def convert_to_mlmodel(model_spec,
tensor_inputs,
backend=("neuralnetwork", "fp32"),
converter_input_type=None,
use_cpu_for_conversion=False):
def _convert_to_inputtype(inputs):
if isinstance(inputs, list):
return [_convert_to_inputtype(x) for x in inputs]
elif isinstance(inputs, tuple):
return tuple([_convert_to_inputtype(x) for x in inputs])
elif isinstance(inputs, TensorType):
return inputs
elif isinstance(inputs, torch.Tensor):
return TensorType(shape=inputs.shape,
dtype=torch_to_mil_types[inputs.dtype])
else:
raise ValueError("Unable to parse type {} into InputType.".format(
type(inputs)))
if converter_input_type is None:
inputs = list(_convert_to_inputtype(tensor_inputs))
else:
inputs = converter_input_type
return ct_convert(model_spec,
inputs=inputs,
convert_to=backend,
source="pytorch",
useCPUOnly=use_cpu_for_conversion)
def tf_graph_to_mlmodel(
graph,
feed_dict,
output_nodes,
frontend="tensorflow",
backend=("neuralnetwork", "fp32"),
use_cpu_for_conversion=False,
):
"""
Parameters
----------
graph: tf.Graph
TensorFlow 1.x model in tf.Graph format.
feed_dict: dict of {tf.placeholder -> np.array or python primitive)
Dict of placeholder and value pairs representing inputs.
output_nodes: tf.node or list[tf.node]
List of names representing outputs.
frontend: str
Frontend to convert from.
backend: str
Backend to convert to.
use_cpu_for_conversion: bool
Argument which is passed as is to the unified converter API.
That is, "ct.convert(...., useCPUOnly=use_cpu_for_conversion)"
It forces the model to be loaded on the CPU context, post conversion.
-----------
Returns MLModel, Input Values, Output Names
"""
if isinstance(output_nodes, tuple):
output_nodes = list(output_nodes)
if not isinstance(output_nodes, list):
output_nodes = [output_nodes]
# Convert TF graph.
input_names = get_tf_node_names(list(feed_dict.keys()), mode="inputs")
output_names = get_tf_node_names(output_nodes, mode="outputs")
input_values = {
name: val
for name, val in zip(input_names, feed_dict.values())
}
mlmodel = ct_convert(
graph,
inputs=None,
outputs=output_names,
source=frontend,
convert_to=backend,
useCPUOnly=use_cpu_for_conversion,
)
return mlmodel, input_values, output_names, output_nodes
def convert_to_mlmodel(model_spec,
tensor_inputs,
backend=("neuralnetwork", "fp32"),
converter_input_type=None,
use_cpu_for_conversion=False,
minimum_deployment_target=None):
def _convert_to_inputtype(inputs):
if isinstance(inputs, list):
return [_convert_to_inputtype(x) for x in inputs]
elif isinstance(inputs, tuple):
return tuple([_convert_to_inputtype(x) for x in inputs])
elif isinstance(inputs, TensorType):
return inputs
elif isinstance(inputs, torch.Tensor):
return TensorType(shape=inputs.shape,
dtype=torch_to_mil_types[inputs.dtype])
else:
raise ValueError("Unable to parse type {} into InputType.".format(
type(inputs)))
if converter_input_type is None:
inputs = list(_convert_to_inputtype(tensor_inputs))
else:
inputs = converter_input_type
if use_cpu_for_conversion:
compute_unit = ComputeUnit.CPU_ONLY
else:
compute_unit = ComputeUnit.ALL
return ct_convert(model_spec,
inputs=inputs,
convert_to=backend,
source="pytorch",
compute_units=compute_unit,
minimum_deployment_target=minimum_deployment_target)
def run_compare_tf2(
model,
input_dict,
output_names,
use_cpu_only=False,
use_cpu_for_conversion=False,
frontend_only=False,
frontend="tensorflow",
backend=("neuralnetwork", "fp32"),
debug=False,
atol=1e-04,
rtol=1e-05,
):
"""
Parameters
----------
model: list of tf.ConcreteFunction
List of TensorFlow 2.x concrete functions.
input_dict: dict of (str, np.array)
Dict of name and value pairs representing inputs.
output_names: list of str
List of output node names.
use_cpu_only: bool
If true, use CPU only for prediction, otherwise, use GPU also.
use_cpu_for_conversion: bool
If true, the converter is invoked using "ct.convert(...., useCPUOnly=True)",
which in turn forces the model to be loaded with the CPU context, which happens
when the converter loads the ML model object from the proto spec
using "ct.models.MLModel(proto_spec, useCPUOnly=True)".
