# mypy: allow-untyped-defs
"""Functions to export models into the ONNX IR format.
These models can be loaded with the ONNX library and then
converted to models which run on other deep learning frameworks.
"""
from __future__ import annotations
import contextlib
import copy
import inspect
import re
import typing
import warnings
from typing import Any, Callable, cast, Collection, Mapping, Sequence
import torch
import torch._C._onnx as _C_onnx
import torch.jit._trace
import torch.serialization
from torch import _C
from torch.onnx import ( # noqa: F401
_constants,
_deprecation,
_exporter_states,
errors,
symbolic_helper,
)
from torch.onnx._globals import GLOBALS
from torch.onnx._internal import diagnostics, jit_utils, onnx_proto_utils, registration
__all__ = [
"is_in_onnx_export",
"select_model_mode_for_export",
"disable_apex_o2_state_dict_hook",
"setup_onnx_logging",
"exporter_context",
"export",
"model_signature",
"warn_on_static_input_change",
"unpack_quantized_tensor",
"export_to_pretty_string",
"unconvertible_ops",
"register_custom_op_symbolic",
"unregister_custom_op_symbolic",
]
def is_in_onnx_export() -> bool:
"""Returns whether it is in the middle of ONNX export."""
return GLOBALS.in_onnx_export
# TODO(justinchuby): Remove dependency to this global variable from constant_fold.cpp
# Skip check due to cannot import IValue from torch._C
_params_dict = {} # type: ignore[var-annotated]
@contextlib.contextmanager
def select_model_mode_for_export(model, mode: _C_onnx.TrainingMode):
r"""A context manager to temporarily set the training mode of ``model``
to ``mode``, resetting it when we exit the with-block.
Args:
model: Same type and meaning as ``model`` arg to :func:`export`.
mode: Same type and meaning as ``training`` arg to :func:`export`.
"""
if not isinstance(mode, _C_onnx.TrainingMode):
raise TypeError(
f"'mode' should be a torch.onnx.TrainingMode enum, but got '{type(mode)}'."
)
originally_training: bool = False
if hasattr(model, "training"):
originally_training = model.training
# ONNX opset 12 has better support for training amenable models, with updated
# versions of the dropout and batch_norm operators
if mode == _C_onnx.TrainingMode.TRAINING or (
mode == _C_onnx.TrainingMode.PRESERVE and originally_training
):
GLOBALS.export_training = True
if GLOBALS.export_onnx_opset_version < 12:
warnings.warn(
"You are exporting the model in training mode with onnx opset "
f"version {GLOBALS.export_onnx_opset_version}. "
"Opset versions lower than opset 12 will not be able to export "
"nodes such as Dropout and BatchNorm correctly."
)
else:
GLOBALS.export_training = False
GLOBALS.training_mode = mode
if mode == _C_onnx.TrainingMode.TRAINING:
model.train(True)
elif mode == _C_onnx.TrainingMode.EVAL:
model.train(False)
# else mode == _C_onnx.TrainingMode.PRESERVE, do nothing
try:
yield
finally:
if hasattr(model, "training") and not mode == _C_onnx.TrainingMode.PRESERVE:
model.train(originally_training)
@contextlib.contextmanager
def disable_apex_o2_state_dict_hook(model: torch.nn.Module | torch.jit.ScriptFunction):
# Apex O2 hook state_dict to return fp16 weights as fp32.
# Exporter cannot identify them as same tensors.
# Since this hook is only used by optimizer, it is safe to
# remove this hook while exporting.
if not isinstance(model, torch.jit.ScriptFunction):
model_hooks = {} # type: ignore[var-annotated]
for module in model.modules():
for key, hook in module._state_dict_hooks.items():
if type(hook).__name__ == "O2StateDictHook":
if module not in model_hooks:
model_hooks[module] = {}
model_hooks[module][key] = hook
if module in model_hooks:
for key in model_hooks[module]:
module._state_dict_hooks.pop(key)
try:
yield
finally:
# Add the hooks back
for module, m_map in model_hooks.items():
for key, hook in m_map.items():
module._state_dict_hooks[key] = hook
else:
try:
yield
finally:
pass
@contextlib.contextmanager
def setup_onnx_logging(verbose: bool):
is_originally_enabled = torch.onnx.is_onnx_log_enabled()
if is_originally_enabled or verbose:
torch.onnx.enable_log()
try:
yield
finally:
if not is_originally_enabled:
torch.onnx.disable_log()
@contextlib.contextmanager
def exporter_context(model, mode: _C_onnx.TrainingMode, verbose: bool):
with select_model_mode_for_export(
model, mode
) as mode_ctx, disable_apex_o2_state_dict_hook(
model
) as apex_ctx, setup_onnx_logging(
verbose
) as log_ctx, diagnostics.create_export_diagnostic_context() as diagnostic_ctx:
yield (mode_ctx, apex_ctx, log_ctx, diagnostic_ctx)
def _get_torch_export_args(
args: tuple[Any, ...],
kwargs: dict[str, Any] | None,
) -> tuple[tuple[Any, ...], dict[str, Any] | None]:
"""Obtain the arguments for torch.onnx.export from the model and the input arguments."""
if not kwargs and args and isinstance(args[-1], dict):
kwargs = args[-1]
args = args[:-1]
return args, kwargs
def export(
model: torch.nn.Module | torch.jit.ScriptModule | torch.jit.ScriptFunction,
args: tuple[Any, ...] | torch.Tensor,
f: str,
*,
kwargs: dict[str, Any] | None = None,
export_params: bool = True,
verbose: bool = False,
training: _C_onnx.TrainingMode = _C_onnx.TrainingMode.EVAL,
input_names: Sequence[str] | None = None,
output_names: Sequence[str] | None = None,
operator_export_type: _C_onnx.OperatorExportTypes = _C_onnx.OperatorExportTypes.ONNX,
opset_version: int | None = None,
do_constant_folding: bool = True,
dynamic_axes: Mapping[str, Mapping[int, str]]
| Mapping[str, Sequence[int]]
| None = None,
keep_initializers_as_inputs: bool | None = None,
custom_opsets: Mapping[str, int] | None = None,
export_modules_as_functions: bool | Collection[type[torch.nn.Module]] = False,
autograd_inlining: bool = True,
) -> None:
r"""Exports a model into ONNX format.
If ``model`` is not a :class:`torch.jit.ScriptModule` nor a
:class:`torch.jit.ScriptFunction`, this runs
``model`` once in order to convert it to a TorchScript graph to be exported
(the equivalent of :func:`torch.jit.trace`). Thus this has the same limited support
for dynamic control flow as :func:`torch.jit.trace`.
Args:
model: The model to be exported.
args:
args can be structured either as:
1. ONLY A TUPLE OF ARGUMENTS::
args = (x, y, z)
The tuple should contain model inputs such that ``model(*args)`` is a valid
invocation of the model. Any non-Tensor arguments will be hard-coded into the
exported model; any Tensor arguments will become inputs of the exported model,
in the order they occur in the tuple.
2. A TENSOR::
args = torch.Tensor([1])
This is equivalent to a 1-ary tuple of that Tensor.
3. A TUPLE OF ARGUMENTS ENDING WITH A DICTIONARY OF NAMED ARGUMENTS::
args = (x, {"y": input_y, "z": input_z})
All but the last element of the tuple will be passed as non-keyword arguments,
and named arguments will be set from the last element. If a named argument is
not present in the dictionary, it is assigned the default value, or None if a
default value is not provided.
.. warning::
This behavior will be deprecated in a future release. Please use the
kwargs argument instead.
