tensorflow/python/compiler/tensorrt/trt_convert.py
# Copyright 2018 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.
# =============================================================================
"""Exposes the Python wrapper conversion to trt_graph."""
import collections
from functools import partial # pylint: disable=g-importing-member
import os
import platform
import sys
import tempfile
import numpy as np
import six as _six
from tensorflow.core.framework import variable_pb2
from tensorflow.core.protobuf import config_pb2
from tensorflow.core.protobuf import meta_graph_pb2
from tensorflow.core.protobuf import rewriter_config_pb2
from tensorflow.python.client import session
from tensorflow.python.compiler.tensorrt import utils as trt_utils
from tensorflow.python.eager import context
from tensorflow.python.eager import wrap_function
from tensorflow.python.framework import convert_to_constants
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import errors
from tensorflow.python.framework import importer
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor
from tensorflow.python.grappler import tf_optimizer
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import gen_resource_variable_ops
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.saved_model import builder
from tensorflow.python.saved_model import load
from tensorflow.python.saved_model import loader
from tensorflow.python.saved_model import save
from tensorflow.python.saved_model import signature_constants
from tensorflow.python.saved_model import tag_constants
from tensorflow.python.trackable import asset
from tensorflow.python.trackable import autotrackable
from tensorflow.python.trackable import resource
from tensorflow.python.training import saver
from tensorflow.python.util import deprecation
from tensorflow.python.util import nest
from tensorflow.python.util.lazy_loader import LazyLoader
from tensorflow.python.util.tf_export import tf_export
# Lazily load the op, since it's not available in cpu-only builds. Importing
# this at top will cause tests that imports TF-TRT fail when they're built
# and run without CUDA/GPU.
gen_trt_ops = LazyLoader(
"gen_trt_ops", globals(),
"tensorflow.compiler.tf2tensorrt.ops.gen_trt_ops")
_pywrap_py_utils = LazyLoader(
"_pywrap_py_utils", globals(),
"tensorflow.compiler.tf2tensorrt._pywrap_py_utils")
# Register TRT ops in python, so that when users import this module they can
# execute a TRT-converted graph without calling any of the methods in this
# module.
#
# This will call register_op_list() in
# tensorflow/python/framework/op_def_registry.py, but it doesn't register
# the op or the op kernel in C++ runtime.
try:
gen_trt_ops.trt_engine_op # pylint: disable=pointless-statement
except AttributeError:
pass
def _to_bytes(s):
"""Encode s if it is a sequence of chars."""
if isinstance(s, _six.text_type):
return s.encode("utf-8", errors="surrogateescape")
return s
def _to_string(s):
"""Decode s if it is a sequence of bytes."""
if isinstance(s, _six.binary_type):
return s.decode("utf-8")
return s
class TrtPrecisionMode(object):
FP32 = "FP32"
FP16 = "FP16"
INT8 = "INT8"
@staticmethod
def supported_precision_modes():
precisions = [
TrtPrecisionMode.FP32, TrtPrecisionMode.FP16, TrtPrecisionMode.INT8
]
return precisions + [p.lower() for p in precisions]
# Use a large enough number as the default max_workspace_size for TRT engines,
# so it can produce reasonable performance results with the default.
# For TRT >= 8.4, the recommendation is MAX_INT.
if (_pywrap_py_utils.is_tensorrt_enabled() and
trt_utils.is_loaded_tensorrt_version_greater_equal(8, 4, 0)):
# We must use `sys.maxsize - 512` to avoid overflow during casting.
DEFAULT_TRT_MAX_WORKSPACE_SIZE_BYTES = sys.maxsize - 512
else:
DEFAULT_TRT_MAX_WORKSPACE_SIZE_BYTES = 1 << 30 # 1,073,741,824
PROFILE_STRATEGY_RANGE = "Range"
PROFILE_STRATEGY_OPTIMAL = "Optimal"
PROFILE_STRATEGY_RANGE_OPTIMAL = "Range+Optimal"
PROFILE_STRATEGY_IMPLICIT_BATCH_MODE_COMPATIBLE = "ImplicitBatchModeCompatible"
def supported_profile_strategies():
return [
PROFILE_STRATEGY_RANGE, PROFILE_STRATEGY_OPTIMAL,
PROFILE_STRATEGY_RANGE_OPTIMAL,
PROFILE_STRATEGY_IMPLICIT_BATCH_MODE_COMPATIBLE
]
@tf_export("experimental.tensorrt.ConversionParams", v1=[])
class TrtConversionParams(
collections.namedtuple("TrtConversionParams", [
"max_workspace_size_bytes", "precision_mode", "minimum_segment_size",
"maximum_cached_engines", "use_calibration", "allow_build_at_runtime"
])):
"""Parameters that are used for TF-TRT conversion.
Fields:
max_workspace_size_bytes: the maximum GPU temporary memory that the TRT
engine can use at execution time. This corresponds to the
'workspaceSize' parameter of nvinfer1::IBuilder::setMaxWorkspaceSize().
precision_mode: one of the strings in
TrtPrecisionMode.supported_precision_modes().
minimum_segment_size: the minimum number of nodes required for a subgraph
to be replaced by TRTEngineOp.
maximum_cached_engines: max number of cached TRT engines for dynamic TRT
ops. Created TRT engines for a dynamic dimension are cached. If the
number of cached engines is already at max but none of them supports the
input shapes, the TRTEngineOp will fall back to run the original TF
subgraph that corresponds to the TRTEngineOp.
use_calibration: this argument is ignored if precision_mode is not INT8.
If set to True, a calibration graph will be created to calibrate the
missing ranges. The calibration graph must be converted to an inference
graph by running calibration with calibrate(). If set to False,
quantization nodes will be expected for every tensor in the graph
(excluding those which will be fused). If a range is missing, an error
will occur. Please note that accuracy may be negatively affected if
there is a mismatch between which tensors TRT quantizes and which
tensors were trained with fake quantization.
allow_build_at_runtime: whether to allow building TensorRT engines during
runtime if no prebuilt TensorRT engine can be found that can handle the
given inputs during runtime, then a new TensorRT engine is built at
runtime if allow_build_at_runtime=True, and otherwise native TF is used.
"""
def __new__(cls,
max_workspace_size_bytes=DEFAULT_TRT_MAX_WORKSPACE_SIZE_BYTES,
precision_mode=TrtPrecisionMode.FP32,
minimum_segment_size=3,
maximum_cached_engines=1,
use_calibration=True,
allow_build_at_runtime=True):
return super(TrtConversionParams,
cls).__new__(cls, max_workspace_size_bytes, precision_mode,
minimum_segment_size, maximum_cached_engines,
use_calibration, allow_build_at_runtime)
DEFAULT_TRT_CONVERSION_PARAMS = TrtConversionParams()
_TRT_ENGINE_OP_NAME = "TRTEngineOp"
def _check_conversion_params(conversion_params, is_v2=False):
"""Validate the provided TrtConversionParams.
Args:
conversion_params: a TrtConversionParams instance.
is_v2: whether we're getting a RewriterConfig for TF 2.0.
Raises:
TypeError: if any of the parameters are of unexpected type.
ValueError: if any of the parameters are of unexpected value.
"""
supported_precision_modes = TrtPrecisionMode.supported_precision_modes()
if conversion_params.precision_mode not in supported_precision_modes:
raise ValueError(
("precision mode '{}' is not supported."
"It should be one of {}").format(conversion_params.precision_mode,
supported_precision_modes))
if (conversion_params.minimum_segment_size <= 0 and
conversion_params.minimum_segment_size != -1):
raise ValueError("minimum segment size should be positive or -1 "
"(to disable main graph conversion).")
def _check_trt_version_compatibility():
"""Check compatibility of TensorRT version.
Raises:
RuntimeError: if the TensorRT library version is incompatible.
