ -29,6 +29,18 @@
 
 using namespace CoreML::Python;
 
+bool Utils::isCompiledModelPath(const std::string& path) {
+    const std::string fileExtension = ".mlmodelc";
+
+    size_t start = path.length() - fileExtension.length();
+    if (path.back() == '/') {
+        start--;
+    }
+    const std::string match = path.substr(start, fileExtension.length());
+
+    return (match == fileExtension);
+}
+
 NSURL * Utils::stringToNSURL(const std::string& str) {
     NSString *nsstr = [NSString stringWithUTF8String:str.c_str()];
     return [NSURL fileURLWithPath:nsstr];

coremltools/__init__.py

@@ -21,6 +21,7 @@
 
 For more information: http://developer.apple.com/documentation/coreml
 """
+
 from enum import Enum as _Enum
 from logging import getLogger as _getLogger
 
@@ -90,15 +91,14 @@ class ComputeUnit(_Enum):
 
 # 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
 from .converters import EnumeratedShapes, ImageType, RangeDim, Shape, TensorType, convert
 from .converters.mil._deployment_compatibility import AvailableTarget as target
 from .converters.mil.mil.passes.defs import quantization as transform
-from .converters.mil.mil.passes.pass_pipeline import PassPipeline
 from .converters.mil.mil.passes.defs.quantization import ComputePrecision as precision
+from .converters.mil.mil.passes.pass_pipeline import PassPipeline
 from .models import utils
 from .models.ml_program import compression_utils
 

coremltools/_deps/__init__.py

@@ -154,7 +154,7 @@ def __get_sklearn_version(version):
 
 # ---------------------------------------------------------------------------------------
 _HAS_TORCH = True
-_TORCH_MAX_VERSION = "2.1.0"
+_TORCH_MAX_VERSION = "2.2.0"
 _HAS_TORCH_EXPORT_API = False
 try:
     import torch