.. note::
If a dictionary is the last element of the args tuple, it will be
interpreted as containing named arguments. In order to pass a dict as the
last non-keyword arg, provide an empty dict as the last element of the args
tuple. For example, instead of::
torch.onnx.export(
model,
(
x,
# WRONG: will be interpreted as named arguments
{y: z},
),
"test.onnx.pb",
)
Write::
torch.onnx.export(model, (x, {y: z}, {}), "test.onnx.pb")
f: Path to the output ONNX model file. E.g. "model.onnx".
kwargs: Named arguments to the model.
export_params: If True, all parameters will
be exported. Set this to False if you want to export an untrained model.
In this case, the exported model will first take all of its parameters
as arguments, with the ordering as specified by ``model.state_dict().values()``
verbose: if True, prints a description of the
model being exported to stdout. In addition, the final ONNX graph will include the
field ``doc_string``` from the exported model which mentions the source code locations
for ``model``. If True, ONNX exporter logging will be turned on.
training:
* ``TrainingMode.EVAL``: export the model in inference mode.
* ``TrainingMode.PRESERVE``: export the model in inference mode if model.training is
False and in training mode if model.training is True.
* ``TrainingMode.TRAINING``: export the model in training mode. Disables optimizations
which might interfere with training.
input_names (list of str, default empty list): names to assign to the
input nodes of the graph, in order.
output_names (list of str, default empty list): names to assign to the
output nodes of the graph, in order.
operator_export_type (enum, default OperatorExportTypes.ONNX):
.. warning::
This option will be deprecated in a future release. Future exported
graphs will always use the default opset domain.
* ``OperatorExportTypes.ONNX``: Export all ops as regular ONNX ops
(in the default opset domain).
* ``OperatorExportTypes.ONNX_FALLTHROUGH``: Try to convert all ops
to standard ONNX ops in the default opset domain. If unable to do so
(e.g. because support has not been added to convert a particular torch op to ONNX),
fall back to exporting the op into a custom opset domain without conversion. Applies
to `custom ops <https://pytorch.org/tutorials/advanced/torch_script_custom_ops.html>`_
as well as ATen ops. For the exported model to be usable, the runtime must support
these non-standard ops.
* ``OperatorExportTypes.ONNX_ATEN``: All ATen ops (in the TorchScript namespace "aten")
are exported as ATen ops (in opset domain "org.pytorch.aten").
`ATen <https://pytorch.org/cppdocs/#aten>`_ is PyTorch's built-in tensor library, so
this instructs the runtime to use PyTorch's implementation of these ops.
.. warning::
Models exported this way are probably runnable only by Caffe2.
This may be useful if the numeric differences in implementations of operators are
causing large differences in behavior between PyTorch and Caffe2 (which is more
common on untrained models).
* ``OperatorExportTypes.ONNX_ATEN_FALLBACK``: Try to export each ATen op
(in the TorchScript namespace "aten") as a regular ONNX op. If we are unable to do so
(e.g. because support has not been added to convert a particular torch op to ONNX),
fall back to exporting an ATen op. See documentation on OperatorExportTypes.ONNX_ATEN for
context.
For example::
graph(%0 : Float):
%3 : int = prim::Constant[value=0]()
# conversion unsupported
%4 : Float = aten::triu(%0, %3)
# conversion supported
%5 : Float = aten::mul(%4, %0)
return (%5)
Assuming ``aten::triu`` is not supported in ONNX, this will be exported as::
graph(%0 : Float):
%1 : Long() = onnx::Constant[value={0}]()
# not converted
%2 : Float = aten::ATen[operator="triu"](%0, %1)
# converted
%3 : Float = onnx::Mul(%2, %0)
return (%3)
.. warning::
Models exported this way are probably runnable only by Caffe2.
opset_version (int, default 17): The version of the
`default (ai.onnx) opset <https://github.com/onnx/onnx/blob/master/docs/Operators.md>`_
to target. Must be >= 7 and <= 17.
do_constant_folding: Apply the constant-folding optimization.
Constant-folding will replace some of the ops that have all constant inputs
with pre-computed constant nodes.
dynamic_axes:
By default the exported model will have the shapes of all input and output tensors
set to exactly match those given in ``args``. To specify axes of tensors as
dynamic (i.e. known only at run-time), set ``dynamic_axes`` to a dict with schema:
* KEY (str): an input or output name. Each name must also be provided in ``input_names`` or
``output_names``.
* VALUE (dict or list): If a dict, keys are axis indices and values are axis names. If a
list, each element is an axis index.
For example::
class SumModule(torch.nn.Module):
def forward(self, x):
return torch.sum(x, dim=1)
torch.onnx.export(
SumModule(),
(torch.ones(2, 2),),
"onnx.pb",
input_names=["x"],
output_names=["sum"],
)
Produces::
input {
name: "x"
...
shape {
dim {
dim_value: 2 # axis 0
}
dim {
dim_value: 2 # axis 1
...
output {
name: "sum"
...
shape {
dim {
dim_value: 2 # axis 0
...
While::
torch.onnx.export(
SumModule(),
(torch.ones(2, 2),),
"onnx.pb",
input_names=["x"],
output_names=["sum"],
dynamic_axes={
# dict value: manually named axes
"x": {0: "my_custom_axis_name"},
# list value: automatic names
"sum": [0],
},
)
Produces::
input {
name: "x"
...
shape {
dim {
dim_param: "my_custom_axis_name" # axis 0
}
dim {
dim_value: 2 # axis 1
...
output {
name: "sum"
...
shape {
dim {
dim_param: "sum_dynamic_axes_1" # axis 0
...
keep_initializers_as_inputs: If True, all the
initializers (typically corresponding to parameters) in the
exported graph will also be added as inputs to the graph. If False,
then initializers are not added as inputs to the graph, and only
the non-parameter inputs are added as inputs.
This may allow for better optimizations (e.g. constant folding) by
backends/runtimes.
If True, `deduplicate_initializers` pass will not be executed. This means
initializers with duplicated values will not be deduplicated and
will be treated as distinct inputs to the graph. This allows different
input initializers to be supplied at the runtime following export.
If ``opset_version < 9``, initializers MUST be part of graph
inputs and this argument will be ignored and the behavior will be
equivalent to setting this argument to True.
custom_opsets (dict[str, int], default empty dict): A dict with schema:
* KEY (str): opset domain name
* VALUE (int): opset version
If a custom opset is referenced by ``model`` but not mentioned in this dictionary,
the opset version is set to 1. Only custom opset domain name and version should be
indicated through this argument.
export_modules_as_functions: Flag to enable
exporting all ``nn.Module`` forward calls as local functions in ONNX. Or a set to indicate the
particular types of modules to export as local functions in ONNX.
This feature requires ``opset_version`` >= 15, otherwise the export will fail. This is because
``opset_version`` < 15 implies IR version < 8, which means no local function support.
Module variables will be exported as function attributes. There are two categories of function
attributes.
1. Annotated attributes: class variables that have type annotations via
`PEP 526-style <https://www.python.org/dev/peps/pep-0526/#class-and-instance-variable-annotations>`_
will be exported as attributes.
Annotated attributes are not used inside the subgraph of ONNX local function because
they are not created by PyTorch JIT tracing, but they may be used by consumers
to determine whether or not to replace the function with a particular fused kernel.
2. Inferred attributes: variables that are used by operators inside the module. Attribute names
will have prefix "inferred::". This is to differentiate from predefined attributes retrieved from
python module annotations. Inferred attributes are used inside the subgraph of ONNX local function.
* ``False`` (default): export ``nn.Module`` forward calls as fine grained nodes.
* ``True``: export all ``nn.Module`` forward calls as local function nodes.
* Set of type of nn.Module: export ``nn.Module`` forward calls as local function nodes,
only if the type of the ``nn.Module`` is found in the set.
autograd_inlining: Flag used to control whether to inline autograd functions.