"""
if not _pywrap_py_utils.is_tensorrt_enabled():
logging.error(
"Tensorflow needs to be built with TensorRT support enabled to allow "
"TF-TRT to operate.")
raise RuntimeError("Tensorflow has not been built with TensorRT support.")
if platform.system() == "Windows":
logging.warn(
"Windows support is provided experimentally. No guarantee is made "
"regarding functionality or engineering support. Use at your own risk.")
linked_version = _pywrap_py_utils.get_linked_tensorrt_version()
loaded_version = _pywrap_py_utils.get_loaded_tensorrt_version()
logging.info("Linked TensorRT version: %s", str(linked_version))
logging.info("Loaded TensorRT version: %s", str(loaded_version))
def raise_trt_version_deprecated(version_type, trt_version):
assert version_type in [
"linked", "loaded"
], ("Incorrect value received for version_type: %s. Accepted: ['linked', "
"'loaded']") % version_type
logging.error(
"The {version_type} version of TensorRT: `{trt_version}` has now "
"been removed. Please upgrade to TensorRT 7 or more recent.".format(
version_type=version_type,
trt_version=trt_utils.version_tuple_to_string(trt_version)))
raise RuntimeError("Incompatible %s TensorRT versions" % version_type)
if not trt_utils.is_linked_tensorrt_version_greater_equal(7, 0, 0):
raise_trt_version_deprecated("linked", linked_version)
if not trt_utils.is_loaded_tensorrt_version_greater_equal(7, 0, 0):
raise_trt_version_deprecated("loaded", loaded_version)
if (loaded_version[0] != linked_version[0] or
not trt_utils.is_loaded_tensorrt_version_greater_equal(*linked_version)):
logging.error(
"Loaded TensorRT %s but linked TensorFlow against TensorRT %s. A few "
"requirements must be met:\n"
"\t-It is required to use the same major version of TensorRT during "
"compilation and runtime.\n"
"\t-TensorRT does not support forward compatibility. The loaded "
"version has to be equal or more recent than the linked version.",
trt_utils.version_tuple_to_string(loaded_version),
trt_utils.version_tuple_to_string(linked_version))
raise RuntimeError("Incompatible TensorRT major version")
elif loaded_version != linked_version:
logging.info(
"Loaded TensorRT %s and linked TensorFlow against TensorRT %s. This is "
"supported because TensorRT minor/patch upgrades are backward "
"compatible.", trt_utils.version_tuple_to_string(loaded_version),
trt_utils.version_tuple_to_string(linked_version))
def _get_tensorrt_rewriter_config(conversion_params,
is_dynamic_op=None,
max_batch_size=None,
is_v2=False,
disable_non_trt_optimizers=False,
use_implicit_batch=True,
profile_strategy=PROFILE_STRATEGY_RANGE):
"""Returns a RewriterConfig proto for TRT transformation.
Args:
conversion_params: a TrtConversionParams instance.
is_dynamic_op: whether to use dynamic engines.
max_batch_size: maximum batch size for static engines.
is_v2: whether we're getting a RewriterConfig for TF 2.0.
disable_non_trt_optimizers: Turn off all default Grappler optimizers.
use_implicit_batch: Whether to use implicit batch or explicit batch.
profile_strategy: dynamic shape optimization profile strategy.
Returns:
A RewriterConfig proto which sets a TensorRTOptimizer to run Grappler.
Raises:
TypeError: if any of the parameters are of unexpected type.
ValueError: if any of the parameters are of unexpected value.
"""
_check_conversion_params(conversion_params, is_v2=is_v2)
if is_v2 and is_dynamic_op is not None and not is_dynamic_op:
raise ValueError("is_dynamic_op is either None or True for TF2")
if not is_v2 and is_dynamic_op is None:
raise ValueError("is_dynamic_op can't be None for TF1")
if (is_dynamic_op is None or is_dynamic_op) and max_batch_size is not None:
raise ValueError("max_batch_size has to be None for TF2"
" or when is_dynamic_op == True in TF1")
if is_dynamic_op is not None and not is_dynamic_op and not isinstance(
max_batch_size, int):
raise ValueError(
"max_batch_size has to be an integer for is_dynamic_op==False in TF1")
rewriter_config_with_trt = rewriter_config_pb2.RewriterConfig()
# Disable Grappler Remapper to avoid that fused OPs that may not be
# beneficial to TF-TRT and are not supported by TF-TRT.
rewriter_config_with_trt.remapping = False
# Prevent folding of Const->QDQ chains.
rewriter_config_with_trt. \
experimental_disable_folding_quantization_emulation = (
trt_utils.is_linked_tensorrt_version_greater_equal(8, 0, 0) or
trt_utils.is_loaded_tensorrt_version_greater_equal(8, 0, 0))
if not disable_non_trt_optimizers:
rewriter_config_with_trt.optimizers.extend([
"pruning", "debug_stripper", "layout", "dependency", "constfold",
"common_subgraph_elimination"
])
rewriter_config_with_trt.meta_optimizer_iterations = (
rewriter_config_pb2.RewriterConfig.ONE)
optimizer = rewriter_config_with_trt.custom_optimizers.add()
if not disable_non_trt_optimizers:
# Add a constfold optimizer to cleanup the unused Const nodes.
rewriter_config_with_trt.custom_optimizers.add().name = "constfold"
optimizer.name = "TensorRTOptimizer"
optimizer.parameter_map[
"minimum_segment_size"].i = conversion_params.minimum_segment_size
optimizer.parameter_map["max_workspace_size_bytes"].i = (
conversion_params.max_workspace_size_bytes)
optimizer.parameter_map["precision_mode"].s = _to_bytes(
conversion_params.precision_mode)
optimizer.parameter_map[
"maximum_cached_engines"].i = conversion_params.maximum_cached_engines
optimizer.parameter_map[
"use_calibration"].b = conversion_params.use_calibration
optimizer.parameter_map["is_dynamic_op"].b = is_dynamic_op
optimizer.parameter_map[
"allow_build_at_runtime"].b = conversion_params.allow_build_at_runtime
if max_batch_size is not None:
optimizer.parameter_map["max_batch_size"].i = max_batch_size
optimizer.parameter_map["use_implicit_batch"].b = use_implicit_batch
# While we accept case insensitive strings from the users, we only pass the
# strings in lower cases to TF-TRT converter.
if not use_implicit_batch:
optimizer.parameter_map["profile_strategy"].s = _to_bytes(
profile_strategy.lower())
# Disabling optimizers should happen after defining the TF-TRT grappler pass
# otherwise the template can overwrite the disablement.
if disable_non_trt_optimizers:
trt_utils.disable_non_trt_optimizers_in_rewriter_config(
rewriter_config_with_trt)
return rewriter_config_with_trt
@deprecation.deprecated(
None, "You shouldn't need a rewriter_config with the current TF-TRT APIs.")
def get_tensorrt_rewriter_config(conversion_params,
is_dynamic_op=None,
max_batch_size=None,
is_v2=False,
disable_non_trt_optimizers=False):
return _get_tensorrt_rewriter_config(conversion_params, is_dynamic_op,
max_batch_size, is_v2,
disable_non_trt_optimizers)
# Remove all scope prefixes in the node name. In TF 2.0, the same concrete
# function can be initialized multiple times with different prefixes, and
# this will result in the same TRTEngineOp being initialized multiple times
# with different cache and duplicate TRT engines.
# TODO(laigd): this may be caused by the fact that TRTEngineOp is not
# stateful, need to investigate.
# TODO(laigd): we rely on the fact that all functions are fully inlined
# before TF-TRT optimizer is called, as otherwise it may generate the same
# name when optimizing a different function graph. Fix this.
def _get_canonical_engine_name(name):
return name.split("/")[-1]
class TrtGraphConverter(object):
"""A converter for TF-TRT transformation for TF 1.x GraphDef/SavedModels.
To run the conversion without quantization calibration (e.g. for FP32/FP16
precision modes):
```python
converter = TrtGraphConverter(
input_saved_model_dir="my_dir",
precision_mode=TrtPrecisionMode.FP16)
converted_graph_def = converter.convert()
converter.save(output_saved_model_dir)
```
To run the conversion with quantization calibration:
```python
converter = TrtGraphConverter(
input_saved_model_dir="my_dir",
precision_mode=TrtPrecisionMode.INT8)
converter.convert()
# Run calibration 10 times.
converted_graph_def = converter.calibrate(
fetch_names=['output:0'],
num_runs=10,
feed_dict_fn=lambda: {'input:0': my_next_data()})
converter.save(output_saved_model_dir)
```
"""
def __init__(self,
input_saved_model_dir=None,
input_saved_model_tags=None,
input_saved_model_signature_key=None,
input_graph_def=None,
nodes_denylist=None,
max_batch_size=1,
max_workspace_size_bytes=DEFAULT_TRT_MAX_WORKSPACE_SIZE_BYTES,
precision_mode=TrtPrecisionMode.FP32,
minimum_segment_size=3,
is_dynamic_op=False,
maximum_cached_engines=1,
use_calibration=True):
"""Initializes the converter.
Args:
input_saved_model_dir: the directory to load the SavedModel which contains
the input graph to transforms. Used only when input_graph_def is None.
input_saved_model_tags: list of tags to load the SavedModel.
input_saved_model_signature_key: the key of the signature to optimize the
graph for.
input_graph_def: a GraphDef object containing a model to be transformed.
