# Copyright 2022 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Converts a frozen graph into a TFLite FlatBuffer.""" import distutils.spawn import enum import hashlib import os as _os import platform as _platform import subprocess as _subprocess import tempfile as _tempfile from typing import Optional import warnings from tensorflow.compiler.mlir.quantization.stablehlo import quantization_config_pb2 from tensorflow.compiler.mlir.quantization.stablehlo import quantization_options_pb2 as quant_opts_pb2 from tensorflow.lite.python import lite_constants from tensorflow.lite.python import util from tensorflow.lite.python import wrap_toco from tensorflow.lite.python.convert_phase import Component from tensorflow.lite.python.convert_phase import convert_phase from tensorflow.lite.python.convert_phase import ConverterError from tensorflow.lite.python.convert_phase import SubComponent from tensorflow.lite.python.metrics import converter_error_data_pb2 from tensorflow.lite.python.metrics.wrapper import metrics_wrapper as _metrics_wrapper from tensorflow.lite.toco import model_flags_pb2 as _model_flags_pb2 from tensorflow.lite.toco import toco_flags_pb2 as _conversion_flags_pb2 from tensorflow.lite.toco import types_pb2 as _types_pb2 from tensorflow.lite.tools import flatbuffer_utils from tensorflow.python.framework import dtypes from tensorflow.python.framework import tensor_shape from tensorflow.python.platform import resource_loader as _resource_loader from tensorflow.python.util import deprecation from tensorflow.python.util.tf_export import tf_export as _tf_export def _is_quantized_input_stats_required( conversion_flags: _conversion_flags_pb2.TocoFlags, ) -> bool: """Checks if the `quantized_input_stats` flag is required for conversion. Args: conversion_flags: A protocol buffer describing the conversion process. Returns: True, if the `inference_type` or the `inference_input_type` is a quantized type and it is not post training quantization, else False. """ quantized_inference_types = [ _types_pb2.QUANTIZED_UINT8, _types_pb2.QUANTIZED_INT8, ] return ( conversion_flags.inference_type in quantized_inference_types or conversion_flags.inference_input_type in quantized_inference_types ) and not conversion_flags.post_training_quantize def convert_tensor_tf_type_to_tflite_type( tf_type: dtypes.DType, usage: str = "" ) -> _types_pb2.IODataType: """Convert tensor type from tf type to tflite type. Args: tf_type: TensorFlow type. usage: Text describing the reason for invoking this function. Raises: ValueError: If `tf_type` is unsupported. Returns: tflite_type: TFLite type. Refer to lite/toco/types.proto. """ mapping = { dtypes.float16: _types_pb2.FLOAT16, dtypes.float32: _types_pb2.FLOAT, dtypes.float64: _types_pb2.FLOAT64, dtypes.int8: _types_pb2.INT8, dtypes.int16: _types_pb2.INT16, dtypes.uint16: _types_pb2.UINT16, dtypes.int32: _types_pb2.INT32, dtypes.int64: _types_pb2.INT64, dtypes.uint8: _types_pb2.UINT8, dtypes.uint32: _types_pb2.UINT32, dtypes.uint64: _types_pb2.UINT64, dtypes.string: _types_pb2.STRING, dtypes.bool: _types_pb2.BOOL, dtypes.complex64: _types_pb2.COMPLEX64, dtypes.complex128: _types_pb2.COMPLEX128, } tflite_type = mapping.get(tf_type) if tflite_type is None: raise ValueError( "Unsupported TensorFlow type `{0}` provided for the {1}".format( tf_type, usage ) ) return tflite_type # Only a few restricted tensor types are allowed for explicitly setting # inference/input/output types. def convert_inference_tf_type_to_tflite_type( tf_type: dtypes.DType, usage: str = "" ) -> _types_pb2.IODataType: """Convert inference type from tf type to tflite type. Args: tf_type: TensorFlow type. usage: Text describing the reason for invoking this function. Raises: ValueError: If `tf_type` is unsupported. Returns: tflite_type: TFLite type. Refer to lite/toco/types.proto. """ mapping = { dtypes.float32: _types_pb2.FLOAT, dtypes.uint8: _types_pb2.QUANTIZED_UINT8, dtypes.int8: _types_pb2.QUANTIZED_INT8, dtypes.int16: _types_pb2.QUANTIZED_INT16, } tflite_type = mapping.get(tf_type) if tflite_type is None: raise ValueError( "Unsupported TensorFlow type `{0}` provided for the {1}".format( tf_type, usage ) ) return tflite_type # Find the deprecated conversion binary using the resource loader if using from # bazel, otherwise we are in a pip where console_scripts already has the tool. if lite_constants.EXPERIMENTAL_USE_TOCO_API_DIRECTLY: _deprecated_conversion_binary = "" else: _deprecated_conversion_binary = _resource_loader.get_path_to_datafile( "../toco/python/toco_from_protos" ) if not _os.path.exists(_deprecated_conversion_binary): _deprecated_conversion_binary = "toco_from_protos" def _try_convert_to_unicode(output): if output is None: return "" if isinstance(output, bytes): try: return output.decode("utf-8") except UnicodeDecodeError: pass return output @_tf_export("lite.OpsSet") class OpsSet(enum.Enum): """Enum class defining the sets of ops available to generate TFLite models. WARNING: Experimental interface, subject to change. """ # Convert model using TensorFlow Lite builtin ops. TFLITE_BUILTINS = "TFLITE_BUILTINS" # Convert model using TensorFlow ops. Not all TensorFlow