7.0b1 release (#1874)

BUILDING.md

@@ -19,7 +19,7 @@ Follow these steps:
 1. Fork and clone the GitHub [coremltools repository](https://github.com/apple/coremltools).
 
 2. Run the [build.sh](scripts/build.sh) script to build `coremltools`.
-	* By default this script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`) as a argument to change the Python version.
+	* By default this script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`, `3.11`) as a argument to change the Python version.
 	* The script creates a new `build` folder with the coremltools distribution, and a `dist` folder with Python wheel files.
 
 3. Run the [test.sh](scripts/test.sh) script to test the build.
@@ -45,7 +45,7 @@ The following build targets help you configure the development environment. If y
 * `test_slow` | Run all non-fast tests.
 * `wheel` | Build wheels in release mode.
 
-The script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`) as a argument to change the Python version.
+The script uses Python 3.7, but you can include `--python=3.8` (or `3.9`, `3.10`, `3.11`) as a argument to change the Python version.
 
 ## Resources
 

CMakeLists.txt

@@ -191,6 +191,19 @@ else()
   message(STATUS "CoreML.framework and dependent frameworks not found. Skipping libcoremlpython build.")
 endif()
 
+# Build kmeans-1d
+set(KMEANS_DIR "${PROJECT_SOURCE_DIR}/deps/kmeans1d")
+execute_process(
+  COMMAND python3 setup.py build_ext --inplace
+  WORKING_DIRECTORY ${KMEANS_DIR}
+)
+
+# Copy kmeans-1d to Python deps folder
+execute_process(
+  COMMAND cp -r kmeans1d ../../coremltools/_deps
+  WORKING_DIRECTORY ${KMEANS_DIR}
+)
+
 set(PYTHON_TAG "cp${PYTHON_VERSION_MAJOR}${PYTHON_VERSION_MINOR}")
 if(APPLE)
   execute_process(COMMAND uname -m OUTPUT_VARIABLE HARDWARE_NAME OUTPUT_STRIP_TRAILING_WHITESPACE)

LICENSE.txt

@@ -1,4 +1,4 @@
-Copyright (c) 2020, Apple Inc. All rights reserved.
+Copyright © 2020-2023, Apple Inc. All rights reserved.
 
 Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:  
 

NOTICE.txt

@@ -0,0 +1,25 @@
+Copyright © 2020-2023, Apple Inc. All rights reserved.
+
+This project contains content adapted from kmeans1d (https://github.com/dstein64/kmeans1d), the license for which follows:
+
+MIT License
+
+Copyright (c) 2019 Daniel Steinberg
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

coremltools/__init__.py

@@ -60,6 +60,10 @@
 # New versions for iOS 16.0
 _SPECIFICATION_VERSION_IOS_16 = 7
 
+# New versions for iOS 17.0
+_SPECIFICATION_VERSION_IOS_17 = 8
+
+
 class ComputeUnit(_Enum):
     '''
     The set of processing-unit configurations the model can use to make predictions.
@@ -76,6 +80,7 @@ class ComputeUnit(_Enum):
     _SPECIFICATION_VERSION_IOS_14: "CoreML4",
     _SPECIFICATION_VERSION_IOS_15: "CoreML5",
     _SPECIFICATION_VERSION_IOS_16: "CoreML6",
+    _SPECIFICATION_VERSION_IOS_17: "CoreML7",
 }
 
 # Default specification version for each backend
@@ -84,7 +89,7 @@ class ComputeUnit(_Enum):
 
 
 # expose sub packages as directories
-from . import converters, models, proto
+from . import converters, models, optimize, proto
 
 # expose unified converter in coremltools package level
 from .converters import ClassifierConfig

coremltools/_deps/.gitignore

@@ -0,0 +1 @@
+kmeans1d/

coremltools/_deps/__init__.py

@@ -16,6 +16,13 @@
 
 from coremltools import _logger as logger
 
+_HAS_KMEANS1D = True
+try:
+    from . import kmeans1d as _kmeans1d
+except:
+    _kmeans1d = None
+    _HAS_KMEANS1D = False
+
 
 def _get_version(version):
     # matching 1.6.1, and 1.6.1rc, 1.6.1.dev
@@ -156,6 +163,14 @@ def __get_sklearn_version(version):
 MSG_TORCH_NOT_FOUND = "PyTorch not found."
 