If set to None, the graph will be read from the SavedModel loaded from
input_saved_model_dir.
nodes_denylist: list of node names to prevent the converter from touching.
max_batch_size: max size for the input batch.
max_workspace_size_bytes: the maximum GPU temporary memory which the TRT
engine can use at execution time. This corresponds to the
'workspaceSize' parameter of nvinfer1::IBuilder::setMaxWorkspaceSize().
precision_mode: one of TrtPrecisionMode.supported_precision_modes().
minimum_segment_size: the minimum number of nodes required for a subgraph
to be replaced by TRTEngineOp.
is_dynamic_op: whether to generate dynamic TRT ops which will build the
TRT network and engine at run time.
maximum_cached_engines: max number of cached TRT engines in dynamic TRT
ops. If the number of cached engines is already at max but none of them
can serve the input, the TRTEngineOp will fall back to run the TF
function based on which the TRTEngineOp is created.
use_calibration: this argument is ignored if precision_mode is not INT8.
If set to True, a calibration graph will be created to calibrate the
missing ranges. The calibration graph must be converted to an inference
graph by running calibration with calibrate(). If set to False,
quantization nodes will be expected for every tensor in the graph
(excluding those which will be fused). If a range is missing, an error
will occur. Please note that accuracy may be negatively affected if
there is a mismatch between which tensors TRT quantizes and which
tensors were trained with fake quantization.
Raises:
ValueError: if the combination of the parameters is invalid.
RuntimeError: if this class is used in TF 2.0.
"""
if context.executing_eagerly():
raise RuntimeError(
"Please use tf.experimental.tensorrt.Converter in TF 2.0.")
if input_graph_def and input_saved_model_dir:
raise ValueError(
"Can only specify one of input_graph_def and input_saved_model_dir")
if not input_graph_def and not input_saved_model_dir:
raise ValueError("Must specify one of input_graph_def and "
"input_saved_model_dir")
_check_trt_version_compatibility()
self._input_graph_def = input_graph_def
self._nodes_denylist = nodes_denylist
self._input_saved_model_dir = input_saved_model_dir
self._converted = False
self._grappler_meta_graph_def = None
self._input_saved_model_tags = (
input_saved_model_tags or [tag_constants.SERVING])
self._input_saved_model_signature_key = (
input_saved_model_signature_key or
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY)
# For calibration usage.
self._calibration_graph = None
self._calibration_data_collected = False
self._need_calibration = (
((precision_mode == TrtPrecisionMode.INT8) or
(precision_mode == TrtPrecisionMode.INT8.lower())) and use_calibration)
if self._need_calibration and not is_dynamic_op:
logging.warn(
"INT8 precision mode with calibration is supported with "
"dynamic TRT ops only. Disregarding is_dynamic_op parameter.")
is_dynamic_op = True
self._is_dynamic_op = is_dynamic_op
if is_dynamic_op:
self._max_batch_size = None
if max_batch_size is not None:
logging.warn("When is_dynamic_op==True max_batch_size should be None")
else:
if not isinstance(max_batch_size, int):
raise ValueError("When is_dynamic_op==False max_batch_size should be "
"an integer")
self._max_batch_size = max_batch_size
self._conversion_params = TrtConversionParams(
max_workspace_size_bytes=max_workspace_size_bytes,
precision_mode=precision_mode,
minimum_segment_size=minimum_segment_size,
maximum_cached_engines=maximum_cached_engines,
use_calibration=use_calibration,
allow_build_at_runtime=True)
_check_conversion_params(self._conversion_params)
self._test_only_disable_non_trt_optimizers = False
def _run_conversion(self):
"""Run Grappler's OptimizeGraph() tool to convert the graph."""
# Create custom ConfigProto for Grappler.
grappler_session_config = config_pb2.ConfigProto()
custom_rewriter_config = _get_tensorrt_rewriter_config(
conversion_params=self._conversion_params,
is_dynamic_op=self._is_dynamic_op,
max_batch_size=self._max_batch_size,
disable_non_trt_optimizers=self._test_only_disable_non_trt_optimizers,
use_implicit_batch=True)
grappler_session_config.graph_options.rewrite_options.CopyFrom(
custom_rewriter_config)
# Run Grappler.
self._converted_graph_def = tf_optimizer.OptimizeGraph(
grappler_session_config,
self._grappler_meta_graph_def,
graph_id=b"tf_graph")
self._converted = True
def _add_nodes_denylist(self):
if self._nodes_denylist:
collection_def = self._grappler_meta_graph_def.collection_def["train_op"]
denylist = collection_def.node_list.value
for i in self._nodes_denylist:
if isinstance(i, tensor.Tensor):
denylist.append(_to_bytes(i.name))
else:
denylist.append(_to_bytes(i))
def _convert_graph_def(self):
"""Convert the input GraphDef."""
graph = ops.Graph()
with graph.as_default():
importer.import_graph_def(self._input_graph_def, name="")
self._grappler_meta_graph_def = saver.export_meta_graph(
graph_def=graph.as_graph_def(add_shapes=True), graph=graph)
self._add_nodes_denylist()
self._run_conversion()
def _collections_to_keep(self, collection_keys):
# TODO(laigd): currently we use the collection key to filter out
# collections that depend on variable ops, but this may miss some
# other user-defined collections. A better way would be to use
# CollectionDef::NodeList for the filtering.
collections_to_remove = (
ops.GraphKeys._VARIABLE_COLLECTIONS + [
ops.GraphKeys.TRAIN_OP, ops.GraphKeys.WHILE_CONTEXT,
ops.GraphKeys.COND_CONTEXT
])
return [key for key in collection_keys if key not in collections_to_remove]
def _convert_saved_model(self):
"""Convert the input SavedModel."""
graph = ops.Graph()
with session.Session(graph=graph) as sess:
input_meta_graph_def = loader.load(sess, self._input_saved_model_tags,
self._input_saved_model_dir)
input_signature_def = input_meta_graph_def.signature_def[
self._input_saved_model_signature_key]
def _gather_names(tensor_info):
"""Get the node names from a TensorInfo."""
return {tensor_info[key].name.split(":")[0] for key in tensor_info}
# Get input and outputs from all SignatureDef.
output_node_names = _gather_names(input_signature_def.inputs).union(
_gather_names(input_signature_def.outputs))
# Preserve nodes in collection
for collection_key in self._collections_to_keep(
input_meta_graph_def.collection_def):
for op in sess.graph.get_collection(collection_key):
if isinstance(op, ops.Operation):
output_node_names.add(op.name.split(":")[0])
# Freeze the variables in the SavedModel graph and copy the frozen
# graph over.
frozen_graph_def = convert_to_constants.convert_variables_to_constants(
sess, sess.graph.as_graph_def(add_shapes=True),
list(output_node_names))
self._grappler_meta_graph_def = meta_graph_pb2.MetaGraphDef()
self._grappler_meta_graph_def.graph_def.CopyFrom(frozen_graph_def)
# Copy the collections that are not variables.
for collection_key in self._collections_to_keep(
input_meta_graph_def.collection_def):
self._grappler_meta_graph_def.collection_def[collection_key].CopyFrom(
input_meta_graph_def.collection_def[collection_key])
self._add_nodes_denylist()
# Copy other information.
self._grappler_meta_graph_def.meta_info_def.CopyFrom(
input_meta_graph_def.meta_info_def)
self._grappler_meta_graph_def.signature_def[
self._input_saved_model_signature_key].CopyFrom(input_signature_def)
# TODO(laigd): maybe add back AssetFileDef.
self._run_conversion()
def convert(self):
"""Run the TF-TRT conversion.
Returns:
The converted GraphDef for TF 1.x.
"""
assert not self._converted
if self._input_graph_def:
self._convert_graph_def()
else:
self._convert_saved_model()
return self._converted_graph_def
def calibrate(self,
fetch_names,
num_runs,
feed_dict_fn=None,
input_map_fn=None):
"""Run the calibration and return the calibrated GraphDef.
Args:
fetch_names: a list of output tensor name to fetch during calibration.
num_runs: number of runs of the graph during calibration.
feed_dict_fn: a function that returns a dictionary mapping input names (as
strings) in the GraphDef to be calibrated to values (e.g. Python list,
numpy arrays, etc). One and only one of `feed_dict_fn` and
`input_map_fn` should be specified.
input_map_fn: a function that returns a dictionary mapping input names (as
strings) in the GraphDef to be calibrated to Tensor objects. The values
of the named input tensors in the GraphDef to be calibrated will be
re-mapped to the respective `Tensor` values during calibration. One and
only one of `feed_dict_fn` and `input_map_fn` should be specified.