ops are available. # WARNING: Experimental interface, subject to change. SELECT_TF_OPS = "SELECT_TF_OPS" # Convert model using only TensorFlow Lite quantized int8 operations. # Specifying this will throw an error for operations that do not yet have # quantized implementations. TFLITE_BUILTINS_INT8 = "TFLITE_BUILTINS_INT8" # Convert model using only TensorFlow Lite operations with quantized int8 # weights, int16 activations and int64 bias. # Specifying this will throw an error for operations that do not yet have # quantized implementations. # This quantization mode may be used in models for super-resolution, # audio signal processing or image de-noising. It improves accuracy # significantly, but only slightly increases the model size. # WARNING: These ops are currently experimental and have not yet been # finalized. # They are only compatible with CPU execution, and have not been optimized for # production. EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = ( "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" ) # Convert model using only stablehlo ops. # This option can not be combined with other OpsSets. # The feature is in early development. # The code to execute StableHLO ops in the runtime is to be implemented # and the serialization format is not stabilized yet. EXPERIMENTAL_STABLEHLO_OPS = "EXPERIMENTAL_STABLEHLO_OPS" def __str__(self): return str(self.value) @staticmethod def get_options(): """Returns a list of OpsSet options as a list of strings.""" return [str(option) for option in list(OpsSet)] @convert_phase(Component.OPTIMIZE_TFLITE_MODEL, SubComponent.QUANTIZE) def mlir_quantize( input_data_str, disable_per_channel=False, fully_quantize=False, inference_type=_types_pb2.QUANTIZED_INT8, input_data_type=dtypes.float32, output_data_type=dtypes.float32, enable_numeric_verify=False, enable_whole_model_verify=False, denylisted_ops=None, denylisted_nodes=None, enable_variable_quantization=False, ): """Quantize `input_data_str` with calibration results. Args: input_data_str: Input data in serialized form (e.g. a TFLITE model with calibration results). disable_per_channel: Bool indicating whether to do per-channel or per-tensor quantization fully_quantize: Bool indicating whether to fully quantize the model. Besides model body, the input/output will be quantized as well. inference_type: Data type for the activations. The default value is int8. input_data_type: Data type for the inputs. The default value is float32. output_data_type: Data type for the outputs. The default value is float32. enable_numeric_verify: Experimental. Subject to change. Bool indicating whether to add NumericVerify ops into the debug mode quantized model. enable_whole_model_verify: Experimental. Subject to change. Bool indicating whether to add verification for layer by layer, or on whole model. When disabled (per-layer) float and quantized ops will be run from same input (output of previous quantized layer). When enabled, float and quantized ops will run with respective float and quantized output of previous ops. denylisted_ops: Experimental. Subject to change. Set of ops to denylist. denylisted_nodes: Experimental. Subject to change. Set of notes to denylist. enable_variable_quantization: Experimental. Subject to change. Bool indicating whether to enable quantization of the residual variables remaining after the variable freezing pass. Returns: Quantized model in serialized form (e.g. a TFLITE model) with floating-point inputs and outputs. """ return wrap_toco.wrapped_experimental_mlir_quantize( input_data_str, disable_per_channel, fully_quantize, inference_type, convert_tensor_tf_type_to_tflite_type(input_data_type), convert_tensor_tf_type_to_tflite_type(output_data_type), enable_numeric_verify, enable_whole_model_verify, denylisted_ops, denylisted_nodes, enable_variable_quantization, ) @convert_phase(Component.OPTIMIZE_TFLITE_MODEL, SubComponent.SPARSIFY) def mlir_sparsify(input_data_str): """Sparsify `input_data_str` to encode sparse tensor with proper format. Args: input_data_str: Input data in serialized form (e.g. a TFLITE model). Returns: Sparsified model in serialized form (e.g. a TFLITE model). """ return wrap_toco.wrapped_experimental_mlir_sparsify(input_data_str) def register_custom_opdefs(custom_opdefs_list): """Register the given custom opdefs to the TensorFlow global op registry. Args: custom_opdefs_list: String representing the custom ops OpDefs that are included in the GraphDef. Returns: True if the registration is successfully completed. """ return wrap_toco.wrapped_register_custom_opdefs(custom_opdefs_list) def convert( model_flags: _model_flags_pb2.ModelFlags, conversion_flags: _conversion_flags_pb2.TocoFlags, input_data_str: Optional[str] = None, debug_info_str: Optional[str] = None, enable_mlir_converter: bool = True, ): """Converts `input_data_str` to a TFLite model. Args: model_flags: Proto describing model properties, see `model_flags.proto`. conversion_flags: 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. enable_mlir_converter: Enables MLIR-based conversion. 