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_tf(
graph,
feed_dict,
output_nodes,
use_cpu_only=False,
frontend_only=False,
frontend="tensorflow",
backend="nn_proto",
atol=1e-04,
rtol=1e-05,
validate_shapes_only=False,
freeze_graph=False,
tf_outputs=None,
):
"""
Utility function to convert and compare a given TensorFlow 1.x model.
Parameters
----------
graph: tf.Graph
TensorFlow 1.x model in tf.Graph format.
feed_dict: dict of (tf.placeholder, np.array)
Dict of placeholder and value pairs representing inputs.
output_nodes: tf.node or list[tf.node]
List of names representing outputs.
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.
validate_shapes_only: bool
If true, skip element-wise value comparision.
tf_outputs: float or list[float]
If present, use it as TensorFlow predictions
Return:
Proto, mlmodel, input dictionay, prediction(if possible)
"""
mlmodel, input_key_values, output_names, output_nodes = tf_graph_to_mlmodel(
graph, feed_dict, output_nodes, frontend, backend)
if frontend_only:
return
if not isinstance(output_nodes, (tuple, list)):
output_nodes = [output_nodes]
if freeze_graph:
model_dir = tempfile.mkdtemp()
graph_def_file = os.path.join(model_dir, "tf_graph.pb")
checkpoint_file = os.path.join(model_dir, "tf_model.ckpt")
static_model_file = os.path.join(model_dir, "tf_static.pb")
coreml_model_file = os.path.join(model_dir, "coreml_model.mlmodel")
with tf.Session(graph=graph) as sess:
sess.run(tf.global_variables_initializer())
tf_outputs = sess.run(output_nodes, feed_dict=feed_dict)
tf.train.write_graph(sess.graph,
model_dir,
graph_def_file,
as_text=False)
saver = tf.train.Saver()
saver.save(sess, checkpoint_file)
freeze_g(
input_graph=graph_def_file,
input_saver="",
input_binary=True,
input_checkpoint=checkpoint_file,
output_node_names=",".join([n.op.name for n in output_nodes]),
restore_op_name="save/restore_all",
filename_tensor_name="save/Const:0",
output_graph=static_model_file,
clear_devices=True,
initializer_nodes="",
)
graph = load_tf_pb(static_model_file)
# Need to convert again using frozen graph
mlmodel, input_key_values, output_names, output_nodes = tf_graph_to_mlmodel(
graph, feed_dict, output_nodes, frontend, backend)
else:
if not tf_outputs:
with tf.Session(graph=graph) as sess:
sess.run(tf.global_variables_initializer())
tf_outputs = sess.run(output_nodes, feed_dict=feed_dict)
expected_outputs = {
name: val
for name, val in zip(output_names, tf_outputs)
}
for k, v in input_key_values.items():
if isinstance(v, np.ndarray) and issubclass(v.dtype.type, np.integer):
input_key_values[k] = v.astype(
np.float) # Core ML only accepts floats
if validate_shapes_only:
compare_shapes(mlmodel, input_key_values, expected_outputs,
use_cpu_only)
else:
compare_backend(
mlmodel,
input_key_values,
expected_outputs,
use_cpu_only,
atol=atol,
rtol=rtol,
also_compare_shapes=True,
)
pred = None
if not coremltoolsutils._has_custom_layer(mlmodel.get_spec()):
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 mlmodel._spec, mlmodel, input_key_values, pred
def run_compare_tf(
graph,
feed_dict,
output_nodes,
use_cpu_only=False,
use_cpu_for_conversion=False,
frontend_only=False,
frontend="tensorflow",
backend=("neuralnetwork", "fp32"),
atol=1e-04,
rtol=1e-05,
validate_shapes_only=False,
freeze_graph=False,
tf_outputs=None,
):
"""
Utility function to convert and compare a given TensorFlow 1.x model.
Parameters
----------
graph: tf.Graph
TensorFlow 1.x model in tf.Graph format.
feed_dict: dict of (tf.placeholder, np.array)
Dict of placeholder and value pairs representing inputs.
output_nodes: tf.node or list[tf.node]
List of names representing outputs.
