Merge pull request #584 from 1duo/master

Mark coremltools 3.2

Author: 1duo

README.md

@@ -74,7 +74,7 @@ sure to install virtualenv using your system pip.
 pip install virtualenv
 ```
 
-The method for installing *coremltools* follows the
+The method for installing `coremltools` follows the
 [standard python package installation steps](https://packaging.python.org/installing/).
 To create a Python virtual environment called `pythonenv` follow these steps:
 
@@ -88,6 +88,7 @@ virtualenv pythonenv
 ```
 
 To activate your new virtual environment and install `coremltools` in this environment, follow these steps:
+
 ```
 # Active your virtual environment
 source pythonenv/bin/activate
@@ -215,6 +216,7 @@ Finally, to run the most important unit tests, you can use:
 ```shell
 pytest -rs
 ```
+
 some tests are marked as slow because they test a lot of combinations.
 If you want to run, all tests, you can use:
 

coremltools/__init__.py

@@ -22,7 +22,7 @@
 For more information: http://developer.apple.com/documentation/coreml
 """
 
-__version__ = '3.1'
+__version__ = '3.2'
 
 # This is the basic Core ML specification format understood by iOS 11.0
 SPECIFICATION_VERSION = 1

coremltools/converters/caffe/_caffe_converter.py

@@ -3,7 +3,7 @@
 # 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 sys as _sys
+import os
 import six as _six
 from ...models import _MLMODEL_FULL_PRECISION, _MLMODEL_HALF_PRECISION, _VALID_MLMODEL_PRECISION_TYPES
 
@@ -191,6 +191,10 @@ def convert(model, image_input_names=[], is_bgr=False,
             green_bias, gray_bias, image_scale, class_labels,
             predicted_feature_name)
     model = MLModel(model_path)
+    try:
+        os.remove(model_path)
+    except OSError:
+        pass
 
     if model_precision == _MLMODEL_HALF_PRECISION and model is not None:
         model = convert_neural_network_weights_to_fp16(model)

coremltools/converters/keras/_keras2_converter.py

@@ -10,6 +10,7 @@
 from ...models.neural_network.update_optimizer_utils import SgdParams
 from ...proto import FeatureTypes_pb2 as _FeatureTypes_pb2
 from collections import OrderedDict as _OrderedDict
+from ...proto import Model_pb2 as _Model_pb2
 from ...models import datatypes
 from ...models import MLModel as _MLModel
 from ...models.utils import save_spec as _save_spec
@@ -32,7 +33,7 @@
 
         _keras.layers.convolutional.Conv2D: _layers2.convert_convolution,
         _keras.layers.convolutional.Conv2DTranspose: _layers2.convert_convolution,
-        _keras.layers.convolutional.SeparableConv2D: _layers2.convert_separable_convolution, 
+        _keras.layers.convolutional.SeparableConv2D: _layers2.convert_separable_convolution,
         _keras.layers.pooling.AveragePooling2D: _layers2.convert_pooling,
         _keras.layers.pooling.MaxPooling2D: _layers2.convert_pooling,
         _keras.layers.pooling.GlobalAveragePooling2D: _layers2.convert_pooling,
@@ -63,7 +64,7 @@
         _keras.layers.Maximum: _layers2.convert_merge,
         _keras.layers.Concatenate: _layers2.convert_merge,
         _keras.layers.Dot: _layers2.convert_merge,
-    
+
         _keras.layers.core.Flatten: _layers2.convert_flatten,
         _keras.layers.core.Permute:_layers2.convert_permute,
         _keras.layers.core.Reshape:_layers2.convert_reshape,
@@ -91,7 +92,7 @@
          _KERAS_LAYER_REGISTRY[_keras.engine.topology.InputLayer] = \
              _layers2.default_skip
     # end if _HAS_KERAS2_TF
-    
+
 
 def _is_merge_layer(layer):
     if _HAS_KERAS2_TF:
@@ -106,7 +107,7 @@ def _is_activation_layer(layer):
             isinstance(layer, _keras.layers.advanced_activations.LeakyReLU) or
             isinstance(layer, _keras.layers.advanced_activations.PReLU) or
             isinstance(layer, _keras.layers.advanced_activations.ELU) or
-            isinstance(layer, 
+            isinstance(layer,
                        _keras.layers.advanced_activations.ThresholdedReLU) or
             isinstance(layer, _keras.layers.advanced_activations.Softmax))
 
@@ -163,7 +164,7 @@ def _load_keras_model(model_network_path, model_weight_path, custom_objects=None
         Path where the model network weights are (hd5 file)
 
     custom_objects:
-        A dictionary of layers or other custom classes 
+        A dictionary of layers or other custom classes
         or functions used by the model
 
     Returns
@@ -300,7 +301,8 @@ def _convert(model,
              custom_objects=None,
              input_shapes=None,
              output_shapes=None,
-             respect_trainable=False):
+             respect_trainable=False,
+             use_float_arraytype=False):
 
     # Check Keras format