Raises:
ValueError: if the input combination is invalid.
RuntimeError: if this method is called in eager mode.
Returns:
The GraphDef after the calibration.
"""
assert self._converted
assert self._need_calibration
assert not self._calibration_data_collected
if (feed_dict_fn and input_map_fn) or (not feed_dict_fn and
not input_map_fn):
raise ValueError(
"Should specify one and only one of feed_dict_fn and input_map_fn.")
if input_map_fn:
for k, v in input_map_fn().items():
if not isinstance(k, str):
raise ValueError("Keys of input_map_fn must be of type str")
if not isinstance(v, tensor.Tensor):
raise ValueError("Values of input_map_fn must be of type tf.Tensor")
self._calibration_graph = ops.Graph()
with self._calibration_graph.as_default():
fetches = importer.import_graph_def(
self._converted_graph_def,
input_map=input_map_fn() if input_map_fn else None,
return_elements=fetch_names,
name="")
calibrate_rewriter_cfg = rewriter_config_pb2.RewriterConfig()
if self._test_only_disable_non_trt_optimizers:
trt_utils.disable_non_trt_optimizers_in_rewriter_config(
calibrate_rewriter_cfg)
# Set allow_soft_placement=True to run the graph for calibration so that
# OPs supported by TensorRT but don't have a GPU implementation are allowed
# to execute on CPU.
calibrate_config = config_pb2.ConfigProto(
allow_soft_placement=True,
graph_options=config_pb2.GraphOptions(
rewrite_options=calibrate_rewriter_cfg))
with session.Session(
graph=self._calibration_graph,
config=calibrate_config) as calibration_sess:
for _ in range(num_runs):
calibration_sess.run(
fetches, feed_dict=feed_dict_fn() if feed_dict_fn else None)
# Maps device name to the corresponding get_calibration_data.
#
# TODO(laigd): a better way would be to use calibration_sess to list
# all the devices, add one get_calibration_data for each device, and
# fetch each such op for every resource until its found. This can work
# even when the device of the TRTEngineOp is empty or not fully specified.
device_to_get_resource_op_map = {}
with self._calibration_graph.as_default():
resource_name_input = array_ops.placeholder(dtypes.string)
for node in self._converted_graph_def.node:
if node.op == _TRT_ENGINE_OP_NAME:
# Adds the get_calibration_data op for the device if not done
# before. We only add one such op for each device.
# TODO(laigd): What if the device is empty?????
if node.device not in device_to_get_resource_op_map:
with self._calibration_graph.device(node.device):
serialized_resources_output = (
gen_trt_ops.get_calibration_data_op(resource_name_input))
device_to_get_resource_op_map[node.device] = (
serialized_resources_output)
# Get the calibration resource.
calibration_result = calibration_sess.run(
device_to_get_resource_op_map[node.device],
feed_dict={
resource_name_input: _get_canonical_engine_name(node.name)
})
node.attr["calibration_data"].s = calibration_result
self._calibration_data_collected = True
return self._converted_graph_def
def save(self, output_saved_model_dir):
"""Save the converted graph as a SavedModel.
Args:
output_saved_model_dir: construct a SavedModel using the converted
GraphDef and save it to the specified directory. This option only works
when the input graph is loaded from a SavedModel, i.e. when
input_saved_model_dir is specified and input_graph_def is None in
__init__().
Raises:
ValueError: if the input to the converter is a GraphDef instead of a
SavedModel.
"""
assert self._converted
if self._need_calibration:
assert self._calibration_data_collected
if self._input_graph_def:
raise ValueError(
"Not able to save to a SavedModel since input is a GraphDef")
def _restore_collections(dest_graph, src_meta_graph_def, collection_keys):
"""Restores collections that we need to keep."""
scope = ""
for key in collection_keys:
collection_def = src_meta_graph_def.collection_def[key]
kind = collection_def.WhichOneof("kind")
if kind is None:
logging.error(
"Cannot identify data type for collection %s. Skipping.", key)
continue
from_proto = ops.get_from_proto_function(key)
if from_proto and kind == "bytes_list":
proto_type = ops.get_collection_proto_type(key)
# It is assumed that there are no Variables Keys in collections
for value in collection_def.bytes_list.value:
proto = proto_type()
proto.ParseFromString(value)
try:
new_value = from_proto(proto, import_scope=scope)
except:
continue
dest_graph.add_to_collection(key, new_value)
else:
field = getattr(collection_def, kind)
if kind == "node_list":
for value in field.value:
name = ops.prepend_name_scope(value, scope)
# Since the graph has been optimized, the node may no longer
# exists
try:
col_op = dest_graph.as_graph_element(name)
except (TypeError, ValueError, KeyError):
continue
dest_graph.add_to_collection(key, col_op)
elif kind == "int64_list":
# NOTE(opensource): This force conversion is to work around the
# fact that Python2 distinguishes between int and long, while
# Python3 has only int.
for value in field.value:
dest_graph.add_to_collection(key, int(value))
else:
for value in field.value:
dest_graph.add_to_collection(key,
ops.prepend_name_scope(value, scope))
# Write the transformed graphdef as SavedModel.
saved_model_builder = builder.SavedModelBuilder(output_saved_model_dir)
with ops.Graph().as_default():
importer.import_graph_def(self._converted_graph_def, name="")
_restore_collections(
ops.get_default_graph(), self._grappler_meta_graph_def,
self._collections_to_keep(
self._grappler_meta_graph_def.collection_def))
# We don't use any specific converter here.
with session.Session() as sess:
saved_model_builder.add_meta_graph_and_variables(
sess,
self._input_saved_model_tags,
signature_def_map=self._grappler_meta_graph_def.signature_def)
# Ignore other meta graphs from the input SavedModel.
saved_model_builder.save()
def _get_resource_handle(name, device):
with ops.device(device):
return gen_trt_ops.create_trt_resource_handle(resource_name=name)
def _remove_native_segments(input_func):
"""Remove native segments from the input TF-TRT Converted Function.
Args:
input_func: provide the concrete function with native segment nodes. The
transformed output func will not contain any native segment nodes. All the
TRTEngineOp references will be deleted and reset to default empty func.
"""
input_graph_def = input_func.graph.as_graph_def()
# Deleting the Native Segment node in each TRTEngineOp node.
nodes_deleted = 0
for func_id in reversed(range(len(input_graph_def.library.function))):
f = input_graph_def.library.function[func_id]
if "native_segment" in f.signature.name:
nodes_deleted += 1
while context.context().has_function(f.signature.name):
context.context().remove_function(f.signature.name)
del input_graph_def.library.function[func_id]
logging.info(
"Found and deleted native segments from "
f"{nodes_deleted} TRTEngineOp nodes."
)
# Deleting the references to `<EngineName>_native_segment`s.
# This helps TRTEngineOp constructor to not look for native segment handles
# during construction of graph for inference.
for node in input_graph_def.node:
if node.op == "TRTEngineOp":
del node.attr["segment_func"]
for func in input_graph_def.library.function:
for node in func.node_def:
if node.op == "TRTEngineOp":
del node.attr["segment_func"]
# Reconstruct the converted_func with the new graph
new_func = _construct_function_from_graph_def(input_func, input_graph_def)
return new_func
class _TRTEngineResource(resource.TrackableResource):
"""Class to track the serialized engines resource."""
def __init__(self,
resource_name,
filename,
maximum_cached_engines,
device="GPU"):
super(_TRTEngineResource, self).__init__(device=device)
self._resource_name = resource_name
# Track the serialized engine file in the SavedModel.
self._filename = self._track_trackable(
asset.Asset(filename), "_serialized_trt_resource_filename")
self._maximum_cached_engines = maximum_cached_engines
def _create_resource(self):
return _get_resource_handle(self._resource_name, self._resource_device)
def _initialize(self):
gen_trt_ops.initialize_trt_resource(
self.resource_handle,
self._filename,
max_cached_engines_count=self._maximum_cached_engines)
def _destroy_resource(self):
handle = _get_resource_handle(self._resource_name, self._resource_device)
with ops.device(self._resource_device):
gen_resource_variable_ops.destroy_resource_op(
handle, ignore_lookup_error=True)
def _print_row(fields, positions, print_fn):
"""Prints a row."""