 
+_HAS_TORCH_VISION = True
+try:
+    import torchvision
+except:
+    _HAS_TORCH_VISION = False
+MSG_TORCH_VISION_NOT_FOUND = "TorchVision not found."
+
+
 # ---------------------------------------------------------------------------------------
 try:
     import scipy

coremltools/converters/_converters_entry.py

@@ -6,7 +6,7 @@
 import collections
 import gc
 import os
-from typing import Optional, Text, Union
+from typing import List, Optional, Text, Union
 
 from coremltools import (
     _LOWEST_ALLOWED_SPECIFICATION_VERSION_FOR_MILPROGRAM,
@@ -23,8 +23,11 @@
 from coremltools.converters.mil.converter import mil_convert
 from coremltools.converters.mil.input_types import (
     ClassifierConfig,
+    EnumeratedShapes,
     ImageType,
     InputType,
+    RangeDim,
+    Shape,
     TensorType,
 )
 from coremltools.converters.mil.mil import Program, types
@@ -395,7 +398,7 @@ def skip_real_div_ops(op):
 
             pipeline = ct.PassPipeline()
             pipeline.remove_passes({"common::fuse_conv_batchnorm"})
-            ct.convert(model, pass_pipeline=pipeline)
+            mlmodel = ct.convert(model, pass_pipeline=pipeline)
 
         * To avoid folding too-large ``const`` ops that lead to a large model, set pass option
           as shown in the following example:
@@ -404,7 +407,34 @@ def skip_real_div_ops(op):
 
             pipeline = ct.PassPipeline()
             pipeline.set_options("common::const_elimination", {"skip_const_by_size": "1e6"})
-            ct.convert(model, pass_pipeline=pipeline)
+            mlmodel = ct.convert(model, pass_pipeline=pipeline)
+
+        We also provide a set of predefined pass pipelines that you can directly call.
+
+        * To avoid running all graph pass, you can use:
+
+          .. sourcecode:: python
+
+             mlmodel = ct.convert(model, pass_pipeline=ct.PassPipeline.EMPTY)
+
+        * To only run the cleanup graph passes, like constant_elimination, dead_code_elimination, etc.
+          You can use:
+
+          .. sourcecode:: python
+
+             mlmodel = ct.convert(model, pass_pipeline=ct.PassPipeline.CLEANUP)
+
+        * To convert a source model with sparse weights to a sparse format Core ML model, you can use:
+
+          .. sourcecode:: python
+
+             mlmodel = ct.convert(model, pass_pipeline=ct.PassPipeline.DEFAULT_PRUNING)
+
+        * To convert a source model with palettized weights to a compressed format Core ML model, you can use:
+
+          .. sourcecode:: python
+
+             mlmodel = ct.convert(model, pass_pipeline=ct.PassPipeline.DEFAULT_PALETTIZATION)
 
     Returns
     -------
@@ -463,9 +493,17 @@ def skip_real_div_ops(op):
                                      outputs_as_tensor_or_image_types,
                                      outputs)
     exact_target = _determine_target(convert_to, minimum_deployment_target)
-    _validate_conversion_arguments(model, exact_source, inputs, outputs_as_tensor_or_image_types,
-                                   classifier_config, compute_precision,
-                                   exact_target, minimum_deployment_target)
+    _validate_conversion_arguments(
+        model,
+        exact_source,
+        exact_target,
+        inputs,
+        outputs_as_tensor_or_image_types,
+        classifier_config,
+        compute_precision,
+        exact_target,
+        minimum_deployment_target,
+    )
 
     if pass_pipeline is None:
         pass_pipeline = PassPipeline()
@@ -504,6 +542,12 @@ def skip_real_div_ops(op):
         main_pipeline=pass_pipeline,
     )
 
+    if exact_target == "mlprogram" and mlmodel._input_has_infinite_upper_bound():
+        raise ValueError(
+            "For mlprogram, inputs with infinite upper_bound is not allowed. Please set upper_bound"
+            ' to a positive value in "RangeDim()" for the "inputs" param in ct.convert().'
+        )
+
     if exact_target == 'milinternal':
         return mlmodel  # Returns the MIL program
 
@@ -539,7 +583,7 @@ def _need_fp16_cast_pass(
         raise ValueError(f"Invalid value of the argument 'compute_precision': {compute_precision}")
 
 
-def _set_default_specification_version(target):
+def _set_default_specification_version(target) -> Optional[AvailableTarget]:
     if target == "neuralnetwork":
         return _LOWEST_ALLOWED_SPECIFICATION_VERSION_FOR_NEURALNETWORK
     elif target == "mlprogram":
@@ -625,18 +669,20 @@ def _validate_outputs_argument(outputs):
                 return output_names, outputs
 