Refer to https://github.com/pytorch/pytorch/pull/74765 for more details.
Raises:
:class:`torch.onnx.errors.CheckerError`: If the ONNX checker detects an invalid ONNX graph.
:class:`torch.onnx.errors.UnsupportedOperatorError`: If the ONNX graph cannot be exported because it
uses an operator that is not supported by the exporter.
:class:`torch.onnx.errors.OnnxExporterError`: Other errors that can occur during export.
All errors are subclasses of :class:`errors.OnnxExporterError`.
"""
if operator_export_type != _C_onnx.OperatorExportTypes.ONNX:
warnings.warn(
"Setting `operator_export_type` to something other than default is deprecated. "
"The option will be removed in a future release.",
category=FutureWarning,
)
if training == _C_onnx.TrainingMode.TRAINING:
warnings.warn(
"Setting `training` to something other than default is deprecated. "
"The option will be removed in a future release. Please set the training mode "
"before exporting the model.",
category=FutureWarning,
)
args = (args,) if isinstance(args, torch.Tensor) else args
if kwargs is not None:
args = args + (kwargs,)
_export(
model,
args,
f,
export_params,
verbose,
training,
input_names,
output_names,
operator_export_type=operator_export_type,
opset_version=opset_version,
do_constant_folding=do_constant_folding,
dynamic_axes=dynamic_axes,
keep_initializers_as_inputs=keep_initializers_as_inputs,
custom_opsets=custom_opsets,
export_modules_as_functions=export_modules_as_functions,
autograd_inlining=autograd_inlining,
)
return None
def _is_constant_tensor_list(node):
if node.kind() != "prim::Constant":
return False
output_type = node.output().type()
if output_type.isSubtypeOf(_C.ListType.ofTensors()):
return True
if output_type.isSubtypeOf(_C.ListType(_C.OptionalType.ofTensor())):
return True
# ONNX can't handle constants that are lists of tensors, which can
# get generated in constant prop. So we split them back into prim::ListConstructs
def _split_tensor_list_constants(g, block):
for node in block.nodes():
for subblock in node.blocks():
_split_tensor_list_constants(g, subblock)
if _is_constant_tensor_list(node):
inputs = []
for val in node.output().toIValue():
input = g.insertConstant(val)
input.node().moveBefore(node)
input.node().copyMetadata(node)
inputs.append(input)
lc = (
g.create("prim::ListConstruct", inputs)
.insertBefore(node)
.output()
.setType(_C.ListType.ofTensors())
)
lc.node().copyMetadata(node)
node.output().replaceAllUsesWith(lc)
def _optimize_graph(
graph: _C.Graph,
operator_export_type: _C_onnx.OperatorExportTypes,
_disable_torch_constant_prop: bool = False,
fixed_batch_size: bool = False,
params_dict=None,
dynamic_axes=None,
input_names=None,
module=None,
):
if params_dict is None:
params_dict = {}
# Inline everything
_C._jit_pass_inline(graph)
# Remove fork/wait nodes
_C._jit_pass_inline_fork_wait(graph)
_C._jit_pass_lint(graph)
if GLOBALS.autograd_inlining:
_C._jit_pass_onnx_autograd_function_process(graph)
_C._jit_pass_lower_all_tuples(graph)
# we now record some ops like ones/zeros
# into a trace where we previously recorded constants.
# use constant prop to maintain our current level of onnx support
# without implementing symbolics for all of them
if _disable_torch_constant_prop is False:
_C._jit_pass_constant_propagation(graph)
_split_tensor_list_constants(graph, graph)
# run dce to eliminate dead parts of the graph that might have been
# left behind by things like symbolic_override
_C._jit_pass_dce(graph)
_C._jit_pass_lint(graph)
# CSE should improve perf when Autocast is used with disabled cache
# Autocast is disabled due to a limitation on tracer as described at https://github.com/pytorch/pytorch/issues/84092
# Must run before _C._jit_pass_erase_number_types to prevent type substitution
if _C._jit_pass_cse(graph):
_C._jit_pass_onnx_lint(graph)
_C._jit_pass_canonicalize_graph_fuser_ops(graph)
_C._jit_pass_lint(graph)
_C._jit_pass_peephole(graph, True)
_C._jit_pass_fuse_addmm(graph)
_C._jit_pass_lint(graph)
_C._jit_pass_peephole(graph, True)
_C._jit_pass_lower_all_tuples(graph)
# in _jit_pass_onnx, symbolic functions are called for each node for conversion.
# However, there are nodes that cannot be converted without additional context.
# For example, the number of outputs from split (and whether it is static or dynamic) is unknown
# until the point where it is unpacked by listUnpack node.
# This pass does a preprocess, and prepares the nodes such that enough context can be received
# by the symbolic function.
_C._jit_pass_onnx_remove_inplace_ops_for_onnx(graph, module)
_C._jit_pass_onnx_preprocess(graph)
# onnx does not support tuples, so try to remove them
_C._jit_pass_lint(graph)
# onnx only supports tensors, but 1 / 2 = 0.5 and tensor(1) / tensor(2) = 0
_C._jit_pass_prepare_division_for_onnx(graph)
_C._jit_pass_onnx_remove_print(graph)
_C._jit_pass_onnx_preprocess_caffe2(graph)
symbolic_helper._quantized_ops.clear()
# Unpack quantized weights for conv and linear ops and insert into graph.
_C._jit_pass_onnx_unpack_quantized_weights(graph, params_dict)
# onnx only supports tensors, so we turn all out number types into tensors
_C._jit_pass_erase_number_types(graph)
if GLOBALS.onnx_shape_inference:
input_names = [] if input_names is None else input_names
dynamic_axes = {} if dynamic_axes is None else dynamic_axes
_C._jit_pass_onnx_set_dynamic_input_shape(graph, dynamic_axes, input_names)
_C._jit_pass_onnx_lint(graph)
graph = _C._jit_pass_onnx(graph, operator_export_type)
_C._jit_pass_onnx_lint(graph)
_C._jit_pass_lint(graph)
_C._jit_pass_onnx_scalar_type_analysis(
graph, True, GLOBALS.export_onnx_opset_version
)
_C._jit_pass_lint(graph)
_C._jit_pass_onnx_peephole(
graph, GLOBALS.export_onnx_opset_version, fixed_batch_size
)
_C._jit_pass_lint(graph)
# graph is not a valid jit graph anymore because types have been replaced
# (e.g. int with Tensor), so it now contains operators that don't actually
# exist. We can't run normal dead code elimination because it'd fail trying
# to look up if an operator has side effects, but we can run a dead code
# elimination variant that doesn't need to look up if an op has side effects.
_C._jit_pass_dce_allow_deleting_nodes_with_side_effects(graph)
_C._jit_pass_lint(graph)
graph = _C._jit_pass_canonicalize(graph)
_C._jit_pass_lint(graph)
if GLOBALS.onnx_shape_inference:
_C._jit_pass_onnx_graph_shape_type_inference(
graph, params_dict, GLOBALS.export_onnx_opset_version
)
return graph
def warn_on_static_input_change(input_states):
"""Warns that changes to input dictionaries and strings won't take effect in the traced ONNX graph.
We accept dictionaries and strings as ONNX inputs, but they should be only for
configuration use. we detect here if these inputs are modified, and if so we warn
the user that the changes won't take effect in the traced ONNX graph.