The other argument, i.e., "use_cpu_only" on the other hand refers to only the compute engine
for prediction purposes. For a model that is loaded on a non-CPU context, it can still be forced
to execute on the CPU at the time of prediction. Hence,
"use_cpu_for_conversion = False && use_cpu_only = True" is valid and results in a case when a model is
loaded for GPU but executed on the CPU.
The scenario, "use_cpu_for_conversion = True && use_cpu_only = False" is invalid though,
since once a model is loaded on a CPU context its context cannot be changed to a non CPU device
at the time of prediction.
frontend_only: bool
If true, skip the prediction call, only validate conversion.
frontend: str
Frontend to convert from.
backend: str
Backend to convert to.
debug: bool
If true, print verbose information and plot intermediate graphs.
atol: float
The absolute tolerance parameter.
rtol: float
The relative tolerance parameter.
"""
if use_cpu_for_conversion and not use_cpu_only:
# use_cpu_for_conversion = True && use_cpu_only = False
raise ValueError(
"use_cpu_for_conversion = True && use_cpu_only = False is an invalid test case"
)
inputs = []
cf_inputs = [t for t in model[0].inputs if t.dtype != dtypes.resource]
for t in cf_inputs:
name = get_tf_node_names(t.name)[0]
shape = [RangeDim() if s is None or s == -1 else s \
for s in list(t.get_shape())]
inputs.append(
TensorType(name=name, shape=shape, dtype=t.dtype.as_numpy_dtype))
outputs = []
for t in output_names:
name = get_tf_node_names(t)[0]
outputs.append(name)
# get TensorFlow 2.x output as reference and run comparison
tf_input_values = [tf.constant(t) for t in input_dict.values()]
tf_outputs = model[0](*tf_input_values)
if isinstance(tf_outputs, (tuple, list)):
ref = [t.numpy() for t in tf_outputs]
else:
ref = [tf_outputs.numpy()]
expected_outputs = {n: v for n, v in zip(outputs, ref)}
mlmodel = ct_convert(
model,
source=frontend,
inputs=inputs,
outputs=outputs,
convert_to=backend,
debug=debug,
useCPUOnly=use_cpu_for_conversion,
)
for k, v in input_dict.items():
if isinstance(v, np.ndarray) and issubclass(v.dtype.type, np.integer):
input_dict[k] = v.astype(np.float) # Core ML only accepts floats
if frontend_only or _macos_version() < (10, 13) \
or (mlmodel.is_package and _macos_version() < (12, 0)):
return mlmodel._spec, mlmodel, input_dict, None
compare_backend(
mlmodel,
input_dict,
expected_outputs,
use_cpu_only,
atol=atol,
rtol=rtol,
also_compare_shapes=True,
dtype=backend[1],
)
pred = None
if not coremltoolsutils._has_custom_layer(mlmodel.get_spec()):
pred = run_core_ml_predict(mlmodel, input_dict, use_cpu_only)
else:
print('Skipping model prediction as it has a custom nn layer!')
return mlmodel._spec, mlmodel, input_dict, pred
def run_compare_tf_keras(
model,
input_values,
use_cpu_only=False,
frontend_only=False,
frontend="tensorflow",
backend=("neuralnetwork", "fp32"),
atol=1e-04,
rtol=1e-05,
):
"""
Parameters
----------
model: TensorFlow 2.x model
TensorFlow 2.x model annotated with @tf.function.
input_values: list of np.array
List of input values in the same order as the input signature.
use_cpu_only: bool
If true, use CPU only for prediction, otherwise, use GPU also.
frontend_only: bool
If true, skip the prediction call, only validate conversion.
frontend: str
Frontend to convert from.
backend: str
Backend to convert to.
atol: float
The absolute tolerance parameter.
rtol: float
The relative tolerance parameter.
"""
mlmodel = ct_convert(model, source=frontend, convert_to=backend)
# assumes conversion preserve the i/o names
proto = mlmodel.get_spec()
inputs = [i.name.split(":")[0].strip() for i in model.inputs]
outputs = [str(o.name) for o in proto.description.output]
# get tf.keras model output as reference and run comparison
keras_outputs = model(input_values)
if not isinstance(keras_outputs, list):
keras_outputs = [keras_outputs]
ref = [output.numpy() for output in keras_outputs]
expected_outputs = {n: v for n, v in zip(outputs, ref)}
input_key_values = {n: v for n, v in zip(inputs, input_values)}
if frontend_only or _macos_version() < (10, 13) \
or (mlmodel.is_package and _macos_version() < (12, 0)):
return proto, mlmodel, input_key_values, None
compare_backend(mlmodel,
input_key_values,
expected_outputs,
use_cpu_only,
atol=atol,
rtol=rtol,
also_compare_shapes=True,
dtype=backend[1])
pred = None
if not coremltoolsutils._has_custom_layer(proto):
pred = run_core_ml_predict(mlmodel, input_key_values, use_cpu_only)
else:
print('Skipping model prediction as it has a custom nn layer!')
return proto, mlmodel, input_key_values, pred
def run_compare_builder(
build,
input_placeholders,
input_values,
expected_output_types=None,
expected_outputs=None,
use_cpu_only=False,
frontend_only=False,
backend=("neuralnetwork", "fp32"),
atol=1e-04,
rtol=1e-05,
inputs=None,
also_compare_shapes=False,
use_cpu_for_conversion=False,
):
"""
Inputs:
- build: python function taking input of Vars and returning Var or
list[Var]. Each input argument in build must match a key in
input_values / input_placeholders.