 
-def _validate_conversion_arguments(model,
-                                   exact_source,
-                                   inputs,
-                                   outputs,
-                                   classifier_config,
-                                   compute_precision,
-                                   convert_to,
-                                   minimum_deployment_target,
-                                   ):
+def _validate_conversion_arguments(
+    model,
+    exact_source,
+    exact_target,
+    inputs,
+    outputs,
+    classifier_config,
+    compute_precision,
+    convert_to,
+    minimum_deployment_target,
+):
     """
     Validate and process model, inputs, classifier_config based on
-    `exact_source` (which cannot be `auto`)
+    `exact_source` (which cannot be `auto`) and `exact_target`.
     """
 
     def raise_if_duplicated(input_list):
@@ -672,10 +718,10 @@ def _flatten_list(_inputs):
 
         # get flattened inputs
         flat_inputs = _flatten_list(inputs)
-        for t in flat_inputs:
-            if not isinstance(t, InputType):
+        for flat_input in flat_inputs:
+            if not isinstance(flat_input, InputType):
                 raise ValueError("inputs must be a list of type ct.TensorType or ct.ImageType")
-            if t.dtype == types.fp16:
+            if flat_input.dtype == types.fp16:
                 if not (
                     minimum_deployment_target is not None
                     and minimum_deployment_target >= AvailableTarget.iOS16
@@ -685,6 +731,24 @@ def _flatten_list(_inputs):
                         "target >= iOS16/macOS13/watchOS9/tvOS16"
                     )
 
+    if exact_target == "mlprogram":
+        err_msg_infinite_bound = (
+            "For mlprogram, inputs with infinite upper_bound is not allowed. Please set upper_bound"
+            ' to a positive value in "RangeDim()" for the "inputs" param in ct.convert().'
+        )
+        if inputs is not None:
+            for flat_input in _flatten_list(inputs):
+                tensor_shapes: List[Optional[Shape]] = (
+                    flat_input.shape.shapes
+                    if isinstance(flat_input.shape, EnumeratedShapes)
+                    else [flat_input.shape]
+                )
+                for tensor_shape in tensor_shapes:
+                    if tensor_shape is not None:
+                        for shape in tensor_shape.shape:
+                            if isinstance(shape, RangeDim) and shape.upper_bound < 0:
+                                raise ValueError(err_msg_infinite_bound)
+
     if outputs is not None:
         for t in outputs:
             if t.dtype == types.fp16:

coremltools/converters/mil/_deployment_compatibility.py

@@ -5,10 +5,13 @@
 
 from enum import IntEnum
 
-from coremltools import (_SPECIFICATION_VERSION_IOS_13,
-                         _SPECIFICATION_VERSION_IOS_14,
-                         _SPECIFICATION_VERSION_IOS_15,
-                         _SPECIFICATION_VERSION_IOS_16)
+from coremltools import (
+    _SPECIFICATION_VERSION_IOS_13,
+    _SPECIFICATION_VERSION_IOS_14,
+    _SPECIFICATION_VERSION_IOS_15,
+    _SPECIFICATION_VERSION_IOS_16,
+    _SPECIFICATION_VERSION_IOS_17,
+)
 
 
 class AvailableTarget(IntEnum):
@@ -17,6 +20,7 @@ class AvailableTarget(IntEnum):
     iOS14 = _SPECIFICATION_VERSION_IOS_14
     iOS15 = _SPECIFICATION_VERSION_IOS_15
     iOS16 = _SPECIFICATION_VERSION_IOS_16
+    iOS17 = _SPECIFICATION_VERSION_IOS_17
 
     # macOS versions (aliases of iOS versions)
     macOS15 = _SPECIFICATION_VERSION_IOS_13
@@ -26,19 +30,22 @@ class AvailableTarget(IntEnum):
     macOS11 = _SPECIFICATION_VERSION_IOS_14
     macOS12 = _SPECIFICATION_VERSION_IOS_15
     macOS13 = _SPECIFICATION_VERSION_IOS_16
+    macOS14 = _SPECIFICATION_VERSION_IOS_17
 
     # watchOS versions (aliases of iOS versions)
     watchOS6 = _SPECIFICATION_VERSION_IOS_13
     watchOS7 = _SPECIFICATION_VERSION_IOS_14
     watchOS8 = _SPECIFICATION_VERSION_IOS_15
     watchOS9 = _SPECIFICATION_VERSION_IOS_16
+    watchOS10 = _SPECIFICATION_VERSION_IOS_17
 
     # tvOS versions (aliases of iOS versions)
     tvOS13 = _SPECIFICATION_VERSION_IOS_13
     tvOS14 = _SPECIFICATION_VERSION_IOS_14
     tvOS15 = _SPECIFICATION_VERSION_IOS_15
     tvOS16 = _SPECIFICATION_VERSION_IOS_16
-
+    tvOS17 = _SPECIFICATION_VERSION_IOS_17
+
     # customized __str__
     def __str__(self):
         original_str = super().__str__()