"""
for input, traced_input in zip(input_states[0], input_states[1]):
if isinstance(input, dict):
if list(input.keys()) != list(traced_input.keys()):
warning = (
"We detected that you are modifying a dictionary that is an input to your "
"model. "
"Note that dictionaries are allowed as inputs in ONNX but they should be "
"handled with care. "
"Usages of dictionaries is not recommended, and should not be used except "
"for configuration use. "
"Also note that the order and values of the keys must remain the same. "
)
warnings.warn(warning)
elif isinstance(input, str):
if input != traced_input:
warning = (
"The model seems to have string inputs/outputs. "
"Note that strings will not appear as inputs/outputs of the ONNX graph. "
)
warnings.warn(warning)
def _resolve_args_by_export_type(arg_name, arg_value, operator_export_type):
"""Resolves the arguments that are ignored when export_type != operator_export_type.ONNX."""
return arg_value
def _decide_keep_init_as_input(
keep_initializers_as_inputs: bool | None,
operator_export_type: _C_onnx.OperatorExportTypes,
opset_version: int,
):
"""Decides whether the initializers in the graph should be listed as ONNX graph inputs.
This method encapsulates the logic to decide whether the initializers in the graph
should be listed as ONNX graph inputs (i.e., whether to choose ONNX IR v3 or v4).
If keep_initializers_as_inputs is not specified (None), then we decide whether to keep
initializers as graph inputs (val_keep_init_as_ip) based on export type. If export type
is ONNX, then do not keep initializers as input (val_keep_init_as_ip=False). For all other
export types keep initializers as input (val_keep_init_as_ip=True).
If keep_initializers_as_inputs is specified, then respect it. Unless opset version <= 8,
in which case it must be ignored because for opset version <= 8, all initializers MUST be
part of graph input (only ONNX IR v3 is allowed), i.e. val_keep_init_as_ip=True.
Special handling is needed for opset version 8 or lower, because irrespective
of user input for keep_initializers_as_inputs, the graph must follow ONNX IR v3
semantics, i.e. all initializers must be listed as ONNX graph input.
"""
if opset_version < 9:
if keep_initializers_as_inputs is False:
warnings.warn(
"Setting 'keep_initializers_as_inputs=False' for opset version"
"8 or lower would lead to an invalid ONNX graph. Therefore, "
"'keep_initializers_as_inputs=False' is ignored during export."
"Exported model will have initializers as graph inputs (compliant "
" to ONNX IR v3)."
)
return True # i.e. True == initializers are part of graph input (ONNX IR v3)
val_keep_init_as_ip = (
True if keep_initializers_as_inputs is None else keep_initializers_as_inputs
)
if (
keep_initializers_as_inputs is None
and operator_export_type is _C_onnx.OperatorExportTypes.ONNX
):
val_keep_init_as_ip = False
return val_keep_init_as_ip
def _decide_add_node_names(add_node_names, operator_export_type):
return _resolve_args_by_export_type(
"add_node_names", add_node_names, operator_export_type
)
def _decide_constant_folding(do_constant_folding, operator_export_type, training):
do_constant_folding = _resolve_args_by_export_type(
"do_constant_folding", do_constant_folding, operator_export_type
)
if do_constant_folding and (
training is not None and training is not _C_onnx.TrainingMode.EVAL
):
warnings.warn(
"It is recommended that constant folding be turned off ('do_constant_folding=False') "
"when exporting the model in training-amenable mode, i.e. with 'training=TrainingMode.TRAIN' "
"or 'training=TrainingMode.PRESERVE' (when model is in training mode). Otherwise, some "
"learnable model parameters may not translate correctly in the exported ONNX model "
"because constant folding mutates model parameters. Please consider "
"turning off constant folding or setting the training=TrainingMode.EVAL."
)
return do_constant_folding
def _signature(model) -> inspect.Signature:
should_be_callable = getattr(model, "forward", model)
if callable(should_be_callable):
return inspect.signature(should_be_callable)
raise ValueError("model has no forward method and is not callable")
def _decide_input_format(model, args):
try:
sig = _signature(model)
except ValueError as e:
warnings.warn(f"{e}, skipping _decide_input_format")
return args
try:
ordered_list_keys = list(sig.parameters.keys())
if ordered_list_keys[0] == "self":
ordered_list_keys = ordered_list_keys[1:]
args_dict: dict = {}
if isinstance(args, list):
args_list = args
elif isinstance(args, tuple):
args_list = list(args)
else:
args_list = [args]
if isinstance(args_list[-1], dict):
args_dict = args_list[-1]
args_list = args_list[:-1]
n_nonkeyword = len(args_list)
for optional_arg in ordered_list_keys[n_nonkeyword:]:
if optional_arg in args_dict:
args_list.append(args_dict[optional_arg])
# Check if this arg has a default value
else:
param = sig.parameters[optional_arg]
if param.default != param.empty:
args_list.append(param.default)
args = args_list if isinstance(args, list) else tuple(args_list)
# Cases of models with no input args
except IndexError:
warnings.warn("No input args, skipping _decide_input_format")
except Exception as e:
warnings.warn(f"Skipping _decide_input_format\n {e.args[0]}")
return args
def _from_dynamic_axes_to_dynamic_shapes(
model,
dynamic_axes: Mapping[str, Mapping[int, str]]
| Mapping[str, Sequence[int]]
| None = None,
input_names: Sequence[str] | None = None,
) -> dict[str, Any] | None:
"""
dynamic_axes examples:
(1) dynamic_axes = {"x": {0: "my_custom_axis_name_1"}, "y": {1: "my_custom_axis_name_2"}}
(2) dynamic_axes = {"x": [0], "y": [1]}
these will be converted to dynamic_shapes respectively:
(1) dynamic_shapes = {"x": {0: Dim("my_custom_axis_name_1")}, "y": {1: Dim("my_custom_axis_name_2")}}
(2) dynamic_shapes = {"x": {0: Dim("x_dim_0")}, "y": {1: Dim("y_dim_1")}} # auto-generated dim names
"""
if dynamic_axes is None:
return None
if input_names is None:
input_names_set = set()
else:
input_names_set = set(input_names)
dynamic_shapes: dict[str, Any | None] = {}
for input_name, axes in dynamic_axes.items():
if input_name in input_names_set:
raise ValueError(
"Assinging new input names is not supported yet. Please use model forward signature "
"to specify input names in dynamix_axes."
)
if isinstance(axes, dict):
dynamic_shapes[input_name] = {
k: torch.export.Dim(v) for k, v in axes.items()
}
elif isinstance(axes, list):
dynamic_shapes[input_name] = {
k: torch.export.Dim(f"{input_name}_dim_{k}") for k in axes
}
else:
raise TypeError(
f"dynamic_axes value must be either a dict or a list, but got {type(axes)}"
)
# torch.export.export needs static dim to present in dynamic_shapes
# for all input tensors, so we need to add them with None
try:
sig = _signature(model)
except ValueError as e:
warnings.warn(f"{e}, skipping auto filling None on static axes...")
return dynamic_shapes
for input_name in sig.parameters.keys():
if input_name not in dynamic_shapes:
dynamic_shapes[input_name] = None
return dynamic_shapes
def _trace(func, args, operator_export_type, return_outs=False):
# Special case for common case of passing a single Tensor
if isinstance(args, torch.Tensor):
args = (args,)
trace_graph, torch_out, inputs_states = torch.jit._get_trace_graph(
func,
args,
strict=False,
_force_outplace=False,
_return_inputs_states=True,
)
warn_on_static_input_change(inputs_states)
trace_graph = _optimize_graph(trace_graph, operator_export_type, params_dict={})
if return_outs:
return trace_graph, torch_out
return trace_graph
def _trace_and_get_graph_from_model(model, args):
# A basic sanity check: make sure the state_dict keys are the same
# before and after running the model. Fail fast!
orig_state_dict_keys = torch.jit._unique_state_dict(model).keys()
# Disable Autocast cache because it replaces kernel's weight and bias
# by (undesired) constants.
# No perf impact for when there are reused weights since https://github.com/pytorch/pytorch/pull/85665
prev_autocast_cache_enabled = torch.is_autocast_cache_enabled()
torch.set_autocast_cache_enabled(False)
trace_graph, torch_out, inputs_states = torch.jit._get_trace_graph(
model,
args,
strict=False,
_force_outplace=False,
_return_inputs_states=True,
)
torch.set_autocast_cache_enabled(prev_autocast_cache_enabled)
warn_on_static_input_change(inputs_states)
if orig_state_dict_keys != torch.jit._unique_state_dict(model).keys():
raise RuntimeError(
"state_dict changed after running the tracer; "
"something weird is happening in your model!"