line = ""
for i, field in enumerate(fields):
field = str(field)
end_line_pos = positions[i]
if i > 0:
line = line + " "
line = "{0:{min_length}}".format(line + field, min_length=end_line_pos)
if len(line) > end_line_pos:
line = line[:(end_line_pos - 4)] + " ..."
print_fn(line)
def _construct_function_from_graph_def(func, graph_def, frozen_func=None):
"""Rebuild function from graph_def."""
if frozen_func is None:
frozen_func = func
# If a function is converted, then the TF context contains the original
# function while the converted_graph_def contains the converted function.
# Remove the original function from the TF context in this case.
for f in graph_def.library.function:
while context.context().has_function(f.signature.name):
context.context().remove_function(f.signature.name)
captures = {
c[1].name.split(":")[0]: c[0]
for c in frozen_func.graph.captures
}
new_func = wrap_function.function_from_graph_def(
graph_def, [tensor.name for tensor in frozen_func.inputs],
[tensor.name for tensor in frozen_func.outputs], captures)
new_func.graph.structured_outputs = nest.pack_sequence_as(
func.graph.structured_outputs, new_func.graph.structured_outputs)
new_func._function_type = func.function_type # pylint: disable=protected-access
# Copy structured input signature from original function (used during
# serialization)
new_func.graph.structured_input_signature = (func.structured_input_signature)
return new_func
def _apply_inlining(func):
"""Apply an inlining optimization to the function's graph definition."""
graph_def = func.graph.as_graph_def()
# In some cases, a secondary implementation of the function (e.g. for GPU) is
# written to the "api_implements" attribute. (e.g. `tf.keras.layers.LSTM` in
# TF2 produces a CuDNN-based RNN for GPU).
# This function suppose to inline all functions calls, but "api_implements"
# prevents this from happening. Removing the attribute solves the problem.
# To learn more about "api_implements", see:
# tensorflow/core/grappler/optimizers/implementation_selector.h
for function in graph_def.library.function:
if "api_implements" in function.attr:
del function.attr["api_implements"]
meta_graph = saver.export_meta_graph(graph_def=graph_def, graph=func.graph)
# Clear the initializer_name for the variables collections, since they are not
# needed after saved to saved_model.
for name in [
"variables", "model_variables", "trainable_variables", "local_variables"
]:
raw_list = []
for raw in meta_graph.collection_def["variables"].bytes_list.value:
variable = variable_pb2.VariableDef()
variable.ParseFromString(raw)
variable.ClearField("initializer_name")
raw_list.append(variable.SerializeToString())
meta_graph.collection_def[name].bytes_list.value[:] = raw_list
# Add a collection 'train_op' so that Grappler knows the outputs.
fetch_collection = meta_graph_pb2.CollectionDef()
for array in func.inputs + func.outputs:
fetch_collection.node_list.value.append(array.name)
meta_graph.collection_def["train_op"].CopyFrom(fetch_collection)
# Initialize RewriterConfig with everything disabled except function inlining.
config = config_pb2.ConfigProto()
rewrite_options = config.graph_options.rewrite_options
rewrite_options.min_graph_nodes = -1 # do not skip small graphs
rewrite_options.optimizers.append("function")
new_graph_def = tf_optimizer.OptimizeGraph(config, meta_graph)
return new_graph_def
def _annotate_variable_ops(func, graph_def):
"""Annotates variable operations with custom `_shape` attribute.
This is required for the converters and shape inference. The graph
definition is modified in-place.
Args:
func: Function represented by the graph definition.
graph_def: Graph definition to be annotated in-place.
Raises:
RuntimeError: if some shapes cannot be annotated.
"""
ph_shape_map = {}
for ph, var in zip(func.graph.internal_captures, func.variables):
ph_shape_map[ph.name] = var.shape
# Construct a mapping of node names to nodes
name_to_node = {node.name: node for node in graph_def.node}
# Go through all the ReadVariableOp nodes in the graph def
for node in graph_def.node:
if node.op == "ReadVariableOp" or node.op == "ResourceGather":
node_ = node
# Go up the chain of identities to find a placeholder
while name_to_node[node_.input[0]].op == "Identity":
node_ = name_to_node[node_.input[0]]
ph_name = node_.input[0] + ":0"
if ph_name in ph_shape_map:
shape = ph_shape_map[ph_name]
node.attr["_shape"].shape.CopyFrom(shape.as_proto())
else:
raise RuntimeError(
"Not found in the function captures: {}".format(ph_name))
def _save_calibration_table(node):
try:
calibration_table = gen_trt_ops.get_calibration_data_op(
_get_canonical_engine_name(node.name))
node.attr["calibration_data"].s = calibration_table.numpy()
except (errors.UnknownError, errors.NotFoundError):
logging.warning("Warning calibration error for %s", node.name)
def _convert_to_tensor(inp):
try:
if isinstance(inp, dict):
args = []
kwargs = {k: ops.convert_to_tensor(v) for k, v in inp.items()}
else:
kwargs = {}
if isinstance(inp, (list, tuple)):
args = map(ops.convert_to_tensor, inp)
else:
args = [ops.convert_to_tensor(inp)]
except:
error_msg = "Failed to convert input to tensor."
logging.error(error_msg + "\ninp = `{0}`\n".format(inp))
raise RuntimeError(error_msg)
return args, kwargs
@tf_export("experimental.tensorrt.Converter", v1=[])
class TrtGraphConverterV2(object):
"""An offline converter for TF-TRT transformation for TF 2.0 SavedModels.
Windows support is provided experimentally. No guarantee is made regarding
functionality or engineering support. Use at your own risk.
There are several ways to run the conversion:
1. FP32/FP16 precision
```python
params = tf.experimental.tensorrt.ConversionParams(
precision_mode='FP16')
converter = tf.experimental.tensorrt.Converter(
input_saved_model_dir="my_dir", conversion_params=params)
converter.convert()
converter.save(output_saved_model_dir)
```
In this case, no TRT engines will be built or saved in the converted
SavedModel. But if input data is available during conversion, we can still
build and save the TRT engines to reduce the cost during inference (see
option 2 below).
2. FP32/FP16 precision with pre-built engines
```python
params = tf.experimental.tensorrt.ConversionParams(
precision_mode='FP16',
# Set this to a large enough number so it can cache all the engines.
maximum_cached_engines=16)
converter = tf.experimental.tensorrt.Converter(
input_saved_model_dir="my_dir", conversion_params=params)
converter.convert()
# Define a generator function that yields input data, and use it to execute
# the graph to build TRT engines.
def my_input_fn():
for _ in range(num_runs):
inp1, inp2 = ...
yield inp1, inp2
converter.build(input_fn=my_input_fn) # Generate corresponding TRT engines
converter.save(output_saved_model_dir) # Generated engines will be saved.
```
In this way, one engine will be built/saved for each unique input shapes of
the TRTEngineOp. This is good for applications that cannot afford building
engines during inference but have access to input data that is similar to
the one used in production (for example, that has the same input shapes).
Also, the generated TRT engines is platform dependent, so we need to run
`build()` in an environment that is similar to production (e.g. with
same type of GPU).
3. INT8 precision and calibration with pre-built engines
```python
params = tf.experimental.tensorrt.ConversionParams(
precision_mode='INT8',
# Currently only one INT8 engine is supported in this mode.
maximum_cached_engines=1,
use_calibration=True)
converter = tf.experimental.tensorrt.Converter(
input_saved_model_dir="my_dir", conversion_params=params)
# Define a generator function that yields input data, and run INT8
# calibration with the data. All input data should have the same shape.
# At the end of convert(), the calibration stats (e.g. range information)
# will be saved and can be used to generate more TRT engines with different
# shapes. Also, one TRT engine will be generated (with the same shape as
# the calibration data) for save later.
def my_calibration_input_fn():
for _ in range(num_runs):
inp1, inp2 = ...
yield inp1, inp2
converter.convert(calibration_input_fn=my_calibration_input_fn)
# (Optional) Generate more TRT engines offline (same as the previous
# option), to avoid the cost of generating them during inference.
def my_input_fn():
for _ in range(num_runs):
inp1, inp2 = ...
yield inp1, inp2
converter.build(input_fn=my_input_fn)
# Save the TRT engine and the engines.
converter.save(output_saved_model_dir)
```
4. To use dynamic shape, we need to call the build method with an input
function to generate profiles. This step is similar to the INT8 calibration
step described above. The converter also needs to be created with
use_dynamic_shape=True and one of the following profile_strategies for
creating profiles based on the inputs produced by the input function:
* `Range`: create one profile that works for inputs with dimension values
in the range of [min_dims, max_dims] where min_dims and max_dims are
derived from the provided inputs.