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: return wrap_toco.wrapped_toco_convert( model_flags.SerializeToString(), conversion_flags.SerializeToString(), input_data_str, debug_info_str, enable_mlir_converter, ) except Exception as e: converter_error = ConverterError(str(e)) for error_data in _metrics_wrapper.retrieve_collected_errors(): converter_error.append_error(error_data) # Seldom we encounter the case where an unsupported # `StatefulPartitionedCallOp` is not inlined and remains in the final # IR. If this occurs we can set `guarantee_all_funcs_one_use` and retry. # This makes the converter copy functions definitions called by # multiple StatefulPartitionedCall, thus allowing them to be properly # inlined. if ( error_data.error_code == converter_error_data_pb2.ConverterErrorData.ERROR_STATEFUL_PARTITIONED_CALL_IN_FINAL_IR and not conversion_flags.guarantee_all_funcs_one_use ): conversion_flags.guarantee_all_funcs_one_use = True return convert( model_flags, conversion_flags, input_data_str, debug_info_str, enable_mlir_converter, ) raise converter_error return _run_deprecated_conversion_binary( model_flags.SerializeToString(), conversion_flags.SerializeToString(), input_data_str, debug_info_str, ) @convert_phase( Component.CONVERT_TF_TO_TFLITE_MODEL, SubComponent.CONVERT_GRAPHDEF_USING_DEPRECATED_CONVERTER, ) def _run_deprecated_conversion_binary( model_flags_str, conversion_flags_str, input_data_str, debug_info_str=None ): """Convert `input_data_str` using deprecated conversion binary. Args: model_flags_str: Serialized proto describing model properties, see `model_flags.proto`. conversion_flags_str: Serialized proto describing TFLite converter 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 deprecated conversion binary. RuntimeError: When conversion fails, an exception is raised with the error message embedded. """ if distutils.spawn.find_executable(_deprecated_conversion_binary) 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//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. conversion_filename: str = None model_filename: str = None input_filename: str = None output_filename: str = None try: # Build all input files with _tempfile.NamedTemporaryFile( delete=False ) as fp_conversion, _tempfile.NamedTemporaryFile( delete=False ) as fp_model, _tempfile.NamedTemporaryFile( delete=False ) as fp_input, _tempfile.NamedTemporaryFile( delete=False ) as fp_debug: conversion_filename = fp_conversion.name input_filename = fp_input.name model_filename = fp_model.name debug_filename = fp_debug.name fp_model.write(model_flags_str) fp_conversion.write(conversion_flags_str) fp_input.write(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 = [ _deprecated_conversion_binary, model_filename, conversion_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 [ conversion_filename, input_filename, model_filename, output_filename, ]: try: _os.unlink(filename) except (OSError, TypeError): pass def build_model_flags( change_concat_input_ranges=False, allow_nonexistent_arrays=False, saved_model_dir=None, saved_model_version=0, saved_model_tags=None, saved_model_exported_names=None, **_ ): """Builds the model flags object from params. Args: change_concat_input_ranges: Boolean to change behavior of min/max ranges for inputs and outputs of the concat operator for quantized models. Changes the ranges of concat operator overlap when true. (default False) allow_nonexistent_arrays: Allow specifying array names that don't exist or are unused in the final graph. (default False) saved_model_dir: Filepath of the saved model to be converted. This value will be non-empty only when the saved model import path will be used. Otherwises, the graph def-based conversion will be processed. saved_model_version: SavedModel file format version of The saved model file to be converted. This value will be set only when the SavedModel import path will be used. saved_model_tags: Set of string saved model tags, formatted in the comma-separated value. This value will be set only when the SavedModel import path will be used. saved_model_exported_names: Names to be exported (default: export all) when the saved model import path is on. This value will be set only when the SavedModel import path will be used. Returns: model_flags: protocol buffer describing the model. """ model_flags = _model_flags_pb2.ModelFlags() model_flags.change_concat_input_ranges = change_concat_input_ranges model_flags.allow_nonexistent_arrays = allow_nonexistent_arrays if saved_model_dir: model_flags.saved_model_dir = saved_model_dir model_flags.saved_model_version = saved_model_version if saved_model_tags: model_flags.saved_model_tags.extend(saved_model_tags) if saved_model_exported_names: model_flags.saved_model_exported_names.extend(saved_model_exported_names) return model_flags def