)
return trace_graph, torch_out
def _get_param_count_list(method_graph, args_params):
param_count_list = []
for input_, arg_params_ in zip(method_graph.inputs(), args_params):
if "PackedParams" in str(input_.type()):
in_vars, _ = torch.jit._flatten(arg_params_)
param_count_list.append(len(in_vars))
else:
param_count_list.append(arg_params_ is not None)
return param_count_list
def _check_flatten_did_not_remove(original, jit_flattened):
"""torch.jit._flatten removes None. Check if it did so in this case."""
def flatten(x):
if isinstance(x, (list, tuple)):
for inner in x:
yield from flatten(inner)
elif isinstance(x, dict):
for inner in x.values():
yield from flatten(inner)
else:
yield x
flattened_with_none = list(flatten(original))
num_none = len(flattened_with_none) - len(jit_flattened)
assert num_none >= 0
if num_none:
raise ValueError(
f"args contained {num_none} None's after flattening. "
"When exporting a ScriptModule or ScriptFunction, no args may "
"be None because that breaks type propagation."
)
def _create_jit_graph(
model: torch.nn.Module | torch.jit.ScriptFunction, args: Sequence[Any]
) -> tuple[_C.Graph, list[_C.IValue], Any | None, _C.ScriptModule | None]:
if isinstance(model, (torch.jit.ScriptFunction, torch.jit.ScriptModule)):
flattened_args = tuple(torch.jit._flatten(tuple(args))[0])
_check_flatten_did_not_remove(args, flattened_args)
torch_out = None
if isinstance(model, torch.jit.ScriptModule):
try:
graph = model.forward.graph # type: ignore[attr-defined]
except AttributeError as e:
raise RuntimeError("'forward' method must be a script method") from e
_C._jit_pass_onnx_function_substitution(graph)
freezed_module = _C._freeze_module(
cast(_C.ScriptModule, model._c), preserveParameters=True
)
module, params = _C._jit_onnx_list_model_parameters(freezed_module)
method_graph = module._get_method("forward").graph
args_params = tuple(args) + tuple(params)
param_count_list = _get_param_count_list(method_graph, args_params)
in_vars, _ = torch.jit._flatten(args_params)
graph = _C._propagate_and_assign_input_shapes(
method_graph, tuple(in_vars), param_count_list, False, False
)
return graph, params, torch_out, module
# torch.jit.ScriptFunction
params = []
graph = model.graph
_C._jit_pass_onnx_function_substitution(graph)
param_count_list = _get_param_count_list(graph, args)
graph = _C._propagate_and_assign_input_shapes(
graph, flattened_args, param_count_list, False, False
)
return graph, params, torch_out, None
graph, torch_out = _trace_and_get_graph_from_model(model, args)
_C._jit_pass_onnx_lint(graph)
state_dict = torch.jit._unique_state_dict(model)
params = list(state_dict.values())
graph_inputs = list(graph.inputs())
user_input_num = len(graph_inputs) - len(state_dict)
param_names = list(state_dict.keys())
for i, inp in enumerate(graph_inputs):
if i >= user_input_num:
inp.setDebugName(param_names[i - user_input_num])
_C._jit_pass_onnx_function_substitution(graph)
return graph, params, torch_out, None
def _get_named_param_dict(graph, params):
input_and_param_names = [val.debugName() for val in graph.inputs()]
param_names = input_and_param_names[len(input_and_param_names) - len(params) :]
_params_dict = dict(zip(param_names, params))
return _params_dict
def _get_example_outputs(model, args):
input_args = copy.deepcopy(args)
input_kwargs = {}
if input_args and isinstance(input_args[-1], dict):
input_kwargs = input_args[-1]
input_args = input_args[:-1]
example_outputs = model(*input_args, **input_kwargs)
if isinstance(example_outputs, list):
example_outputs = [example_outputs]
elif not isinstance(example_outputs, tuple):
example_outputs = (example_outputs,)
return example_outputs
_qtype_vtype_map = {
torch.quint8: torch.uint8,
torch.qint8: torch.int8,
torch.qint32: torch.int32,
torch.quint4x2: torch.int8,
}
def unpack_quantized_tensor(value, cast_onnx_accepted=True):
if isinstance(value, torch.Tensor) and value.dtype in _qtype_vtype_map:
q_value_dequantize = value.dequantize()
q_scale = (
torch.tensor(value.q_scale(), dtype=torch.double)
if cast_onnx_accepted
else torch.tensor(value.q_scale(), dtype=torch.float32)
)
q_zero_point = (
torch.tensor(value.q_zero_point(), dtype=torch.int64)
if cast_onnx_accepted
else torch.tensor(value.q_zero_point(), dtype=_qtype_vtype_map[value.dtype])
)
q_value = q_value_dequantize / q_scale + q_zero_point
q_value = q_value.to(dtype=_qtype_vtype_map[value.dtype])
return q_value, q_scale, q_zero_point
else:
return (value,)
def _pre_trace_quant_model(model, args):
r"""Returns `torch.jit.trace(model, args)` if model is quantized. Otherwise do nothing and return
original model.
This is due to https://github.com/pytorch/pytorch/issues/75761.
"""
if any(
hasattr(m, "_packed_params") for m in getattr(model, "modules", list)()
) or any(getattr(arg, "is_quantized", False) for arg in args):
return torch.jit.trace(model, args)
return model
def _model_to_graph(
model,
args,
verbose=False,
input_names=None,
output_names=None,
operator_export_type=_C_onnx.OperatorExportTypes.ONNX,
do_constant_folding=True,
_disable_torch_constant_prop=False,
fixed_batch_size=False,
training=_C_onnx.TrainingMode.EVAL,
dynamic_axes=None,
) -> tuple[
_C.Graph,
dict[str, torch.Tensor],
torch.Tensor
| tuple[torch.Tensor, ...]
| list[torch.Tensor]
| dict[str, torch.Tensor]
| Any
| None,
]:
"""Converts model into an ONNX graph.
Returns:
graph: A TorchScript IR Graph with ONNX nodes.
params_dict: Dict from input param name to param value.
torch_out: The output tensors resulting from the trace of ``model``.
If ``model`` is a :class:`torch.jit.ScriptModule` or :class:`torch.jit.ScriptFunction`,
this will be None, since we are not doing any tracing.