     if _keras.backend.image_data_format() == 'channels_first':
@@ -309,48 +311,48 @@ def _convert(model,
               "Changing to 'channels_last', but your model may not be converted "
               "converted properly.")
         _keras.backend.set_image_data_format('channels_last')
-    
+
     # Check custom conversion functions / custom objects
     add_custom_layers = custom_conversion_functions is not None
 
     if isinstance(model, _string_types):
         model = _keras.models.load_model(model, custom_objects = custom_objects)
     elif isinstance(model, tuple):
         model = _load_keras_model(model[0], model[1])
-    
+
     # Check valid versions
     _check_unsupported_layers(model, add_custom_layers)
-    
+
     # Build network graph to represent Keras model
     graph = _topology2.NetGraph(model)
     graph.build()
 
     # The graph should be finalized before executing this
     graph.generate_blob_names()
     graph.add_recurrent_optionals()
-    
+
     inputs = graph.get_input_layers()
     outputs = graph.get_output_layers()
-    
+
     # check input / output names validity
-    if input_names is not None: 
+    if input_names is not None:
         if isinstance(input_names, _string_types):
             input_names = [input_names]
-    else: 
+    else:
         input_names = ['input' + str(i+1) for i in range(len(inputs))]
 
-    if output_names is not None: 
+    if output_names is not None:
         if isinstance(output_names, _string_types):
             output_names = [output_names]
-    else: 
+    else:
         output_names = ['output' + str(i+1) for i in range(len(outputs))]
-    
+
     if image_input_names is not None and isinstance(image_input_names, _string_types):
         image_input_names = [image_input_names]
-    
+
     graph.reset_model_input_names(input_names)
     graph.reset_model_output_names(output_names)
-    
+
     # Keras -> Core ML input dimension dictionary
     # (None, None) -> [1, 1, 1, 1, 1]
     # (None, D) -> [D] or [D, 1, 1, 1, 1]
@@ -478,7 +480,8 @@ def _convert(model,
     input_features = list(zip(input_names, input_types))
     output_features = list(zip(output_names, output_types))
 
-    builder = _NeuralNetworkBuilder(input_features, output_features, mode = mode)
+    builder = _NeuralNetworkBuilder(input_features, output_features, mode = mode,
+                                    use_float_arraytype=use_float_arraytype)
 
     for iter, layer in enumerate(graph.layer_list):
         keras_layer = graph.keras_layer_map[layer]
@@ -548,4 +551,21 @@ def _convert(model,
 
     # Return the protobuf spec
     spec = builder.spec
+
+    # If the model has multi-arrays of type double, recommend to the user the utility function
+    # coremltools.models.utils.convert_double_to_float_multiarray_type(spec)
+    has_double_multiarray = False
+    for feature in list(spec.description.input) + list(spec.description.output):
+        if feature.type.HasField('multiArrayType'):
+            if feature.type.multiArrayType.dataType == _Model_pb2.ArrayFeatureType.DOUBLE:
+                has_double_multiarray = True
+                break
+
+    if has_double_multiarray:
+        print("\n\nRecommendation: This model has at least one multiarray input/output of type double.\n"
+              "For large sized arrays, multiarrays of type float32 are more efficient.\n"
+              "In future, float input/output multiarrays will be produced by default by the converter.\n"
+              "Please use, either the flag 'use_float_arraytype' during the call to convert or\n"
+              "the utility 'coremltools.utils.convert_double_to_float_multiarray_type(spec)', post-conversion.\n\n")
+
     return spec

coremltools/converters/keras/_keras_converter.py

@@ -361,7 +361,8 @@ def convertToSpec(model,
                   custom_objects=None,
                   input_shapes = None,
                   output_shapes = None,
-                  respect_trainable = False):
+                  respect_trainable = False,
+                  use_float_arraytype = False):
 
     """
     Convert a Keras model to Core ML protobuf specification (.mlmodel).
@@ -485,6 +486,10 @@ def convertToSpec(model,
         If True, then Keras layers that are marked 'trainable' will
         automatically be marked updatable in the Core ML model.
 
+    use_float_arraytype: bool
+        If true, the datatype of input/output multiarrays is set to Float32 instead
+        of double.
+
     Returns
     -------
     model: MLModel
@@ -576,7 +581,8 @@ def convertToSpec(model,
                                           custom_objects=custom_objects,
