tf.lite.TFLiteConverter

TensorFlow 2 version View source on GitHub

Convert a TensorFlow model into output_format.

tf.lite.TFLiteConverter(
    graph_def, input_tensors, output_tensors, input_arrays_with_shape=None,
    output_arrays=None, experimental_debug_info_func=None
)

This is used to convert from a TensorFlow GraphDef, SavedModel or tf.keras model into either a TFLite FlatBuffer or graph visualization.

Example usage:

# Converting a GraphDef from session.
converter = lite.TFLiteConverter.from_session(sess, in_tensors, out_tensors)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

# Converting a GraphDef from file.
converter = lite.TFLiteConverter.from_frozen_graph(
  graph_def_file, input_arrays, output_arrays)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

# Converting a SavedModel.
converter = lite.TFLiteConverter.from_saved_model(saved_model_dir)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

# Converting a tf.keras model.
converter = lite.TFLiteConverter.from_keras_model_file(keras_model)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

graph_def Frozen TensorFlow GraphDef.
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).
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 and output_tensors are None. (default None)
output_arrays List of output tensors to freeze graph with. Use only when graph cannot be loaded into TensorFlow and when input_tensors and output_tensors are None. (default None)
experimental_debug_info_func An experimental function to retrieve the graph debug info for a set of nodes from the graph_def.

ValueError Invalid arguments.

inference_type Target data type of real-number arrays in the output file. Must be {tf.float32, tf.uint8}. If optimzations are provided, this parameter is ignored. (default tf.float32)
inference_input_type Target data type of real-number input arrays. Allows for a different type for input arrays. If an integer type is provided and optimizations are not used, quantized_inputs_stats must be provided. If inference_type is tf.uint8, signaling conversion to a fully quantized model from a quantization-aware trained input model, then inference_input_type defaults to tf.uint8. In all other cases, inference_input_type defaults to tf.float32. Must be {tf.float32, tf.uint8, tf.int8}
inference_output_type Target data type of real-number output arrays. Allows for a different type for output arrays. If inference_type is tf.uint8, signaling conversion to a fully quantized model from a quantization-aware trained output model, then inference_output_type defaults to tf.uint8. In all other cases, inference_output_type must be tf.float32, an error will be thrown otherwise. Must be {tf.float32, tf.uint8, tf.int8}
output_format Output file format. Currently must be {TFLITE, GRAPHVIZ_DOT}. (default TFLITE)
quantized_input_stats Dict of strings representing input tensor names mapped to tuple of floats representing the mean and standard deviation of the training data (e.g., {"foo" : (0., 1.)}). Only need if inference_input_type is QUANTIZED_UINT8. real_input_value = (quantized_input_value - mean_value) / std_dev_value. (default {})
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)
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_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 Deprecated. Please specify [Optimize.DEFAULT] for optimizations instead. 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)
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 Deprecated. Please specify target_spec.supported_ops instead. Set of OpsSet options indicating which converter to use. (default set([OpsSet.TFLITE_BUILTINS]))
target_spec Experimental flag, subject to change. Specification of target device.
optimizations Experimental flag, subject to change. A list of optimizations to apply when converting the model. E.g. [Optimize.DEFAULT]
representative_dataset A representative dataset that can be used to generate input and output samples for the model. The converter can use the dataset to evaluate different optimizations.
experimental_enable_mlir_converter Experimental flag, subject to change. Enables the MLIR converter instead of the TOCO converter.

Methods

convert

View source

convert()

Converts a TensorFlow GraphDef based on instance variables.

Returns
The converted data in serialized format. Either a TFLite Flatbuffer or a Graphviz graph depending on value in output_format.

Raises
ValueError Input shape is not specified. None value for dimension in input_tensor.

from_frozen_graph

View source

@classmethod
from_frozen_graph(
    graph_def_file, input_arrays, output_arrays, input_shapes=None
)

Creates a TFLiteConverter class from a file containing a frozen GraphDef.

Args
graph_def_file Full filepath of file containing frozen GraphDef.
input_arrays List of input tensors to freeze graph with.
output_arrays List of output tensors to freeze graph with.
input_shapes Dict of strings representing input tensor names to list of integers representing input shapes (e.g., {"foo" : [1, 16, 16, 3]}). Automatically determined when input shapes is None (e.g., {"foo" : None}). (default None)

Returns
TFLiteConverter class.

Raises
IOError File not found. Unable to parse input file.
ValueError The graph is not frozen. input_arrays or output_arrays contains an invalid tensor name. input_shapes is not correctly defined when required

from_keras_model_file

View source

@classmethod
from_keras_model_file(
    model_file, input_arrays=None, input_shapes=None, output_arrays=None,
    custom_objects=None
)

Creates a TFLiteConverter class from a tf.keras model file.

Args
model_file Full filepath of HDF5 file containing the tf.keras model.
input_arrays List of input tensors to freeze graph with. Uses input arrays from SignatureDef when none are provided. (default None)
input_shapes Dict of strings representing input tensor names to list of integers representing input shapes (e.g., {"foo" : [1, 16, 16, 3]}). Automatically determined when input shapes is None (e.g., {"foo" : None}). (default None)
output_arrays List of output tensors to freeze graph with. Uses output arrays from SignatureDef when none are provided. (default None)
custom_objects Dict mapping names (strings) to custom classes or functions to be considered during model deserialization. (default None)

Returns
TFLiteConverter class.

from_saved_model

View source

@classmethod
from_saved_model(
    saved_model_dir, input_arrays=None, input_shapes=None, output_arrays=None,
    tag_set=None, signature_key=None
)

Creates a TFLiteConverter class from a SavedModel.

Args
saved_model_dir SavedModel directory to convert.
input_arrays List of input tensors to freeze graph with. Uses input arrays from SignatureDef when none are provided. (default None)
input_shapes Dict of strings representing input tensor names to list of integers representing input shapes (e.g., {"foo" : [1, 16, 16, 3]}). Automatically determined when input shapes is None (e.g., {"foo" : None}). (default None)
output_arrays List of output tensors to freeze graph with. Uses output arrays from SignatureDef when none are provided. (default None)
tag_set Set of tags identifying the MetaGraphDef within the SavedModel to analyze. All tags in the tag set must be present. (default set("serve"))
signature_key Key identifying SignatureDef containing inputs and outputs. (default DEFAULT_SERVING_SIGNATURE_DEF_KEY)

Returns
TFLiteConverter class.

from_session

View source

@classmethod
from_session(
    sess, input_tensors, output_tensors
)

Creates a TFLiteConverter class from a TensorFlow Session.

Args
sess TensorFlow Session.
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).

Returns
TFLiteConverter class.

get_input_arrays

View source

get_input_arrays()

Returns a list of the names of the input tensors.

Returns
List of strings.