def test_saved_model_v2_quant_float16(directory, **kwargs):
"""Validates the TensorFlow SavedModel converts to a TFLite model."""
converter = _lite.TFLiteConverterV2.from_saved_model(directory)
tflite_model_float = _convert(converter, version=2, **kwargs)
interpreter_float = _lite.Interpreter(model_content=tflite_model_float)
interpreter_float.allocate_tensors()
float_tensors = interpreter_float.get_tensor_details()
tflite_model_quant = _convert(
converter,
version=2,
post_training_quantize=True,
quantize_to_float16=True,
**kwargs)
interpreter_quant = _lite.Interpreter(model_content=tflite_model_quant)
interpreter_quant.allocate_tensors()
quant_tensors = interpreter_quant.get_tensor_details()
quant_tensors_map = {
tensor_detail["name"]: tensor_detail for tensor_detail in quant_tensors
}
# Check if weights are of different types in the float and quantized models.
num_tensors_float = len(float_tensors)
num_tensors_same_dtypes = sum(
float_tensor["dtype"] == quant_tensors_map[float_tensor["name"]]["dtype"]
for float_tensor in float_tensors)
has_quant_tensor = num_tensors_float != num_tensors_same_dtypes
if not has_quant_tensor:
raise ValueError("--post_training_quantize flag was unable to quantize the "
"graph as expected.")
def _evaluate_tflite_model(tflite_model, input_data):
"""Returns evaluation of input data on TFLite model.
Args:
tflite_model: Serialized TensorFlow Lite model.
input_data: List of np.ndarray.
Returns:
List of np.ndarray.
"""
interpreter = _lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
for input_tensor, tensor_data in zip(input_details, input_data):
interpreter.set_tensor(input_tensor["index"], tensor_data)
interpreter.invoke()
output_data = [
interpreter.get_tensor(output_tensor["index"])
for output_tensor in output_details
]
output_labels = [output_tensor["name"] for output_tensor in output_details]
return output_data, output_labels
def _get_input_data_map(tflite_model, input_data):
"""Generates a map of input data based on the TFLite model.
Args:
tflite_model: Serialized TensorFlow Lite model.
input_data: List of np.ndarray.
Returns:
{str: [np.ndarray]}.
"""
interpreter = _lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
return {
input_tensor["name"]: data
for input_tensor, data in zip(input_details, input_data)
}
def _generate_random_input_data(tflite_model, seed=None):
"""Generates input data based on the input tensors in the TFLite model.
Args:
tflite_model: Serialized TensorFlow Lite model.
seed: Integer seed for the random generator. (default None)
Returns:
List of np.ndarray.
"""
interpreter = _lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
if seed:
np.random.seed(seed=seed)
return [
np.array(np.random.random_sample(input_tensor["shape"]),
dtype=input_tensor["dtype"]) for input_tensor in input_details
]
def _generate_random_input_data(tflite_model,
seed=None,
input_data_range=None):
"""Generates input data based on the input tensors in the TFLite model.
Args:
tflite_model: Serialized TensorFlow Lite model.
seed: Integer seed for the random generator. (default None)
input_data_range: A map where the key is the input tensor name and
the value is a tuple (min_val, max_val) which specifies the value range of
the corresponding input tensor. For example, '{'input1': (1, 5)}' means to
generate a random value for tensor `input1` within range [1.0, 5.0)
(half-inclusive). (default None)
Returns:
([np.ndarray], {str : [np.ndarray]}).
"""
interpreter = _lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
if seed:
np.random.seed(seed=seed)
# Generate random input data. If a tensor's value range is specified, say
# [a, b), then the generated value will be (b - a) * Unif[0.0, 1.0) + a,
# otherwise it's Unif[0.0, 1.0).
input_data = []
for input_tensor in input_details:
val = np.random.random_sample(input_tensor["shape"])
if (input_data_range is not None
and input_tensor["name"] in input_data_range):
val = (input_data_range[input_tensor["name"]][1] -
input_data_range[input_tensor["name"]][0]
) * val + input_data_range[input_tensor["name"]][0]
input_data.append(np.array(val, dtype=input_tensor["dtype"]))
input_data_map = _get_input_data_map(tflite_model, input_data)
return input_data, input_data_map
def _get_tflite_interpreter(tflite_model, input_shapes_resize=None):
"""Creates a TFLite interpreter with resized input tensors.
Args:
tflite_model: Serialized TensorFlow Lite model.
input_shapes_resize: A map where the key is the input tensor name and the
value is the shape of the input tensor. This resize happens after model
conversion, prior to calling allocate tensors. (default None)
Returns:
lite.Interpreter
"""
interpreter = _lite.Interpreter(model_content=tflite_model)
if input_shapes_resize:
input_details = interpreter.get_input_details()
input_details_map = {
detail["name"]: detail["index"] for detail in input_details
}
for name, shape in input_shapes_resize.items():
idx = input_details_map[name]
interpreter.resize_tensor_input(idx, shape)
return interpreter
def test_frozen_graph_quant(filename,
input_arrays,
output_arrays,
input_shapes=None,
**kwargs):
"""Sanity check to validate post quantize flag alters the graph.
This test does not check correctness of the converted model. It converts the
TensorFlow frozen graph to TFLite with and without the post_training_quantized
flag. It ensures some tensors have different types between the float and
quantized models in the case of an all TFLite model or mix-and-match model.
It ensures tensor types do not change in the case of an all Flex model.
Args:
filename: 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)
**kwargs: Additional arguments to be passed into the converter.
Raises:
ValueError: post_training_quantize flag doesn't act as intended.
"""
# Convert and load the float model.
converter = _lite.TFLiteConverter.from_frozen_graph(
filename, input_arrays, output_arrays, input_shapes)
tflite_model_float = _convert(converter, **kwargs)
interpreter_float = _lite.Interpreter(model_content=tflite_model_float)
interpreter_float.allocate_tensors()
float_tensors = interpreter_float.get_tensor_details()
# Convert and load the quantized model.
converter = _lite.TFLiteConverter.from_frozen_graph(filename, input_arrays,
output_arrays)
tflite_model_quant = _convert(
converter, post_training_quantize=True, **kwargs)
interpreter_quant = _lite.Interpreter(model_content=tflite_model_quant)
interpreter_quant.allocate_tensors()
quant_tensors = interpreter_quant.get_tensor_details()
quant_tensors_map = {
tensor_detail["name"]: tensor_detail for tensor_detail in quant_tensors
}
# Check if weights are of different types in the float and quantized models.
num_tensors_float = len(float_tensors)
num_tensors_same_dtypes = sum(
float_tensor["dtype"] == quant_tensors_map[float_tensor["name"]]["dtype"]
for float_tensor in float_tensors)
has_quant_tensor = num_tensors_float != num_tensors_same_dtypes
if ("target_ops" in kwargs and
set(kwargs["target_ops"]) == set([_lite.OpsSet.SELECT_TF_OPS])):
if has_quant_tensor:
raise ValueError("--post_training_quantize flag unexpectedly altered the "
"full Flex mode graph.")
elif not has_quant_tensor:
raise ValueError("--post_training_quantize flag was unable to quantize the "
"graph as expected in TFLite and mix-and-match mode.")