* `Optimal`: create one profile for each input. The profile only works for
inputs with the same dimensions as the input it is created for. The GPU
engine will be run with optimal performance with such inputs.
* `Range+Optimal`: create the profiles for both `Range` and `Optimal`.
"""
def _verify_profile_strategy(self, strategy):
supported_strategies = [s.lower() for s in supported_profile_strategies()]
if strategy.lower() not in supported_strategies:
raise ValueError(
("profile_strategy '{}' is not supported. It should be one of {}"
).format(strategy, supported_profile_strategies()))
if strategy == "ImplicitBatchModeCompatible":
logging.warn(
"ImplicitBatchModeCompatible strategy is deprecated, and"
" using it may result in errors during engine building. Please"
" consider using a different profile strategy.")
@deprecation.deprecated_args(None,
"Use individual converter parameters instead",
"conversion_params")
def __init__(self,
input_saved_model_dir=None,
input_saved_model_tags=None,
input_saved_model_signature_key=None,
use_dynamic_shape=None,
dynamic_shape_profile_strategy=None,
max_workspace_size_bytes=DEFAULT_TRT_MAX_WORKSPACE_SIZE_BYTES,
precision_mode=TrtPrecisionMode.FP32,
minimum_segment_size=3,
maximum_cached_engines=1,
use_calibration=True,
allow_build_at_runtime=True,
conversion_params=None):
"""Initialize the converter.
Args:
input_saved_model_dir: the directory to load the SavedModel which contains
the input graph to transforms. Required.
input_saved_model_tags: list of tags to load the SavedModel.
input_saved_model_signature_key: the key of the signature to optimize the
graph for.
use_dynamic_shape: whether to enable dynamic shape support. None is
equivalent to False in the current implementation.
dynamic_shape_profile_strategy: one of the strings in
supported_profile_strategies(). None is equivalent to Range in the
current implementation.
max_workspace_size_bytes: the maximum GPU temporary memory that the TRT
engine can use at execution time. This corresponds to the
'workspaceSize' parameter of nvinfer1::IBuilder::setMaxWorkspaceSize().
precision_mode: one of the strings in
TrtPrecisionMode.supported_precision_modes().
minimum_segment_size: the minimum number of nodes required for a subgraph
to be replaced by TRTEngineOp.
maximum_cached_engines: max number of cached TRT engines for dynamic TRT
ops. Created TRT engines for a dynamic dimension are cached. If the
number of cached engines is already at max but none of them supports the
input shapes, the TRTEngineOp will fall back to run the original TF
subgraph that corresponds to the TRTEngineOp.
use_calibration: this argument is ignored if precision_mode is not INT8.
If set to True, a calibration graph will be created to calibrate the
missing ranges. The calibration graph must be converted to an inference
graph by running calibration with calibrate(). If set to False,
quantization nodes will be expected for every tensor in the graph
(excluding those which will be fused). If a range is missing, an error
will occur. Please note that accuracy may be negatively affected if
there is a mismatch between which tensors TRT quantizes and which
tensors were trained with fake quantization.
allow_build_at_runtime: whether to allow building TensorRT engines during
runtime if no prebuilt TensorRT engine can be found that can handle the
given inputs during runtime, then a new TensorRT engine is built at
runtime if allow_build_at_runtime=True, and otherwise native TF is used.
conversion_params: a TrtConversionParams instance (deprecated).
Raises:
ValueError: if the combination of the parameters is invalid.
"""
assert context.executing_eagerly()
if conversion_params is None:
conversion_params = TrtConversionParams(
max_workspace_size_bytes=max_workspace_size_bytes,
precision_mode=precision_mode,
minimum_segment_size=minimum_segment_size,
maximum_cached_engines=maximum_cached_engines,
use_calibration=use_calibration,
allow_build_at_runtime=allow_build_at_runtime)
_check_trt_version_compatibility()
_check_conversion_params(conversion_params, is_v2=True)
self._conversion_params = conversion_params
self._input_saved_model_dir = input_saved_model_dir
self._input_saved_model_tags = (
input_saved_model_tags or [tag_constants.SERVING])
self._input_saved_model_signature_key = (
input_saved_model_signature_key or
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY)
self.freeze = not trt_utils.is_experimental_feature_activated(
"disable_graph_freezing")
self._need_calibration = ((
(conversion_params.precision_mode == TrtPrecisionMode.INT8) or
(conversion_params.precision_mode == TrtPrecisionMode.INT8.lower())) and
conversion_params.use_calibration)
self._calibration_input_fn = None
self._converted = False
self._device = None
self._build_called_once = False
self._calibrated = False
if use_dynamic_shape is None:
self._use_dynamic_shape = False
else:
self._use_dynamic_shape = use_dynamic_shape
if not self.freeze and not self._use_dynamic_shape:
logging.warn(
"Disabling graph freezing is only possible in dynamic shape mode."
" The graph will be frozen.")
self.freeze = True
self._profile_strategy = "Unknown"
if self._use_dynamic_shape:
if dynamic_shape_profile_strategy is None:
self._profile_strategy = PROFILE_STRATEGY_RANGE
else:
self._verify_profile_strategy(dynamic_shape_profile_strategy)
self._profile_strategy = dynamic_shape_profile_strategy
# Fields to support TF-TRT testing and shouldn't be used for other purpose.
self._test_only_disable_non_trt_optimizers = False
def _need_trt_profiles(self):
return self._use_dynamic_shape
def _run_conversion(self, meta_graph_def):
"""Run Grappler's OptimizeGraph() tool to convert the graph.
Args:
meta_graph_def: the MetaGraphDef instance to run the optimizations on.
Returns:
The optimized GraphDef.
"""
grappler_session_config = config_pb2.ConfigProto()
# Always set `allow_build_at_runtime` for offline TensorRT engine building.
custom_rewriter_config = _get_tensorrt_rewriter_config(
conversion_params=self._conversion_params._replace(
allow_build_at_runtime=True),
is_dynamic_op=True,
max_batch_size=None,
disable_non_trt_optimizers=self._test_only_disable_non_trt_optimizers,
use_implicit_batch=not self._use_dynamic_shape,
profile_strategy=self._profile_strategy)
grappler_session_config.graph_options.rewrite_options.CopyFrom(
custom_rewriter_config)
return tf_optimizer.OptimizeGraph(
grappler_session_config, meta_graph_def, graph_id=b"tf_graph")
def _for_each_trt_node(self, graph_def, fn):
"""Helper method to manipulate all TRTEngineOps in a GraphDef."""
for node in graph_def.node:
if node.op == _TRT_ENGINE_OP_NAME:
fn(node)
for func in graph_def.library.function:
for node in func.node_def:
if node.op == _TRT_ENGINE_OP_NAME:
fn(node)
def _execute_calibration(self, calibration_input_fn):
"""Run INT8 calibration with the provided input generator function."""
for inp in calibration_input_fn():
args, kwargs = _convert_to_tensor(inp)
self._converted_func(*args, **kwargs)
self._for_each_trt_node(self._converted_graph_def, _save_calibration_table)
# Rebuild the function since calibration has changed the graph.
self._converted_func = _construct_function_from_graph_def(
self._converted_func, self._converted_graph_def)
self._calibrated = True
# TODO(laigd): provide a utility function to optimize a ConcreteFunction and
# use it here (b/124792963).
def convert(self, calibration_input_fn=None):
"""Convert the input SavedModel in 2.0 format.
Args:
calibration_input_fn: a generator function that yields input data as a
list or tuple or dict, which will be used to execute the converted
signature for calibration. All the returned input data should have the
same shape. Example: `def input_fn(): yield input1, input2, input3`
If dynamic_shape_mode==False, (or if the graph has static input shapes)
then we run calibration and build the calibrated engine during
conversion.
If dynamic_shape_mode==True (and the graph has any unknown input
shape), then the reference to calibration_input_fn is stored, and the
calibration is actually performed when we build the engine (see
build()).
Raises:
ValueError: if the input combination is invalid.
Returns:
The TF-TRT converted Function.
