def toco_convert_protos(model_flags_str,
toco_flags_str,
input_data_str,
debug_info_str=None,
enable_mlir_converter=False):
"""Convert `input_data_str` according to model and toco parameters.
Unless you know what you are doing consider using
the more friendly `tf.compat.v1.lite.toco_convert`.
Args:
model_flags_str: Serialized proto describing model properties, see
`toco/model_flags.proto`.
toco_flags_str: Serialized proto describing conversion properties, see
`toco/toco_flags.proto`.
input_data_str: Input data in serialized form (e.g. a graphdef is common)
debug_info_str: Serialized `GraphDebugInfo` proto describing logging
information. (default None)
enable_mlir_converter: Enables MLIR-based conversion instead of the default
TOCO conversion. (default False)
Returns:
Converted model in serialized form (e.g. a TFLITE model is common).
Raises:
ConverterError: When conversion fails in TFLiteConverter, usually due to
ops not being supported.
RuntimeError: When conversion fails, an exception is raised with the error
message embedded.
"""
# Historically, TOCO conversion failures would trigger a crash, so we would
# attempt to run the converter out-of-process. The MLIR conversion pipeline
# surfaces errors instead, and can be safely run in-process.
if enable_mlir_converter or not _toco_from_proto_bin:
try:
model_str = wrap_toco.wrapped_toco_convert(model_flags_str,
toco_flags_str,
input_data_str,
debug_info_str,
enable_mlir_converter)
return model_str
except Exception as e:
converter_error = ConverterError(str(e))
for error_data in _metrics_wrapper.get_collected_errors():
converter_error.append_error(error_data)
raise converter_error
return _run_toco_binary(model_flags_str, toco_flags_str, input_data_str,
debug_info_str)
def convert(model_flags_str,
conversion_flags_str,
input_data_str,
debug_info_str=None,
enable_mlir_converter=True):
"""Converts `input_data_str` to a TFLite model.
Args:
model_flags_str: Serialized proto describing model properties, see
`model_flags.proto`.
conversion_flags_str: Serialized proto describing conversion properties, see
`toco/toco_flags.proto`.
input_data_str: Input data in serialized form (e.g. a graphdef is common, or
it can be hlo text or proto)
debug_info_str: Serialized `GraphDebugInfo` proto describing logging
information. (default None)
enable_mlir_converter: Enables MLIR-based conversion. (default True)
Returns:
Converted model in serialized form (e.g. a TFLITE model is common).
Raises:
ConverterError: When conversion fails in TFLiteConverter, usually due to
ops not being supported.
RuntimeError: When conversion fails, an exception is raised with the error
message embedded.
"""
# Historically, deprecated conversion failures would trigger a crash, so we
# attempt to run the converter out-of-process. The current MLIR conversion
# pipeline surfaces errors instead, and can be safely run in-process.
if enable_mlir_converter or not _deprecated_conversion_binary:
try:
model_str = wrap_toco.wrapped_toco_convert(model_flags_str,
conversion_flags_str,
input_data_str,
debug_info_str,
enable_mlir_converter)
return model_str
except Exception as e:
converter_error = ConverterError(str(e))
for error_data in _metrics_wrapper.retrieve_collected_errors():
converter_error.append_error(error_data)
raise converter_error
return _run_deprecated_conversion_binary(model_flags_str,
conversion_flags_str,
input_data_str, debug_info_str)
def _run_toco_binary(model_flags_str,
toco_flags_str,
input_data_str,
debug_info_str=None):
"""Convert `input_data_str` using TOCO converter binary.
Args:
model_flags_str: Serialized proto describing model properties, see
`toco/model_flags.proto`.
toco_flags_str: Serialized proto describing conversion properties, see
`toco/toco_flags.proto`.
input_data_str: Input data in serialized form (e.g. a graphdef is common)
debug_info_str: Serialized `GraphDebugInfo` proto describing logging
information. (default None)
Returns:
Converted model in serialized form (e.g. a TFLITE model is common).
Raises:
ConverterError: When cannot find the toco binary.
RuntimeError: When conversion fails, an exception is raised with the error
message embedded.
"""
if distutils.spawn.find_executable(_toco_from_proto_bin) is None:
raise ConverterError(
"""Could not find toco_from_protos binary, make sure
your virtualenv bin directory or pip local bin directory is in your path.
In particular, if you have installed TensorFlow with --user, make sure you
add the install directory to your path.
For example:
Linux: export PATH=$PATH:~/.local/bin/
Mac: export PATH=$PATH:~/Library/Python/<version#>/bin
Alternative, use virtualenv.""")
# Windows and TemporaryFile are not that useful together,
# since you cannot have two readers/writers. So we have to
# make the temporaries and close and delete them explicitly.
toco_filename, model_filename, input_filename, output_filename = (None,
None,
None,
None)
try:
# Build all input files
with _tempfile.NamedTemporaryFile(delete=False) as fp_toco, \
_tempfile.NamedTemporaryFile(delete=False) as fp_model, \
_tempfile.NamedTemporaryFile(delete=False) as fp_input, \
_tempfile.NamedTemporaryFile(delete=False) as fp_debug:
toco_filename = fp_toco.name
input_filename = fp_input.name
model_filename = fp_model.name
debug_filename = fp_debug.name
fp_model.write(model_flags_str)
fp_toco.write(toco_flags_str)
fp_input.write(six.ensure_binary(input_data_str))
debug_info_str = debug_info_str if debug_info_str else ""
# if debug_info_str contains a "string value", then the call to
# fp_debug.write(debug_info_str) will fail with the following error
#
# TypeError: a bytes-like object is required, not 'str'
#
# Some of the subtests within the "convert_test" unit-test fail
# with the error shown above. So watch out for that scenario and
# convert debug_info_str to bytes where needed
if not isinstance(debug_info_str, bytes):
fp_debug.write(debug_info_str.encode("utf-8"))
else:
fp_debug.write(debug_info_str)
# Reserve an output file
with _tempfile.NamedTemporaryFile(delete=False) as fp:
output_filename = fp.name
# Run
cmd = [
_toco_from_proto_bin,
model_filename,
toco_filename,
input_filename,
output_filename,
"--debug_proto_file={}".format(debug_filename),
]
cmdline = " ".join(cmd)
is_windows = _platform.system() == "Windows"
proc = _subprocess.Popen(cmdline,
shell=True,
stdout=_subprocess.PIPE,
stderr=_subprocess.STDOUT,
close_fds=not is_windows)
stdout, stderr = proc.communicate()
exitcode = proc.returncode
if exitcode == 0:
with open(output_filename, "rb") as fp:
return fp.read()
else:
stdout = _try_convert_to_unicode(stdout)
stderr = _try_convert_to_unicode(stderr)
raise ConverterError("See console for info.\n%s\n%s\n" %
(stdout, stderr))
finally:
# Must manually cleanup files.
for filename in [
toco_filename, input_filename, model_filename, output_filename
]:
try:
_os.unlink(filename)
except (OSError, TypeError):
pass