build_conversion_flags( inference_type=dtypes.float32, inference_input_type=None, input_format=lite_constants.TENSORFLOW_GRAPHDEF, output_format=lite_constants.TFLITE, default_ranges_stats=None, drop_control_dependency=True, reorder_across_fake_quant=False, allow_custom_ops=False, post_training_quantize=False, quantize_to_float16=False, dump_graphviz_dir=None, dump_graphviz_video=False, target_ops=None, conversion_summary_dir=None, select_user_tf_ops=None, allow_all_select_tf_ops=False, enable_tflite_resource_variables=True, unfold_batchmatmul=False, legalize_custom_tensor_list_ops=False, lower_tensor_list_ops=True, default_to_single_batch_in_tensor_list_ops=False, accumulation_type=None, allow_bfloat16=False, unfold_large_splat_constant=False, supported_backends=None, disable_per_channel_quantization=False, enable_mlir_dynamic_range_quantizer=False, tf_quantization_mode=None, disable_infer_tensor_range=False, use_fake_quant_num_bits=False, enable_dynamic_update_slice=False, preserve_assert_op=False, guarantee_all_funcs_one_use=False, enable_mlir_variable_quantization=False, disable_fuse_mul_and_fc=False, quantization_options: Optional[quant_opts_pb2.QuantizationOptions] = None, ir_dump_dir=None, ir_dump_pass_regex=None, ir_dump_func_regex=None, enable_timing=None, print_ir_before=None, print_ir_after=None, print_ir_module_scope=None, elide_elementsattrs_if_larger=None, quantization_config: Optional[ quantization_config_pb2.QuantizationConfig ] = None, use_buffer_offset=False, reduce_type_precision=False, qdq_conversion_mode=None, **_ ): """Builds protocol buffer describing a conversion of a model. Typically this is to convert from TensorFlow GraphDef to TFLite, in which case the default `input_format` and `output_format` are sufficient. Args: inference_type: Data type of numeric arrays, excluding the input layer. (default tf.float32, must be in {tf.float32, tf.int8, tf.uint8}) inference_input_type: Data type of the numeric arrays in the input layer. If `inference_input_type` is in {tf.int8, tf.uint8}, then `quantized_input_stats` must be provided. (default is the value assigned to `inference_type`, must be in {tf.float32, tf.int8, tf.uint8}) input_format: Type of data to read. (default TENSORFLOW_GRAPHDEF, must be in {TENSORFLOW_GRAPHDEF}) output_format: Output file format. (default TFLITE, must be in {TFLITE, GRAPHVIZ_DOT}) default_ranges_stats: Tuple of integers representing (min, max) range values for all arrays without a specified range. Intended for experimenting with quantization via "dummy quantization". (default None) drop_control_dependency: Boolean indicating whether to drop control dependencies silently. This is due to TFLite not supporting control dependencies. (default True) reorder_across_fake_quant: Boolean indicating whether to reorder FakeQuant nodes in unexpected locations. Used when the location of the FakeQuant nodes is preventing graph transformations necessary to convert the graph. Results in a graph that differs from the quantized training graph, potentially causing differing arithmetic behavior. (default False) allow_custom_ops: Boolean indicating whether to allow custom operations. When false any unknown operation is an error. When true, custom ops are created for any op that is unknown. The developer will need to provide these to the TensorFlow Lite runtime with a custom resolver. (default False) post_training_quantize: Boolean indicating whether to quantize the weights of the converted float model. Model size will be reduced and there will be latency improvements (at the cost of accuracy). (default False) If quantization_options is set, all quantization arg will be ignored. quantize_to_float16: Boolean indicating whether to convert float buffers to float16. (default False) dump_graphviz_dir: Full filepath of folder to dump the graphs at various stages of processing GraphViz .dot files. Preferred over --output_format=GRAPHVIZ_DOT in order to keep the requirements of the output file. (default None) dump_graphviz_video: Boolean indicating whether to dump the graph after every graph transformation. (default False) target_ops: Experimental flag, subject to change. Set of OpsSet options indicating which converter to use. (default set([OpsSet.TFLITE_BUILTINS])) conversion_summary_dir: A string, the path to the generated conversion logs. select_user_tf_ops: List of user's defined TensorFlow ops need to be supported in the TensorFlow Lite runtime. These ops will be supported as select TensorFlow ops. allow_all_select_tf_ops: If True, automatically add all TF ops (including custom TF ops) to the converted model as flex ops. enable_tflite_resource_variables: Experimental flag, subject to change. Enables conversion of resource variables. (default False) unfold_batchmatmul: Whether to unfold tf.BatchMatMul to a set of tfl.fully_connected ops. If not, translate to tfl.batch_matmul. legalize_custom_tensor_list_ops: Whether to legalize `tf.TensorList*` ops to tfl custom if they can all be supported. lower_tensor_list_ops: Whether to lower tensor list ops to builtin ops. If not, use Flex tensor list ops. default_to_single_batch_in_tensor_list_ops: Whether to force to use batch size one when the tensor list ops has the unspecified batch size. accumulation_type: Data type of the accumulators in quantized inference. Typically used for float16 quantization and is either fp16 or fp32. allow_bfloat16: Whether the converted model supports reduced precision inference with the bfloat16 type. unfold_large_splat_constant: Whether to unfold large splat constant