"""
assert not self._converted
# Creating an empty tensor to fetch queried device
device_requested = array_ops.zeros([]).device
if "gpu" not in device_requested.lower():
raise ValueError(f"Specified device is not a GPU: {device_requested}")
if "gpu:0" not in device_requested.lower():
self._device = device_requested
logging.info(f"Placing imported graph from "
f"`{self._input_saved_model_dir}` on device: {self._device}")
if (self._need_calibration and not calibration_input_fn):
raise ValueError("Should specify calibration_input_fn because INT8 "
"calibration is needed")
if (not self._need_calibration and calibration_input_fn):
raise ValueError("Should not specify calibration_input_fn because INT8 "
"calibration is not needed")
self._saved_model = load.load(self._input_saved_model_dir,
self._input_saved_model_tags)
func = self._saved_model.signatures[self._input_saved_model_signature_key]
if self.freeze:
frozen_func = convert_to_constants.convert_variables_to_constants_v2(func)
else:
inlined_graph_def = _apply_inlining(func)
_annotate_variable_ops(func, inlined_graph_def)
frozen_func = _construct_function_from_graph_def(func, inlined_graph_def)
frozen_graph_def = frozen_func.graph.as_graph_def()
# Clear any prior device assignments
logging.info("Clearing prior device assignments in loaded saved model")
for node in frozen_graph_def.node:
node.device = ""
if self._device is None:
grappler_meta_graph_def = saver.export_meta_graph(
graph_def=frozen_graph_def, graph=frozen_func.graph)
else:
with ops.Graph().as_default() as graph, ops.device(self._device):
importer.import_graph_def(frozen_graph_def, name="")
grappler_meta_graph_def = saver.export_meta_graph(
graph_def=graph.as_graph_def(), graph=graph)
# Add a collection 'train_op' so that Grappler knows the outputs.
fetch_collection = meta_graph_pb2.CollectionDef()
for array in frozen_func.inputs + frozen_func.outputs:
fetch_collection.node_list.value.append(array.name)
grappler_meta_graph_def.collection_def["train_op"].CopyFrom(
fetch_collection)
# Run TRT optimizer in Grappler to convert the graph.
self._converted_graph_def = self._run_conversion(grappler_meta_graph_def)
self._converted_func = _construct_function_from_graph_def(
func, self._converted_graph_def, frozen_func)
if self._need_calibration:
# Execute calibration here only if not in dynamic shape mode.
if not self._need_trt_profiles():
self._execute_calibration(calibration_input_fn)
else:
self._calibration_input_fn = calibration_input_fn
self._converted = True
graphviz_path = os.environ.get("TF_TRT_EXPORT_GRAPH_VIZ_PATH", default=None)
if graphviz_path is not None:
try:
trt_utils.draw_graphdef_as_graphviz(
graphdef=self._converted_func.graph.as_graph_def(add_shapes=True),
dot_output_filename=graphviz_path)
except Exception as e:
logging.error(
"An Exception occurred during the export of the graph "
f"visualization: {e}"
)
return self._converted_func
def build(self, input_fn):
"""Run inference with converted graph in order to build TensorRT engines.
If the conversion requires INT8 calibration, then a reference to the
calibration function was stored during the call to convert(). Calibration
will be performed while we build the TensorRT engines.
Args:
input_fn: a generator function that provides the input data as a single
array, OR a list or tuple of the arrays OR a dict, which will be used
to execute the converted signature to generate TRT engines.
Example 1:
`def input_fn():
# Let's assume a network with 1 input tensor.
# We generate 2 sets of dummy input data:
input_shapes = [(1, 16), # 1st shape
(2, 32)] # 2nd shape
for shapes in input_shapes:
# return an input tensor
yield np.zeros(shape).astype(np.float32)'
Example 2:
`def input_fn():
# Let's assume a network with 2 input tensors.
# We generate 3 sets of dummy input data:
input_shapes = [[(1, 16), (2, 16)], # 1st input list
[(2, 32), (4, 32)], # 2nd list of two tensors
[(4, 32), (8, 32)]] # 3rd input list
for shapes in input_shapes:
# return a list of input tensors
yield [np.zeros(x).astype(np.float32) for x in shapes]`
Raises:
NotImplementedError: build() is already called.
RuntimeError: the input_fx is None.
"""
if self._build_called_once:
raise NotImplementedError("build() is already called. It is not "
"supported to call build() more than once.")
if not input_fn:
raise RuntimeError("input_fn is None. Method build() needs input_fn "
"to be specified in order to build TensorRT engines")
if not self._converted:
raise RuntimeError("Need to call convert() before build()")
if (self._need_calibration and not self._calibrated and
self._calibration_input_fn is None):
raise RuntimeError("Need to provide the calibration_input_fn arg while "
"calling convert().")
def _set_profile_generation_mode(value, node):
node.attr["_profile_generation_mode"].b = value
if self._need_trt_profiles():
# Enable profile generation.
self._for_each_trt_node(self._converted_graph_def,
partial(_set_profile_generation_mode, True))
# Profile generation is enabled using the _profile_generation_mode
# attribute of the TRTEngineOps. We need to rebuild the function to
# change this attribute.
func = _construct_function_from_graph_def(self._converted_func,
self._converted_graph_def)
else:
func = self._converted_func
first_input = None
# Run inference:
# Builds TRT engines if self._need_trt_profiles is False.
# Builds TRT optimization profiles if self._need_trt_profiles is True.
for inp in input_fn():
if first_input is None:
first_input = inp
args, kwargs = _convert_to_tensor(inp)
func(*args, **kwargs)
if self._need_trt_profiles():
# Disable profile generation.
self._for_each_trt_node(self._converted_graph_def,
partial(_set_profile_generation_mode, False))
# Run calibration if required, this would have been skipped in
# the convert step
if self._need_calibration and not self._calibrated:
self._execute_calibration(self._calibration_input_fn)
# calibration also builds the engine
else:
# Use the first input in explicit batch mode to build TensorRT engines
# after generating all the profiles. The first input is used but any of
# the inputs can be used because the shape of this input does not
# determine the engine and instead the shapes collected in profiles
# determine the engine.
args, kwargs = _convert_to_tensor(first_input)
self._converted_func(*args, **kwargs)
self._build_called_once = True
def save(self,
output_saved_model_dir,
save_gpu_specific_engines=True,
options=None):
"""Save the converted SavedModel.
Args:
output_saved_model_dir: directory to saved the converted SavedModel.
save_gpu_specific_engines: whether to save TRT engines that have been
built. When True, all engines are saved and when False, the engines
are not saved and will be rebuilt at inference time. By using
save_gpu_specific_engines=False after doing INT8 calibration, inference
can be done on different GPUs than the GPU that the model was calibrated
and saved on.
options: `tf.saved_model.SaveOptions` object for configuring save options.
Raises:
RuntimeError: if the needed calibration hasn't been done.
"""
assert self._converted
# 'remove_native_segments': setting this value to True removes native segments
# associated with each TRT engine. This option can be used to reduce the size
# of the converted model. Please note that a converted model without native
# segments can't be used for collecting profiles, building or re-converting.
# The reduced model can only be used for inference when no native segments
# are required for computation. When remove_native_segments flag is set to
# True, the converted_graph_def needs to be reduced before saved_model
# function serialization.
if trt_utils.is_experimental_feature_activated("remove_native_segments"):
logging.info(
"'remove_native_segments' experimental feature is enabled"
" during saving of converted SavedModel."
)
self._converted_func = _remove_native_segments(self._converted_func)
self._converted_graph_def = self._converted_func.graph.as_graph_def()
if self._need_calibration and not self._calibrated:
raise RuntimeError("A model that requires INT8 calibration has to be "
"built before saving it. Call build() to build and "
"calibrate the TensorRT engines.")
# Serialize the TRT engines in the cache if any, and create trackable
# resource to track them.
engine_asset_dir = tempfile.mkdtemp()
resource_map = {}
def _serialize_and_track_engine(node):
"""Serialize TRT engines in the cache and track them."""