- input_placeholders: str -> placeholder. It may not be an empty
dict as MLModel doesn't support function with
no input.
- input_values: str -> np.array or PIL.Image. Keys must match those in
input_placeholders.
- expected_output_types: list[(shape, builtin_type)] or (shape,
builtin_type). None skips type inference validation.
- expected_outputs: list[np.array] or np.array. Required iff
frontend_only == False
- frontend_only: True to test up to proto generation.
- inputs: type of inputs (either None (defaults to tensor) or [ct.ImageType])
- use_cpu_for_conversion: bool
Argument which is passed as is to the unified converter API.
That is, "ct.convert(...., useCPUOnly=use_cpu_for_conversion)"
It forces the model to be loaded on the CPU context, post conversion.
Returns:
The converted mlmodel
"""
if not isinstance(expected_output_types, list):
expected_output_types = [expected_output_types]
if expected_outputs is not None and not isinstance(expected_outputs, list):
expected_outputs = [expected_outputs]
prog = Program()
with Function(input_placeholders) as ssa_func:
output_vars = build(**ssa_func.inputs)
if isinstance(output_vars, tuple):
output_vars = list(output_vars)
elif not isinstance(output_vars, list):
output_vars = [output_vars]
ssa_func.set_outputs(output_vars)
prog.add_function("main", ssa_func)
# get output names for output_vars
output_names = [x.name for x in output_vars]
# Validate type inference
msg = ("Provided expected outputs types {} should match number of output" +
" variables {}")
assert_msg = msg.format(len(expected_output_types), len(output_vars))
assert len(output_vars) == len(expected_output_types), assert_msg
for out_var, s in zip(output_vars, expected_output_types):
if out_var.dtype != s[-1]:
raise ValueError(
"Output {} type: expect {}, got {}. Program:\n{}".format(
out_var.name, s[-1].__type_info__(),
out_var.dtype.__type_info__(), prog))
if UNK_VARIADIC in s[:-1]:
msg = "Skip type checking for UNK_VARIADIC. Output shape: {} vs expected shape: {}"
logging.debug(msg.format(out_var.shape, s[:-1]))
continue
expected_shape = s[:-1]
msg = "Output {} shape: expect {}, got {}. Program:\n{}".format(
out_var.name, expected_shape, out_var.shape, prog)
# No more variadic here.
if len(out_var.shape) != len(expected_shape):
raise ValueError(msg)
# replace UNK_SYM in out_var.shape.
output_shape = [
0 if es == UNK_SYM else os
for os, es in zip(out_var.shape, expected_shape)
]
expected_shape = [0 if es == UNK_SYM else es for es in expected_shape]
# convert float etc to int.
output_shape = [i if is_symbolic(i) else int(i) for i in output_shape]
expected_shape = [
i if is_symbolic(i) else int(i) for i in expected_shape
]
if output_shape != expected_shape:
raise ValueError(msg)
mlmodel = ct_convert(prog,
source="milinternal",
convert_to=backend,
inputs=inputs,
useCPUOnly=use_cpu_for_conversion)
if frontend_only:
return mlmodel
if expected_outputs:
assert len(output_vars) == len(expected_outputs), (
"Provided expected_outputs {}"
" should match number of output"
" variables {}".format(len(expected_outputs), len(output_vars)))
expected_outputs = {
name: val
for name, val in zip(output_names, expected_outputs)
}
compare_backend(mlmodel=mlmodel,
input_key_values=input_values,
expected_outputs=expected_outputs,
use_cpu_only=use_cpu_only,
atol=atol,
rtol=rtol,
also_compare_shapes=also_compare_shapes,
dtype=backend[1])
return mlmodel
def run_compare_tf2(
model,
input_dict,
output_names,
inputs_for_conversion=None,
use_cpu_for_conversion=False,
frontend_only=False,
frontend="tensorflow",
backend=("neuralnetwork", "fp32"),
debug=False,
atol=1e-04,
rtol=1e-05,
minimum_deployment_target=None,
):
"""
Parameters
----------
model: list of tf.ConcreteFunction
List of TensorFlow 2.x concrete functions.
input_dict: dict of (str, np.array)
Dict of name and value pairs representing inputs.
output_names: list of str
List of output node names.