tensors in the flatbuffer model to reduce size. supported_backends: List of TFLite backends which needs to check compatibility. disable_per_channel_quantization: Disable per-channel quantized weights for dynamic range quantization. Only per-tensor quantization will be used. enable_mlir_dynamic_range_quantizer: Enable MLIR dynamic range quantization. If False, the old converter dynamic range quantizer is used. tf_quantization_mode: Indicates the mode of TF Quantization when the output model is used for TF Quantization. disable_infer_tensor_range: Disable infering tensor ranges. use_fake_quant_num_bits: Allow quantization parameters to be calculated from num_bits attribute. enable_dynamic_update_slice: Enable to convert to DynamicUpdateSlice op. (default: False). preserve_assert_op: Whether to preserve `TF::AssertOp` (default: False). guarantee_all_funcs_one_use: Whether to clone functions so that each function only has a single use. This option will be helpful if the conversion fails when the `PartitionedCall` or `StatefulPartitionedCall` can't be properly inlined (default: False). enable_mlir_variable_quantization: Enable MLIR variable quantization. There is a variable freezing pass, but some variables may not be fully frozen by it. This flag enables quantization of those residual variables in the MLIR graph. disable_fuse_mul_and_fc: Disable fusing input multiplication with fullyconnected operations. Useful when quantizing weights. quantization_options: [Deprecated] Config to indicate quantization options of each components (ex: weight, bias, activation). This can be a preset method or a custom method, and allows finer, modular control. This option will override any other existing quantization flags. We plan on gradually migrating all quantization-related specs into this option. ir_dump_dir: A string specifying the target directory to output MLIR dumps produced during conversion. If populated, enables MLIR dumps. ir_dump_pass_regex: A string containing a regular expression for filtering the pass names to be dumped. Effective only if `ir_dump_dir` is populated. ir_dump_func_regex: A string containing a regular expression for filtering the function names to be dumped. Effective only if `ir_dump_dir` is populated. enable_timing: A boolean, if set to true reports the execution time of each MLIR pass. print_ir_before: A string containing a regular expression. If specified, prints MLIR before passes which match. print_ir_after: A string containing a regular expression. If specified, prints MLIR after passes which match. print_ir_module_scope: A boolean, if set to true always print the top-level operation when printing IR for print_ir_[before|after]. elide_elementsattrs_if_larger: An int, if specified elides ElementsAttrs with '...' that have more elements than the given upper limit. quantization_config: Configures the StableHLO Quantizer. See the comments in `QuantizationConfig` protobuf definition for details. use_buffer_offset: Force the model use buffer_offset & buffer_size fields instead of data. i.e. store the constant tensor and custom op binaries outside of Flatbuffers reduce_type_precision: Convert some tensor types to a lower precision if all values within that tensor are within the range of the lower precision. This could have side effects e.g. reduced flatbuffer size. qdq_conversion_mode: If set, assume input model is a quantized model represented with QDQ ops and convert to quantized kernels. Returns: conversion_flags: protocol buffer describing the conversion process. Raises: ValueError, if the input tensor type is unknown. """ conversion_flags = _conversion_flags_pb2.TocoFlags() conversion_flags.inference_type = convert_inference_tf_type_to_tflite_type( inference_type, usage="inference_type flag" ) if inference_input_type: conversion_flags.inference_input_type = ( convert_inference_tf_type_to_tflite_type( inference_input_type, usage="inference_input_type flag" ) ) else: conversion_flags.inference_input_type = conversion_flags.inference_type conversion_flags.input_format = input_format conversion_flags.output_format = output_format if default_ranges_stats: conversion_flags.default_ranges_min = default_ranges_stats[0] conversion_flags.default_ranges_max = default_ranges_stats[1] conversion_flags.drop_control_dependency = drop_control_dependency conversion_flags.reorder_across_fake_quant = reorder_across_fake_quant conversion_flags.allow_custom_ops = allow_custom_ops conversion_flags.post_training_quantize = post_training_quantize conversion_flags.quantize_to_float16 = quantize_to_float16 if dump_graphviz_dir: conversion_flags.dump_graphviz_dir = dump_graphviz_dir conversion_flags.dump_graphviz_include_video = dump_graphviz_video if target_ops: if OpsSet.SELECT_TF_OPS in target_ops: conversion_flags.enable_select_tf_ops = True if set(target_ops) == {OpsSet.SELECT_TF_OPS}: conversion_flags.force_select_tf_ops = True if OpsSet.EXPERIMENTAL_STABLEHLO_OPS in target_ops: conversion_flags.convert_to_stablehlo = True if OpsSet.EXPERIMENTAL_STABLEHLO_OPS in target_ops and len(target_ops) > 1: raise ValueError( "StableHLO Ops set can not be specified with other