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.
atol: float
The absolute tolerance parameter.
rtol: float
The relative tolerance parameter.
validate_shapes_only: bool
If true, skip element-wise value comparision.
freeze_graph: bool
If True, use the "tensorflow.python.tools.freeze_graph" function
to freeze the TF graph prior to conversion. This will ensure that
all the variables in the graph have been converted to constants.
tf_outputs: float or list[float]
If present, use it as TensorFlow predictions
Return:
Proto, mlmodel, input dictionay, prediction(if possible)
"""
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"
)
if not isinstance(output_nodes, (tuple, list)):
output_nodes = [output_nodes]
if freeze_graph:
with tempfile.TemporaryDirectory() as model_dir:
graph_def_file = os.path.join(model_dir, "tf_graph.pb")
checkpoint_file = os.path.join(model_dir, "tf_model.ckpt")
static_model_file = os.path.join(model_dir, "tf_static.pb")
with tf.Session(graph=graph) as sess:
sess.run(tf.global_variables_initializer())
if tf_outputs is None:
tf_outputs = sess.run(output_nodes, feed_dict=feed_dict)
tf.train.write_graph(sess.graph,
model_dir,
graph_def_file,
as_text=False)
saver = tf.train.Saver()
saver.save(sess, checkpoint_file)
output_node_names = get_tf_node_names(output_nodes,
mode="outputs")
output_node_names = [
name.split(":")[0] for name in output_node_names
]
output_op_names = ",".join(output_node_names)
freeze_g(
input_graph=graph_def_file,
input_saver="",
input_binary=True,
input_checkpoint=checkpoint_file,
output_node_names=output_op_names,
restore_op_name="save/restore_all",
filename_tensor_name="save/Const:0",
output_graph=static_model_file,
clear_devices=True,
initializer_nodes="",
)
graph = load_tf_pb(static_model_file)
mlmodel, input_key_values, output_names, output_nodes = tf_graph_to_mlmodel(
graph,
feed_dict,
output_nodes,
frontend,
backend,
use_cpu_for_conversion=use_cpu_for_conversion,
)
if frontend_only or coremltoolsutils._macos_version() < (10, 13) \
or (mlmodel.is_package and coremltoolsutils._macos_version() < (12, 0)):
return mlmodel._spec, mlmodel, input_key_values, None
if tf_outputs is None:
with tf.Session(graph=graph) as sess:
sess.run(tf.global_variables_initializer())
tf_outputs = sess.run(output_nodes, feed_dict=feed_dict)
expected_outputs = {
name: val
for name, val in zip(output_names, tf_outputs)
}
for k, v in input_key_values.items():
if isinstance(v, np.ndarray) and issubclass(v.dtype.type, np.integer):
input_key_values[k] = v.astype(
np.float) # Core ML only accepts floats
if validate_shapes_only:
compare_shapes(mlmodel, input_key_values, expected_outputs,
use_cpu_only)
else:
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(mlmodel.get_spec()):
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 mlmodel._spec, mlmodel, input_key_values, pred
def run_compare_tf2(
model,
input_dict,
output_names,
use_cpu_only=False,
frontend_only=False,
frontend="tensorflow",
backend="nn_proto",
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.
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.
"""
# Avoid circular dependency
import coremltools as ct
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,
)
if frontend_only:
return
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
compare_backend(
mlmodel,
input_dict,
expected_outputs,
use_cpu_only,
atol=atol,
rtol=rtol,
also_compare_shapes=True,
)
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="nn_proto",
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.
"""
# Avoid circular dependency
from coremltools.converters._converters_entry import convert
mlmodel = convert(model, source=frontend, convert_to=backend)
# assumes conversion preserve the i/o names
proto = mlmodel.get_spec()
inputs = sorted([str(i.name) for i in proto.description.input])
outputs = [str(o.name) for o in proto.description.output]
if frontend_only:
return
# 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)}
compare_backend(
mlmodel,
input_key_values,
expected_outputs,
use_cpu_only,
atol=atol,
rtol=rtol,
also_compare_shapes=True,
)
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