"""
# TODO: can we simplify this to always return a tuple of Tensor or None?
# Special case for common case of passing a single Tensor
if isinstance(args, (torch.Tensor, int, float, bool)):
args = (args,)
model = _pre_trace_quant_model(model, args)
graph, params, torch_out, module = _create_jit_graph(model, args)
params_dict = _get_named_param_dict(graph, params)
try:
graph = _optimize_graph(
graph,
operator_export_type,
_disable_torch_constant_prop=_disable_torch_constant_prop,
fixed_batch_size=fixed_batch_size,
params_dict=params_dict,
dynamic_axes=dynamic_axes,
input_names=input_names,
module=module,
)
except Exception as e:
torch.onnx.log("Torch IR graph at exception: ", graph)
raise
is_script = isinstance(model, (torch.jit.ScriptFunction, torch.jit.ScriptModule))
if is_script:
example_outputs = _get_example_outputs(model, args)
example_outputs_final = ()
for example_output in example_outputs:
example_outputs_final += unpack_quantized_tensor(example_output)
out_vars, desc = torch.jit._flatten(example_outputs_final)
_C._jit_pass_onnx_assign_output_shape(
graph,
out_vars,
desc,
GLOBALS.onnx_shape_inference,
is_script,
GLOBALS.export_onnx_opset_version,
)
# NB: ONNX requires complete information about output types, which might be
# erased by some optimizations, so we need to set it explicitly again.
else:
if not isinstance(torch_out, (list, tuple)):
output_wrapped = [torch_out]
else:
output_wrapped = torch_out # type: ignore[assignment]
output_tensors, out_desc = torch.jit._flatten(tuple(output_wrapped))
# assign_output_shape pass is not compatible with quantized outputs.
# Quantized outputs are flattened to 3 values in ONNX, while packed as
# single value in PyTorch.
if not any(getattr(out, "is_quantized", False) for out in output_tensors):
_C._jit_pass_onnx_assign_output_shape(
graph,
output_tensors,
out_desc,
GLOBALS.onnx_shape_inference,
is_script,
GLOBALS.export_onnx_opset_version,
)
_set_input_and_output_names(graph, input_names, output_names)
params_dict = _get_named_param_dict(graph, params)
if (
do_constant_folding
and GLOBALS.export_onnx_opset_version
>= _constants.ONNX_CONSTANT_FOLDING_MIN_OPSET
):
if training is None or training == _C_onnx.TrainingMode.EVAL:
params_dict = _C._jit_pass_onnx_eval_peephole(graph, params_dict)
params_dict = _C._jit_pass_onnx_constant_fold(
graph, params_dict, GLOBALS.export_onnx_opset_version
)
_C._jit_pass_dce_allow_deleting_nodes_with_side_effects(graph)
if GLOBALS.onnx_shape_inference:
_C._jit_pass_onnx_graph_shape_type_inference(
graph, params_dict, GLOBALS.export_onnx_opset_version
)
params_dict = _C._jit_pass_onnx_eliminate_unused_items(graph, params_dict)
# For ONNX opset < 9, constants only have three data types: float16, float, double.
# In this pass transform constants of other data types to float/double + cast operator.
if GLOBALS.export_onnx_opset_version < 9:
_C._jit_pass_onnx_cast_all_constant_to_floating(graph)
params_dict = _C._jit_pass_filter_non_tensor_arguments(params_dict)
_C._jit_decay_packed_param_input_types(graph)
# If output names lack a proper name and are identified only by their unique
# give them a legible name for debugging purposes
_apply_friendly_debug_names(graph, params_dict)
return graph, params_dict, torch_out
@torch._disable_dynamo
@_deprecation.deprecated("2.5", "the future", "use onnx.printer.to_text() instead")
def export_to_pretty_string(
model,
args,
export_params=True,
verbose=False,
training=_C_onnx.TrainingMode.EVAL,
input_names=None,
output_names=None,
operator_export_type=_C_onnx.OperatorExportTypes.ONNX,
export_type=None,
google_printer=False,
opset_version=None,
keep_initializers_as_inputs=None,
custom_opsets=None,
add_node_names=True,
do_constant_folding=True,
dynamic_axes=None,
):
"""Similar to :func:`export`, but returns a text representation of the ONNX model.
Only differences in args listed below. All other args are the same
as :func:`export`.
Args:
add_node_names (bool, default True): Whether or not to set
NodeProto.name. This makes no difference unless
``google_printer=True``.
google_printer (bool, default False): If False, will return a custom,
compact representation of the model. If True will return the
protobuf's `Message::DebugString()`, which is more verbose.
Returns:
A UTF-8 str containing a human-readable representation of the ONNX model.
"""
if opset_version is None:
opset_version = _constants.ONNX_DEFAULT_OPSET
if custom_opsets is None:
custom_opsets = {}
GLOBALS.export_onnx_opset_version = opset_version
GLOBALS.operator_export_type = operator_export_type
with exporter_context(model, training, verbose):
val_keep_init_as_ip = _decide_keep_init_as_input(
keep_initializers_as_inputs, operator_export_type, opset_version
)
val_add_node_names = _decide_add_node_names(
add_node_names, operator_export_type
)
val_do_constant_folding = _decide_constant_folding(
do_constant_folding, operator_export_type, training
)
args = _decide_input_format(model, args)
graph, params_dict, torch_out = _model_to_graph(
model,
args,
verbose,
input_names,
output_names,
operator_export_type,
val_do_constant_folding,
training=training,
dynamic_axes=dynamic_axes,
)
return graph._pretty_print_onnx( # type: ignore[attr-defined]
params_dict,
opset_version,
False,
operator_export_type,
google_printer,
val_keep_init_as_ip,
custom_opsets,
val_add_node_names,
)
@_deprecation.deprecated("2.5", "the future", "avoid using this function")
def unconvertible_ops(
model,
args,
training: _C_onnx.TrainingMode = _C_onnx.TrainingMode.EVAL,
opset_version: int | None = None,
) -> tuple[_C.Graph, list[str]]:
"""Returns an approximated list of all ops that are yet supported by :mod:`torch.onnx`.
The list is approximated because some ops may be removed during the conversion
process and don't need to be converted. Some other ops may have partial support
that will fail conversion with particular inputs. Please open a Github Issue
for op support requests.
Args:
model: Same as the `model` parameter in :func:`torch.onnx.export`.
args: Same as the `args` parameter in :func:`torch.onnx.export`.
training: Same as the `training` parameter in :func:`torch.onnx.export`.
opset_version: Same as the `opset_version` parameter in :func:`torch.onnx.export`.
Returns:
The JIT graph and a list of unconvertible ops in the format of "domain::op".
"""
opset_version = opset_version or _constants.ONNX_DEFAULT_OPSET
GLOBALS.export_onnx_opset_version = opset_version
try:
with exporter_context(model, training, verbose=False):
# Create a mostly clean JIT graph that contains the plain aten and
# other ops we can check with the symbolic registry.
# NOTE: We don't want to actually convert any ops to ONNX or run any
# symbolic functions because there is a higher chance that a pass
# fails or an unconvertible op messes up the graph during ONNX conversion.
# This way we can always generate a list just by looking at the names
# of the ops in the graph.
args = _decide_input_format(model, args)
model = _pre_trace_quant_model(model, args)
graph, _, _, module = _create_jit_graph(model, args)
_C._jit_pass_inline(graph)
_C._jit_pass_onnx_remove_inplace_ops_for_onnx(graph, module)
_C._jit_pass_erase_number_types(graph)
_C._jit_pass_dce_allow_deleting_nodes_with_side_effects(graph)
except Exception as e:
raise errors.OnnxExporterError(
"Failed to discover unconvertible ops because of errors during the JIT graph "
"generation process."
) from e
unsupported_ops = []
for node in graph.nodes():
domain_op = node.kind()
if domain_op.startswith(("onnx::", "prim::")):
# We consider onnx and prim ops as supported ops, even though some "prim"
# ops are not implemented as symbolic functions, because they may be
# eliminated in the conversion passes. Users may still see errors caused
# by prim ops even though they don't show up in the list.
continue
if not registration.registry.is_registered_op(
domain_op.rstrip("_"), opset_version
):
# We consider all registered ops supported, even though some of them are
# only partially supported, because there is not yet a good way to check
# if an op is fully supported.
# TODO(justinchuby): Create a way to check if an op is fully supported.
unsupported_ops.append(domain_op)
return graph, unsupported_ops
def _setup_trace_module_map(
model: torch.nn.Module | torch.jit.ScriptModule,
export_modules_as_functions: bool | Collection[type[torch.nn.Module]],
) -> set[str]:
def __register_attribute_hook():
attr_name = "_onnx_attrs"
def _track_module_attributes_forward_pre_hook(module, input):
setattr(module, attr_name, _get_module_attributes(module))
def _track_module_attributes_forward_hook(module, input, output):
tracing_state = _C._get_tracing_state()
if not tracing_state:
return
graph = tracing_state.graph()
onnx_attrs = {}
if hasattr(module, attr_name):
onnx_attrs = getattr(module, attr_name)
delattr(module, attr_name)
_C._jit_pass_onnx_track_scope_attributes(graph, onnx_attrs)
for m in model.modules():
m.register_forward_hook(_track_module_attributes_forward_hook)
m.register_forward_pre_hook(_track_module_attributes_forward_pre_hook)
def _unqualified_variable_name(qualified_name: str) -> str:
"""
Parse qualified variable name and return the unqualified version.