                                           input_shapes=input_shapes,
                                           output_shapes=output_shapes,
-                                          respect_trainable=respect_trainable)
+                                          respect_trainable=respect_trainable,
+                                          use_float_arraytype=use_float_arraytype)
     else:
         raise RuntimeError(
             'Keras not found or unsupported version or backend found. keras conversion API is disabled.')
@@ -606,7 +612,8 @@ def convert(model,
                   custom_conversion_functions = None,
                   input_shapes = None,
                   output_shapes = None,
-                  respect_trainable = False):
+                  respect_trainable = False,
+                  use_float_arraytype = False):
 
     """
     Convert a Keras model to Core ML protobuf specification (.mlmodel).
@@ -719,6 +726,10 @@ def convert(model,
         If yes, then Keras layers marked 'trainable' will automatically be
         marked updatable in the Core ML model.
 
+    use_float_arraytype: bool
+        If true, the datatype of input/output multiarrays is set to Float32 instead
+        of double.
+
     Returns
     -------
     model: MLModel
@@ -789,6 +800,7 @@ def convert(model,
                          custom_conversion_functions=custom_conversion_functions,
                          input_shapes=input_shapes,
                          output_shapes=output_shapes,
-                         respect_trainable=respect_trainable)
+                         respect_trainable=respect_trainable,
+                         use_float_arraytype=use_float_arraytype)
 
     return _MLModel(spec)

coremltools/converters/nnssa/commons/basic_graph_ops.py

@@ -153,7 +153,7 @@ def check_connections(gd):
         for i in v.control_inputs:
             assert (k in gd[i].control_outputs)
         for i in v.control_outputs:
-            message = f"Node {k} not in {i} control_inputs"
+            message = "Node " + k + " not in " + i + " control_inputs"
             assert (k in gd[i].control_inputs), message
 
 

coremltools/converters/nnssa/coreml/graph_pass/__init__.py

@@ -3,14 +3,6 @@
 from __future__ import division as _
 from __future__ import absolute_import as _
 
-from .op_removals import remove_no_ops_and_shift_control_dependencies
-from .op_removals import constant_weight_link_removal
-from .op_removals import remove_single_isolated_node
-from .op_removals import remove_identity, remove_oneway_split
-
-from .op_fusions import fuse_bias_add, transform_nhwc_to_nchw, \
-    onehot_matmul_to_embedding, fuse_layer_norm, fuse_gelu, \
-    fuse_batch_norm, fuse_pad_into_conv, \
-    spatial_reduce_to_global_pool, fuse_batch_to_space_or_space_to_batch
-
-from .mlmodel_passes import remove_disconnected_layers
+from .op_removals import *
+from .op_fusions import *
+from .mlmodel_passes import *

coremltools/converters/nnssa/coreml/graph_pass/op_fusions.py

@@ -285,8 +285,8 @@ def transform_nhwc_to_nchw(nnssa):
                 inp_node_format = graph[inp_node_name].attr.get('data_format')
                 symbolic_value = graph[inp_node_name].attr['symbolic_value']
                 if (graph[inp_node_name].op == 'Const' or
-                    len(graph[inp_node_name].datatype.get_shape()) != 4 or
-                    ( symbolic_value and not any_symbolic_or_unknown(symbolic_value))):
+                        len(graph[inp_node_name].datatype.get_shape()) != 4 or
+                        (symbolic_value and not any_symbolic_or_unknown(symbolic_value))):
                     # Const weights and parameters
                     continue
 
@@ -649,7 +649,11 @@ def _match_batch_norm_pattern(graph, entry_node, pattern_ops):
                 return None
             if not _check_number_inputs(node, 2):
                 return None
-            const_node = graph[node.inputs[1]]
+            node_inputs = [graph[n].op.lower() for n in node.inputs]
+            try:
+                const_node = graph[node.inputs[node_inputs.index('const')]]
+            except ValueError:
+                return None
             if not _check_single_out_vector_constant_node(const_node):
                 return None
             if not _check_rank_matches(const_node, node):
@@ -702,7 +706,6 @@ def _merge_batch_norm(graph, nodes, pattern_id=1):
         control_inputs = []
         control_outputs = []
         bn_node_names = [x.name for x in nodes]
-        print(f"\n\nProcessing Fused Batch Norm: {fused_bn_node.name}")
 
         for name in bn_node_names:
             control_inputs += graph[name].control_inputs