Ops set together" ) if conversion_summary_dir: conversion_flags.conversion_summary_dir = conversion_summary_dir if select_user_tf_ops: conversion_flags.select_user_tf_ops.extend(select_user_tf_ops) conversion_flags.allow_all_select_tf_ops = allow_all_select_tf_ops conversion_flags.enable_tflite_resource_variables = ( enable_tflite_resource_variables ) conversion_flags.unfold_batchmatmul = unfold_batchmatmul conversion_flags.legalize_custom_tensor_list_ops = ( legalize_custom_tensor_list_ops ) conversion_flags.lower_tensor_list_ops = lower_tensor_list_ops conversion_flags.default_to_single_batch_in_tensor_list_ops = ( default_to_single_batch_in_tensor_list_ops ) if accumulation_type: conversion_flags.accumulation_type = convert_tensor_tf_type_to_tflite_type( accumulation_type, usage="accumulation_type flag" ) conversion_flags.allow_bfloat16 = allow_bfloat16 conversion_flags.unfold_large_splat_constant = unfold_large_splat_constant if supported_backends: conversion_flags.supported_backends.extend(supported_backends) conversion_flags.disable_per_channel_quantization = ( disable_per_channel_quantization ) conversion_flags.enable_mlir_dynamic_range_quantizer = ( enable_mlir_dynamic_range_quantizer ) conversion_flags.enable_dynamic_update_slice = enable_dynamic_update_slice conversion_flags.preserve_assert_op = preserve_assert_op conversion_flags.guarantee_all_funcs_one_use = guarantee_all_funcs_one_use if tf_quantization_mode: conversion_flags.tf_quantization_mode = tf_quantization_mode conversion_flags.disable_infer_tensor_range = disable_infer_tensor_range conversion_flags.use_fake_quant_num_bits = use_fake_quant_num_bits conversion_flags.enable_mlir_variable_quantization = ( enable_mlir_variable_quantization ) conversion_flags.disable_fuse_mul_and_fc = disable_fuse_mul_and_fc if quantization_options: # Deprecated conversion_flags.quantization_options.CopyFrom(quantization_options) if quantization_config: conversion_flags.quantization_config.CopyFrom(quantization_config) # Transfer debug options. Check for existence before populating in order to # leverage defaults specified in proto definition. # TODO: b/319329480 - Match the debug_options fields with the user-facing # flags. if ir_dump_dir is not None: conversion_flags.debug_options.ir_dump_dir = ir_dump_dir if ir_dump_pass_regex is not None: conversion_flags.debug_options.ir_dump_pass_regex = ir_dump_pass_regex if ir_dump_func_regex is not None: conversion_flags.debug_options.ir_dump_func_regex = ir_dump_func_regex if enable_timing is not None: conversion_flags.debug_options.enable_timing = enable_timing if print_ir_before is not None: conversion_flags.debug_options.print_ir_before = print_ir_before if print_ir_after is not None: conversion_flags.debug_options.print_ir_after = print_ir_after if print_ir_module_scope is not None: conversion_flags.debug_options.print_ir_module_scope = print_ir_module_scope if elide_elementsattrs_if_larger is not None: conversion_flags.debug_options.elide_elementsattrs_if_larger = ( elide_elementsattrs_if_larger ) if use_buffer_offset is not None: conversion_flags.use_buffer_offset = use_buffer_offset if reduce_type_precision is not None: conversion_flags.reduce_type_precision = reduce_type_precision if qdq_conversion_mode is not None: conversion_flags.qdq_conversion_mode = qdq_conversion_mode return conversion_flags @convert_phase( Component.CONVERT_TF_TO_TFLITE_MODEL, SubComponent.CONVERT_GRAPHDEF ) def convert_graphdef_with_arrays( input_data, input_arrays_with_shape, output_arrays, control_output_arrays, **kwargs ): """Convert a frozen GraphDef that can't be loaded in TF. Conversion can be customized by providing arguments that are forwarded to `build_model_flags` and `build_conversion_flags` (see documentation). Args: input_data: Input data (i.e. often `sess.graph_def`), input_arrays_with_shape: Tuple of strings representing input tensor names and list of integers representing input shapes (e.g., [("foo" : [1, 16, 16, 3])]). Use only when graph cannot be loaded into TensorFlow and when `input_tensors` is None. output_arrays: List of output tensors to freeze graph with. Use only when graph cannot be loaded into TensorFlow and when `output_tensors` is None. control_output_arrays: Control output node names. This is used when converting a Graph with no output tensors. For example, if the graph's last operation is a Print op, just specify that op's name in this field. This can be used together with the `output_arrays` parameter. **kwargs: See `build_model_flags` and `build_conversion_flags`. Returns: The converted data. For example if TFLite was the destination, then this will be a tflite flatbuffer in a bytes array. Raises: Defined in `build_conversion_flags`. """ model_flags = build_model_flags(**kwargs) conversion_flags = build_conversion_flags(**kwargs) enable_mlir_converter = kwargs.get("enable_mlir_converter", True) quantized_input_stats = kwargs.get("quantized_input_stats", None) for idx, (name, shape) in enumerate(input_arrays_with_shape): input_array = model_flags.input_arrays.add() if _is_quantized_input_stats_required(conversion_flags): if quantized_input_stats: input_array.mean_value, input_array.std_value = quantized_input_stats[ idx ] else: raise ValueError( "The `quantized_input_stats` flag must be defined when either " "`inference_type` flag or `inference_input_type` flag is set to " "tf.int8 or tf.uint8." ) input_array.name = name input_array.shape.dims.extend(list(map(int, shape))) if output_arrays: for name in output_arrays: model_flags.output_arrays.append(name) if control_output_arrays: for name in control_output_arrays: model_flags.control_output_arrays.append(name) data = convert( model_flags, conversion_flags, input_data.SerializeToString(), debug_info_str=None, enable_mlir_converter=enable_mlir_converter, ) return data @convert_phase( Component.CONVERT_TF_TO_TFLITE_MODEL, SubComponent.CONVERT_GRAPHDEF ) def convert_graphdef(input_data, input_tensors, output_tensors, **kwargs): """Convert a frozen GraphDef model using the TF Lite converter. Conversion can be customized by providing arguments that are forwarded to `build_model_flags` and `build_conversion_flags` (see documentation). Args: input_data: Input data (i.e. often `sess.graph_def`), input_tensors: List of input tensors. Type and shape are computed using `foo.shape` and `foo.dtype`. output_tensors: List of output tensors (only .name is used from this). **kwargs: See `build_model_flags` and `build_conversion_flags`. Returns: The converted data. For example if TFLite was the destination, then this will be a tflite flatbuffer in a bytes array. Raises: Defined in `build_conversion_flags`. """ model_flags = build_model_flags(**kwargs) conversion_flags = build_conversion_flags(**kwargs) saved_model_dir = kwargs.get("saved_model_dir", None) input_shapes = kwargs.get("input_shapes", None) enable_mlir_converter = kwargs.get("enable_mlir_converter", True) quantized_input_stats = kwargs.get("quantized_input_stats", None) debug_info = kwargs.get("debug_info", None) for idx, input_tensor in enumerate(input_tensors): input_array = model_flags.input_arrays.add() if saved_model_dir: input_array.name = input_tensor.name else: input_array.name = util.get_tensor_name(input_tensor) input_array.data_type = convert_tensor_tf_type_to_tflite_type( input_tensor.dtype, usage="input type of the TensorFlow model" ) if _is_quantized_input_stats_required(conversion_flags): if quantized_input_stats: input_array.mean_value, input_array.std_value = quantized_input_stats[ idx ] else: # We should ideally raise an error here, but we don't as it would break # several models/projects that depend on this workflow. warnings.warn( "Statistics for quantized inputs were expected, but not " "specified; continuing anyway." ) if input_shapes is None: shape = input_tensor.shape else: shape = input_shapes[idx] if shape.rank is not None: # Create shapes with -1 for unknown dimensions. dims = [] for dim in shape: if dim is None or ( isinstance(dim, tensor_shape.Dimension) and dim.value is None ): dims.append(-1) else: dims.append(int(dim)) input_array.shape.dims.extend(dims) input_array.shape.unknown_rank = False else: input_array.shape.unknown_rank = True for output_tensor in output_tensors: if saved_model_dir: model_flags.output_arrays.append(output_tensor.name) else: model_flags.output_arrays.append(util.get_tensor_name(output_tensor)) data = convert( model_flags, conversion_flags, input_data.SerializeToString(), debug_info_str=debug_info.SerializeToString() if debug_info else None, enable_mlir_converter=enable_mlir_converter, ) return data @convert_phase( Component.CONVERT_TF_TO_TFLITE_MODEL, SubComponent.CONVERT_SAVED_MODEL ) def convert_saved_model(**kwargs): """Converts a SavedModel using TF Lite converter.""" model_flags = build_model_flags(**kwargs) conversion_flags = build_conversion_flags(**kwargs) data = convert( model_flags, conversion_flags, input_data_str=None, debug_info_str=None, enable_mlir_converter=True, ) return data @convert_phase( Component.CONVERT_TF_TO_TFLITE_MODEL, SubComponent.CONVERT_JAX_HLO ) def convert_jax_hlo(input_content, input_names, is_proto_format, **kwargs): """Converts a Jax hlo-based model using TFLite converter.""" model_flags = _model_flags_pb2.ModelFlags() model_flags.use_hlo_import = True if is_proto_format: model_flags.hlo_file_type = _model_flags_pb2.ModelFlags.HLO_PROTO else: model_flags.hlo_file_type = _model_flags_pb2.ModelFlags.HLO_TEXT # Build input names. for input_name in input_names: input_array = model_flags.input_arrays.add() input_array.name = input_name conversion_flags = build_conversion_flags(**kwargs) data = convert( model_flags, conversion_flags, input_data_str=input_content, debug_info_str=None, enable_mlir_converter=True, ) return data @_tf_export(v1=["lite.toco_convert"]) @deprecation.deprecated(None, "Use `lite.TFLiteConverter` instead.") def toco_convert(input_data, input_tensors, output_tensors, *args, **kwargs): """Convert a TensorFlow GraphDef to TFLite. This function is deprecated. Please use `tf.lite.TFLiteConverter` API instead. Conversion can be customized by providing arguments that are forwarded to `build_model_flags` and `build_conversion_flags` (see documentation for details). Args: input_data: Input data (i.e. often `sess.graph_def`). input_tensors: List of input tensors. Type and shape are computed using `foo.shape` and `foo.dtype`. output_tensors: List of output tensors (only .name is used from this). *args: See `build_model_flags` and `build_conversion_flags`. **kwargs: See `build_model_flags` and `build_conversion_flags`. Returns: The converted TensorFlow Lite model in a bytes array. Raises: Defined in `convert`. """ kwargs["enable_mlir_converter"] = kwargs.get("enable_mlir_converter", False) return convert_graphdef( input_data, input_tensors, output_tensors, *args, **kwargs ) def deduplicate_readonly_buffers(tflite_model): """Generates a new model byte array after deduplicating readonly buffers. This function should be invoked after the model optimization toolkit. The model optimization toolkit assumes that each tensor object owns its each buffer separately. Args: tflite_model: TFLite flatbuffer in a byte array to be deduplicated. Returns: TFLite flatbuffer in a bytes array, processed with the deduplication method. """ # Load TFLite Flatbuffer byte array into an object. model = flatbuffer_utils.convert_bytearray_to_object(tflite_model) # Get all the read-only buffers, which can be modified without causing any # issue in the graph invocation stage. read_only_buffer_indices = set() for subgraph in model.subgraphs: # To get all the read-only buffers: # (1) Get all read-only input tensors. # (2) Discard intermediate or output tensors. # (3) Discard the subgraph's input/output tensors. # (4) Gather the buffers of the read-only input tensors. # (1) Get read-only input tensors. read_only_input_tensor_indices = set() for op in subgraph.operators: if op.inputs is None: continue for i, input_tensor_idx in enumerate(op.inputs): # Ignore mutable tensors. if op.mutatingVariableInputs is not None: # Ignore invalid tensors. if ( i < len(op.mutatingVariableInputs) and op.mutatingVariableInputs[i] ): continue # Ignore variable tensors. if subgraph.tensors[input_tensor_idx].isVariable: continue read_only_input_tensor_indices.add(input_tensor_idx) # (2) Discard intermediate or output tensors. for op in subgraph.operators: if op.outputs is not None: for output_tensor_idx in op.outputs: read_only_input_tensor_indices.discard(output_tensor_idx) if op.intermediates is not None: for intermediate_tensor_idx in op.intermediates: read_only_input_tensor_indices.discard(intermediate_tensor_idx) # (3) Discard the subgraph's input and output tensors. if subgraph.inputs is not None: for input_tensor_idx in subgraph.inputs: read_only_input_tensor_indices.discard(input_tensor_idx) if subgraph.outputs is not None: for output_tensor_idx in subgraph.outputs: read_only_input_tensor_indices.discard(output_tensor_idx) # (4) Gather the buffers of the read-only input tensors. for tensor_idx in read_only_input_tensor_indices: read_only_buffer_indices.add(subgraph.tensors[tensor_idx].buffer) # Ignore invalid negative index or zero-sized buffers. for buffer_idx in read_only_buffer_indices.copy(): if buffer_idx < 0 or ( model.buffers[buffer_idx].data is None or isinstance(model.buffers[buffer_idx].data, list) or model.buffers[buffer_idx].data.size == 0 ): read_only_buffer_indices.discard(buffer_idx) class BufferIndex: """A class to store index, size, hash of the buffers in TFLite model.""" def __init__(self, idx, size, hash_value): self.idx = idx self.size = size self.hash_value = hash_value read_only_buffers = list( map( lambda index: BufferIndex( # pylint: disable=g-long-lambda index, model.buffers[index].data.size, hashlib.md5(model.buffers[index].data.data.tobytes()).hexdigest(), ), read_only_buffer_indices, ) ) # Sort read_only_buffers by buffer size & hash in descending order. read_only_buffers = sorted( read_only_buffers, key=lambda buffer: (buffer.size, buffer.hash_value), reverse=True, ) # Create a map of duplicate buffers (same size and same type). # eg: In [1, 2, 3, 4, 5, 6] if (1, 4, 6) and (2, 5) are each, groups of buffer # indices of the same size and type, then the map would be {4:1, 6:1, 5:2} duplicate_buffer_map = {} for i, buffer_i in enumerate(read_only_buffers): # This buffer is a duplicate. if buffer_i.idx in duplicate_buffer_map: continue # This buffer is unique. Scan rest of the list to find duplicates # of this buffer and mark them accordingly. for buffer_j in read_only_buffers[i + 1 :]: if buffer_j.idx in duplicate_buffer_map: continue if buffer_i.size != buffer_j.size: break if buffer_i.hash_value != buffer_j.hash_value: continue # Found duplicate. Nullify j-th buffer and use i-th buffer instead. duplicate_buffer_map[buffer_j.idx] = buffer_i.idx # Make the duplicated tensors use the single shared buffer index. for subgraph in model.subgraphs: for op in subgraph.operators: if op.inputs is None: continue for input_tensor in op.inputs: buffer_idx = subgraph.tensors[input_tensor].buffer if buffer_idx in duplicate_buffer_map: subgraph.tensors[input_tensor].buffer = duplicate_buffer_map[ buffer_idx ] # Nullify the unused buffers. for idx in duplicate_buffer_map: model.buffers[idx].data = None # Return a TFLite flatbuffer as a byte array. return flatbuffer_utils.convert_object_to_bytearray(model)