inputs_for_conversion: list of coremltools.TensorType() or coremltools.ImageType() objects
Defaults to None. It is passed as is to the "inputs" argument of the converter.
use_cpu_for_conversion: bool
If True, forces the model to be loaded with the CPU context.
frontend_only: bool
If True, skip the prediction call, only validate conversion.
frontend: str
Frontend to convert from.
backend: str
Backend to convert to.
debug: bool
If True, print verbose information and plot intermediate graphs.
atol: float
The absolute tolerance parameter.
rtol: float
The relative tolerance parameter.
minimum_deployment_target: coremltools.target enumeration
The spec version for the mlmodel
"""
inputs = []
if inputs_for_conversion is None:
cf_inputs = [t for t in model[0].inputs if t.dtype != dtypes.resource]
for t in cf_inputs:
name = get_tf_node_names(t.name)[0]
shape = [RangeDim() if s is None or s == -1 else s \
for s in list(t.get_shape())]
inputs.append(
TensorType(name=name,
shape=shape,
dtype=t.dtype.as_numpy_dtype))
else:
inputs = inputs_for_conversion
outputs = []
for t in output_names:
name = get_tf_node_names(t)[0]
outputs.append(name)
# get TensorFlow 2.x output as reference and run comparison
tf_input_values = [tf.constant(t) for t in input_dict.values()]
tf_outputs = model[0](*tf_input_values)
if isinstance(tf_outputs, (tuple, list)):
ref = [t.numpy() for t in tf_outputs]
else:
ref = [tf_outputs.numpy()]
expected_outputs = {n: v for n, v in zip(outputs, ref)}
if use_cpu_for_conversion:
compute_unit = ct.ComputeUnit.CPU_ONLY
else:
compute_unit = ct.ComputeUnit.ALL
mlmodel = ct_convert(
model,
source=frontend,
inputs=inputs,
outputs=outputs,
convert_to=backend,
debug=debug,
compute_units=compute_unit,
minimum_deployment_target=minimum_deployment_target,
)
for k, v in input_dict.items():
if isinstance(v, np.ndarray) and issubclass(v.dtype.type, np.integer):
input_dict[k] = v.astype(np.float) # Core ML only accepts floats
if frontend_only or _macos_version() < (10, 13) \
or (mlmodel.is_package and _macos_version() < (12, 0)):
return mlmodel._spec, mlmodel, input_dict, None
pred = None
if not coremltoolsutils._has_custom_layer(mlmodel._spec):
pred = compare_backend(
mlmodel,
input_dict,
expected_outputs,
atol=atol,
rtol=rtol,
also_compare_shapes=True,
dtype=backend[1],
)
else:
print('Skipping model prediction as it has a custom nn layer!')
return mlmodel._spec, mlmodel, input_dict, pred
def tf_graph_to_mlmodel(
graph,
feed_dict,
output_nodes,
frontend="tensorflow",
backend=("neuralnetwork", "fp32"),
use_cpu_for_conversion=False,
inputs_for_conversion=None,
minimum_deployment_target=None,
):
"""
Parameters
----------
graph: tf.Graph
TensorFlow 1.x model in tf.Graph format.
feed_dict: dict of {tf.placeholder -> np.array or python primitive)
Dict of placeholder and value pairs representing inputs.
output_nodes: tf.node or list[tf.node]
List of names representing outputs.
frontend: str
Frontend to convert from.
backend: str
Backend to convert to.
use_cpu_for_conversion: bool
Argument which is passed as is to the unified converter API.
It forces the model to be loaded on the CPU context, post conversion.
inputs_for_conversion: list of coremltools.TensorType() or coremltools.ImageType() objects
Defaults to None. It is passed as is to the "inputs" argument of the converter.
minimum_deployment_target : coremltools.target enumeration
It set the minimum_deployment_target argument in the coremltools.convert functino.
-----------
Returns MLModel, Input Values, Output Names
"""
if isinstance(output_nodes, tuple):
output_nodes = list(output_nodes)
if not isinstance(output_nodes, list):
output_nodes = [output_nodes]
# Convert TF graph.
input_names = get_tf_node_names(list(feed_dict.keys()), mode="inputs")
output_names = get_tf_node_names(output_nodes, mode="outputs")
input_values = {
name: val
for name, val in zip(input_names, feed_dict.values())
}
if use_cpu_for_conversion:
compute_unit = ct.ComputeUnit.CPU_ONLY
else:
compute_unit = ct.ComputeUnit.ALL
inputs = inputs_for_conversion if inputs_for_conversion is not None else None
mlmodel = ct_convert(
graph,
inputs=inputs,
outputs=output_names,
source=frontend,
convert_to=backend,
compute_units=compute_unit,
minimum_deployment_target=minimum_deployment_target,
)
return mlmodel, input_values, output_names, output_nodes