# Don't dump the same cache twice.
canonical_engine_name = _get_canonical_engine_name(node.name)
if canonical_engine_name in resource_map:
return
filename = os.path.join(engine_asset_dir,
"trt-serialized-engine." + canonical_engine_name)
try:
gen_trt_ops.serialize_trt_resource(
resource_name=canonical_engine_name,
filename=filename,
delete_resource=True,
save_gpu_specific_engines=save_gpu_specific_engines)
except errors.NotFoundError:
logging.info(
"Could not find %s in TF-TRT cache. "
"This can happen if build() is not called, "
"which means TensorRT engines will be built "
"and cached at runtime.", canonical_engine_name)
return
# TODO(laigd): add an option for the user to choose the device.
resource_map[canonical_engine_name] = _TRTEngineResource(
canonical_engine_name, filename,
self._conversion_params.maximum_cached_engines)
self._for_each_trt_node(self._converted_graph_def,
_serialize_and_track_engine)
# If the graph is frozen, tracked variables are not needed by the converted model.
trackable = autotrackable.AutoTrackable(
) if self.freeze else self._saved_model
trackable.trt_engine_resources = resource_map
# Set allow_build_at_runtime=False if asked by user.
#
# This attribute is set here because build() needs it to be True in order to
# build engines.
if not self._conversion_params.allow_build_at_runtime:
def _reset_allow_build_at_runtime(node):
node.attr["_allow_build_at_runtime"].b = False
self._for_each_trt_node(self._converted_graph_def,
_reset_allow_build_at_runtime)
# Rebuild the function since a node attribute changed above
reset_converted_func = wrap_function.function_from_graph_def(
self._converted_graph_def,
[tensor.name for tensor in self._converted_func.inputs],
[tensor.name for tensor in self._converted_func.outputs])
reset_converted_func.graph.structured_outputs = nest.pack_sequence_as(
self._converted_func.graph.structured_outputs,
reset_converted_func.graph.structured_outputs)
reset_converted_func.graph.structured_input_signature = (
self._converted_func.structured_input_signature)
self._converted_func = reset_converted_func
# Rewrite the signature map using the optimized ConcreteFunction.
signatures = {self._input_saved_model_signature_key: self._converted_func}
save.save(trackable, output_saved_model_dir, signatures, options=options)
def summary(self, line_length=160, detailed=True, print_fn=None):
"""This method describes the results of the conversion by TF-TRT.
It includes information such as the name of the engine, the number of nodes
per engine, the input and output dtype, along with the input shape of each
TRTEngineOp.
Args:
line_length: Default line length when printing on the console. Minimum 160
characters long.
detailed: Whether or not to show the nodes inside each TRTEngineOp.
print_fn: Print function to use. Defaults to `print`. It will be called on
each line of the summary. You can set it to a custom function in order
to capture the string summary.
Raises:
RuntimeError: if the graph is not converted.
"""
if not self._converted:
raise RuntimeError(
f"Impossible to call `{self.__class__.__name__}.summary()` before "
f"calling {self.__class__.__name__}.convert()`.")
if line_length < 160:
raise ValueError(f"Invalid `line_length` value has been received: "
f"{line_length}. Minimum: 160.")
if print_fn is None:
print_fn = print
# positions are percentage of `line_length`. positions[i]+1 is the starting
# position for (i+1)th field. We also make sure that the last char printed
# for each field is a space.
columns = [
# (column name, column size in % of line)
("TRTEngineOP Name", .20), # 20%
("Device", .09), # 29%
("# Nodes", .05), # 34%
("# Inputs", .09), # 43%
("# Outputs", .09), # 52%
("Input DTypes", .12), # 64%
("Output Dtypes", .12), # 76%
("Input Shapes", .12), # 88%
("Output Shapes", .12) # 100%
]
positions = [int(line_length * p) for _, p in columns]
positions = np.cumsum(positions).tolist()
headers = [h for h, _ in columns]
_print_row(headers, positions, print_fn=print_fn)
print_fn("=" * line_length)
n_engines = 0
n_ops_converted = 0
n_ops_not_converted = 0
graphdef = self._converted_func.graph.as_graph_def(add_shapes=True)
trtengineops_dict = dict()
for node in graphdef.node:
if node.op != "TRTEngineOp":
n_ops_not_converted += 1
continue
else:
trtengineops_dict[node.name] = node
n_engines += 1
for name, node in sorted(trtengineops_dict.items()):
node_device = node.device.split("/")[-1]
in_shapes = trt_utils.get_node_io_shapes(node, "input_shapes")
out_shapes = trt_utils.get_node_io_shapes(node, "_output_shapes")
in_dtypes = trt_utils.get_trtengineop_io_dtypes(node, "InT")
out_dtypes = trt_utils.get_trtengineop_io_dtypes(node, "OutT")
in_nodes_count = trt_utils.get_trtengineop_io_nodes_count(node, "InT")
out_nodes_count = trt_utils.get_trtengineop_io_nodes_count(node, "OutT")
node_count, converted_ops_dict = trt_utils.get_trtengineop_node_op_count(
graphdef, name)
n_ops_converted += node_count
if n_engines != 1:
print_fn(f"\n{'-'*40}\n")
_print_row(
fields=[
name, node_device, node_count, in_nodes_count, out_nodes_count,
in_dtypes, out_dtypes, in_shapes, out_shapes
],
positions=positions,
print_fn=print_fn)
if detailed:
print_fn()
for key, value in sorted(dict(converted_ops_dict).items()):
print_fn(f"\t- {key}: {value}x")
print_fn(f"\n{'='*line_length}")
print_fn(f"[*] Total number of TensorRT engines: {n_engines}")
total_ops = n_ops_not_converted + n_ops_converted
conversion_ratio = n_ops_converted / total_ops * 100
print_fn(f"[*] % of OPs Converted: {conversion_ratio:.2f}% "
f"[{n_ops_converted}/{total_ops}]\n")
# TODO(laigd): use TrtConversionParams here.
def create_inference_graph(
input_graph_def,
outputs,
max_batch_size=1,
max_workspace_size_bytes=DEFAULT_TRT_MAX_WORKSPACE_SIZE_BYTES,
precision_mode=TrtPrecisionMode.FP32,
minimum_segment_size=3,
is_dynamic_op=False,
maximum_cached_engines=1,
input_saved_model_dir=None,
input_saved_model_tags=None,
input_saved_model_signature_key=None,
output_saved_model_dir=None):
"""Python wrapper for the TRT transformation.
Args:
input_graph_def: a GraphDef object containing a model to be transformed. If
set to None, the graph will be read from the SavedModel loaded from
input_saved_model_dir.
outputs: list of tensors or node names for the model outputs. Only used when
input_graph_def is not None.
max_batch_size: max size for the input batch.
max_workspace_size_bytes: the maximum GPU temporary memory which the TRT
engine can use at execution time. This corresponds to the 'workspaceSize'
parameter of nvinfer1::IBuilder::setMaxWorkspaceSize().
precision_mode: one of TrtPrecisionMode.supported_precision_modes().
minimum_segment_size: the minimum number of nodes required for a subgraph to
be replaced by TRTEngineOp.
is_dynamic_op: whether to generate dynamic TRT ops which will build the TRT
network and engine at run time.
maximum_cached_engines: max number of cached TRT engines in dynamic TRT ops.
If the number of cached engines is already at max but none of them can
serve the input, the TRTEngineOp will fall back to run the TF function
based on which the TRTEngineOp is created.
input_saved_model_dir: the directory to load the SavedModel which contains
the input graph to transforms. Used only when input_graph_def is None.
input_saved_model_tags: list of tags to load the SavedModel.
input_saved_model_signature_key: the key of the signature to optimize the
graph for.
output_saved_model_dir: if not None, construct a SavedModel using the
returned GraphDef and save it to the specified directory. This option only
works when the input graph is loaded from a SavedModel, i.e. when
input_saved_model_dir is specified and input_graph_def is None.
Returns:
A GraphDef transformed from input_graph_def (or the SavedModel graph def
loaded from input_saved_model_dir, if input_graph_def is not present), where
all TRT compatible subgraphs are replaced with TRTEngineOps, and a TF
function is added for each of the subgraphs.
If is_dynamic_op is True, each TRTEngineOp will contain a serialized
subgraph GraphDef, which will be converted to a TRT engine at execution time
and the TRT engine will be cached for future usage. A new TRT engine will be
created each time when none of the cached engines match the input shapes. If
it fails to execute the TRT engine or the number of cached engines reaches
maximum_cached_engines, the op will fall back to call the corresponding TF
function.
If is_dynamic_op is False, each TRTEngineOp will contain a serialized TRT
engine created from the corresponding subgraph. No more engines will be
created on the fly, and the op will fall back to call the corresponding TF
function when it fails to execute the engine.
Raises:
ValueError: if the combination of the parameters is invalid.
"""
trt_converter = TrtGraphConverter(
input_saved_model_dir=input_saved_model_dir,
input_saved_model_tags=input_saved_model_tags,
input_saved_model_signature_key=input_saved_model_signature_key,
input_graph_def=input_graph_def,
nodes_denylist=outputs,
max_batch_size=max_batch_size,
max_workspace_size_bytes=max_workspace_size_bytes,
precision_mode=precision_mode,
minimum_segment_size=minimum_segment_size,
is_dynamic_op=is_dynamic_op,
maximum_cached_engines=maximum_cached_engines,
use_calibration=False)
converted_graph_def = trt_converter.convert()
if output_saved_model_dir:
trt_converter.save(output_saved_model_dir)
return converted_graph_def