Pure numeric atoms are considered inadequate, so this function will look past them,
and start from the first non-numeric atom.
Example:
>>> _unqualified_variable_name("__main__.Foo.bar")
'bar'
>>> _unqualified_variable_name("__main__.Foo.bar.0")
'bar.0'
"""
name_atoms = qualified_name.split(".")
for i, atom in reversed(list(enumerate(name_atoms))):
if not atom.isnumeric():
return ".".join(name_atoms[i:])
return qualified_name
trace_module_map = {
_m: torch._C._jit_onnx_create_full_scope_name(
torch.typename(type(_m)), _unqualified_variable_name(_n)
)
for _n, _m in model.named_modules()
}
torch.jit._trace._trace_module_map = trace_module_map
if isinstance(export_modules_as_functions, bool) and export_modules_as_functions:
module_typenames = {torch.typename(type(module)) for module in trace_module_map}
elif isinstance(export_modules_as_functions, set) and export_modules_as_functions:
def _find_typename(v):
if isinstance(v, type):
return torch.typename(v)
else:
raise RuntimeError(
"Only type of the `nn.Module` should be "
"passed in the set for argument `export_modules_as_functions`. "
f"Got `{type(v).__name__}`."
)
module_typenames = {_find_typename(v) for v in export_modules_as_functions}
else:
module_typenames = set()
if module_typenames:
__register_attribute_hook()
return module_typenames
def _reset_trace_module_map():
torch.jit._trace._trace_module_map = None
_C._jit_pass_onnx_clear_scope_records()
def _get_module_attributes(module):
annotations = typing.get_type_hints(type(module))
base_m_annotations = typing.get_type_hints(torch.nn.Module)
[annotations.pop(k, None) for k in base_m_annotations]
# Check whether module attributes can be accessed. Some classes
# define attributes but don't provide access to them in their
# constructor.
#
# For example, torch.nn.Embedding has the `freeze` variable and its
# type specified in the class but the attribute is not created in the
# constructor. In other words, there is no `self.freeze = <True | False>`
# in the constructor.
#
# Reference: https://github.com/pytorch/pytorch/blob/92de1d322223fb5584e384971b32c46b93bc2f4b/torch/nn/modules/sparse.py#L120
attrs = {}
for k in annotations:
try:
attrs[k] = getattr(module, k)
except AttributeError:
torch.onnx.log(f"Skipping module attribute '{k}'")
continue
return attrs
def _export(
model,
args,
f,
export_params=True,
verbose=False,
training=_C_onnx.TrainingMode.EVAL,
input_names=None,
output_names=None,
operator_export_type=_C_onnx.OperatorExportTypes.ONNX,
export_type=None,
opset_version=None,
do_constant_folding=True,
dynamic_axes=None,
keep_initializers_as_inputs=None,
fixed_batch_size=False,
custom_opsets=None,
add_node_names=True,
onnx_shape_inference=True,
export_modules_as_functions: Any = False,
autograd_inlining=True,
):
assert GLOBALS.in_onnx_export is False
if export_type is None:
export_type = _exporter_states.ExportTypes.PROTOBUF_FILE
if isinstance(model, torch.nn.DataParallel):
raise ValueError(
"torch.nn.DataParallel is not supported by ONNX "
"exporter, please use 'attribute' module to "
"unwrap model from torch.nn.DataParallel. Try "
"torch.onnx.export(model.module, ...)"
)
GLOBALS.onnx_shape_inference = onnx_shape_inference
if opset_version is None:
opset_version = _constants.ONNX_DEFAULT_OPSET
# torch.onnx.export does not support opset versions >=18
if opset_version > _constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET:
# We do not want to fail because we should still allow users to create
# custom symbolic functions for opset>17
warnings.warn(
f"Exporting to ONNX opset version {opset_version} is not supported. "
f"by 'torch.onnx.export()'. "
f"The highest opset version supported is {_constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET}. "
f"To use a newer opset version, consider 'torch.onnx.export(..., dynamo=True)'. ",
category=errors.OnnxExporterWarning,
)
if export_modules_as_functions and opset_version < 15:
raise ValueError(
"`export_modules_as_functions` is not supported for `opset_version` < 15."
"This is because `opset_version` < 15 implies IR version < 8, which means "
"no local function support. "
)
if not operator_export_type:
operator_export_type = _C_onnx.OperatorExportTypes.ONNX
# By default, training=TrainingMode.EVAL,
# which is good because running a model in training mode could result in
# internal buffers getting updated, dropout getting applied, etc.
# If you really know what you're doing, you can turn
# training=TrainingMode.TRAINING or training=TrainingMode.PRESERVE,
# (to preserve whatever the original training mode was.)
GLOBALS.export_onnx_opset_version = opset_version
GLOBALS.operator_export_type = operator_export_type
try:
GLOBALS.in_onnx_export = True
_autograd_inlining_previous = GLOBALS.autograd_inlining
GLOBALS.autograd_inlining = autograd_inlining
module_typenames_to_export_as_functions: set[str] = set()
if isinstance(model, (torch.nn.Module, torch.jit.ScriptModule)):
module_typenames_to_export_as_functions = _setup_trace_module_map(
model, export_modules_as_functions
)
with exporter_context(model, training, verbose):
val_keep_init_as_ip = _decide_keep_init_as_input(
keep_initializers_as_inputs,
operator_export_type,
opset_version,
)
val_add_node_names = _decide_add_node_names(
add_node_names, operator_export_type
)
val_do_constant_folding = _decide_constant_folding(
do_constant_folding, operator_export_type, training
)
# Normally f can be a file-like object, but for large models, the external data format requires a
# valid `model_file_location`. Code in export.cpp will enforce this.
if isinstance(f, str):
model_file_location = f
else:
model_file_location = ""
args = _decide_input_format(model, args)
if dynamic_axes is None:
dynamic_axes = {}
_validate_dynamic_axes(dynamic_axes, model, input_names, output_names)
graph, params_dict, torch_out = _model_to_graph(
model,
args,
verbose,
input_names,
output_names,
operator_export_type,
val_do_constant_folding,
fixed_batch_size=fixed_batch_size,
training=training,
dynamic_axes=dynamic_axes,
)
# TODO: Don't allocate a in-memory string for the protobuf
defer_weight_export = (
export_type is not _exporter_states.ExportTypes.PROTOBUF_FILE
)
if custom_opsets is None:
custom_opsets = {}
_C._jit_pass_dce_allow_deleting_nodes_with_side_effects(graph)
node_attr_to_name = {} # type: ignore[var-annotated]
if module_typenames_to_export_as_functions:
# NOTE: cannot call DCE after this pass. DCE will remove function definition nodes.
node_attr_to_name = _C._jit_pass_onnx_function_extraction(
graph,
module_typenames_to_export_as_functions,
list(params_dict.keys()),
)
if keep_initializers_as_inputs is not True:
params_dict = _C._jit_pass_onnx_deduplicate_initializers( # type: ignore[assignment]
graph,
params_dict, # type: ignore[arg-type]
getattr(model, "training", False), # type: ignore[arg-type]
)
_C._jit_pass_onnx_assign_scoped_names_for_node_and_value(graph)
if export_params:
(
proto,
export_map,
val_use_external_data_format,
node_names,
) = graph._export_onnx( # type: ignore[attr-defined]
params_dict,
opset_version,
dynamic_axes,
defer_weight_export,
operator_export_type,
not verbose,
val_keep_init_as_ip,
custom_opsets,
val_add_node_names,
model_file_location,
node_attr_to_name,
)
else:
(
proto,
export_map,
val_use_external_data_format,
node_names,
) = graph._export_onnx( # type: ignore[attr-defined]
{},
opset_version,
dynamic_axes,
False,
operator_export_type,
not verbose,
val_keep_init_as_ip,
custom_opsets,
val_add_node_names,
model_file_location,
node_attr_to_name,
)
# insert function_proto into model_proto.
proto = onnx_proto_utils._add_onnxscript_fn(
proto,
custom_opsets,
)
if verbose:
torch.onnx.log("Exported graph: ", graph)
onnx_proto_utils._export_file(proto, f, export_type, export_map)
finally:
assert GLOBALS.in_onnx_export
GLOBALS.in_onnx_export = False
GLOBALS.autograd_inlining = _autograd_inlining_previous
_reset_trace_module_map()
return torch_out
def _apply_friendly_debug_names(graph, params):
for n in graph.nodes():
for v in n.inputs():
old_name = v.debugName()
if old_name != str(v.unique()):
continue
new_name = f"{n.kind()}_{v.unique()}"
v.setDebugName(new_name)
if old_name in params:
params[new_name] = params.pop(old_name)
def _set_input_and_output_names(graph, input_names, output_names):
def set_names(node_list, name_list, descriptor):
if name_list is None:
return
if len(name_list) > len(node_list):
raise RuntimeError(
"number of %s names provided (%d) exceeded number of %ss (%d)"
% (descriptor, len(name_list), descriptor, len(node_list))
)
# Mark if the output node DebugName is set before.
output_node_set = set()
for i, (name, node) in enumerate(zip(name_list, node_list)):
# Duplicated output node, insert onnx::Identity to avoid setting the same DebugName after setDebugName().
if descriptor == "output":
if node in output_node_set:
identity_node = graph.create("onnx::Identity")
identity_node.insertAfter(node.node())
identity_node.addInput(node)
identity_node.output().setType(node.type())
graph.return_node().replaceInput(i, identity_node.output())
node = identity_node.output()
output_node_set.add(node)
if node.debugName() != name:
node.setDebugName(name)
set_names(list(graph.inputs()), input_names, "input")
set_names(list(graph.outputs()), output_names, "output")
def _run_symbolic_method(g, op_name, symbolic_fn, args):
r"""
This trampoline function gets invoked for every symbolic method
call from C++.
"""
try:
graph_context = jit_utils.GraphContext(
graph=g,
block=g.block(),
opset=GLOBALS.export_onnx_opset_version,
original_node=None, # type: ignore[arg-type]
params_dict=_params_dict,
env={},
values_in_env=set(),
new_nodes=[],
)
return symbolic_fn(graph_context, *args)
except TypeError as e:
# Handle the specific case where we didn't successfully dispatch
# to symbolic_fn. Otherwise, the backtrace will have the clues
# you need.
e.args = (f"{e.args[0]} (occurred when translating {op_name})",)
raise
def _add_block(node: _C.Node) -> _C.Block:
return node.addBlock()
def _add_input_to_block(block: _C.Block):
return block.addInputToBlock() # type: ignore[attr-defined]
def _add_output_to_block(block: _C.Block, value: _C.Value) -> int:
return block.registerOutput(value)
def _should_aten_fallback(
name: str, opset_version: int, operator_export_type: _C_onnx.OperatorExportTypes
):
# For all builds, if domain=="aten" and operator_export_type==ONNX_ATEN,
# an aten::ATen operator is created regardless of symbolics existence
is_exportable_aten_op = registration.registry.is_registered_op(name, opset_version)
is_onnx_aten_export = operator_export_type == _C_onnx.OperatorExportTypes.ONNX_ATEN
is_aten_fallback_export = (
operator_export_type == _C_onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK
)
if not name.startswith("aten::"):
return False
if is_onnx_aten_export or (is_aten_fallback_export and not is_exportable_aten_op):
return True
return False
def _get_aten_op_overload_name(n: _C.Node) -> str:
# Returns `overload_name` attribute to ATen ops on non-Caffe2 builds
schema = n.schema()
if not schema.startswith("aten::"):
return ""
return _C.parse_schema(schema).overload_name
def _run_symbolic_function(
graph: _C.Graph,
block: _C.Block,
node: _C.Node,
inputs: Any,
env: dict[_C.Value, _C.Value],
values_in_env: set[_C.Value],
new_nodes: list[_C.Node],
operator_export_type=_C_onnx.OperatorExportTypes.ONNX,
) -> _C.Value | Sequence[_C.Value | None] | None:
"""Runs a symbolic function.
The function is used in C++ to export the node to ONNX.
Returns:
A single or a tuple of Values.
None when the node gets cloned as is into the new graph.
"""
opset_version = GLOBALS.export_onnx_opset_version
# See Note [Export inplace]
node_kind = node.kind()
if node_kind.endswith("_"):
# Treat relu_ -> relu; add_ -> add etc.
ns_op_name = node_kind[:-1]
else:
ns_op_name = node_kind
namespace, op_name = jit_utils.parse_node_kind(ns_op_name)
graph_context = jit_utils.GraphContext(
graph=graph,
block=block,
opset=opset_version,
original_node=node,
params_dict=_params_dict,
env=env,
values_in_env=values_in_env,
new_nodes=new_nodes,
)
# Direct ATen export requested
if _should_aten_fallback(ns_op_name, opset_version, operator_export_type):
attrs = {
k + "_" + node.kindOf(k)[0]: symbolic_helper._node_get(node, k)
for k in node.attributeNames()
}
outputs = node.outputsSize()
attrs["outputs"] = outputs
return graph_context.aten_op(
op_name,
*inputs,
overload_name=_get_aten_op_overload_name(node),
**attrs,
)
try:
domain = namespace
symbolic_function_name = f"{domain}::{op_name}"
symbolic_function_group = registration.registry.get_function_group(
symbolic_function_name
)
if symbolic_function_group is not None:
symbolic_fn = symbolic_function_group.get(opset_version)
if symbolic_fn is not None:
# TODO Wrap almost identical attrs assignment or comment the difference.
attrs = {
k: symbolic_helper._node_get(node, k) for k in node.attributeNames()
}
return symbolic_fn(graph_context, *inputs, **attrs)
attrs = {
k + "_" + node.kindOf(k)[0]: symbolic_helper._node_get(node, k)
for k in node.attributeNames()
}
if namespace == "onnx":
# Clone node to trigger ONNX shape inference
return graph_context.op(
op_name, *inputs, **attrs, outputs=node.outputsSize()
) # type: ignore[attr-defined]
raise errors.UnsupportedOperatorError(
symbolic_function_name,
opset_version,
symbolic_function_group.get_min_supported()
if symbolic_function_group
else None,
)
except RuntimeError:
if operator_export_type == _C_onnx.OperatorExportTypes.ONNX_FALLTHROUGH:
return None
elif operator_export_type == _C_onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
# Emit ATen op for non-Caffe2 builds when `operator_export_type==ONNX_ATEN_FALLBACK`
attrs = {
k + "_" + node.kindOf(k)[0]: symbolic_helper._node_get(node, k)
for k in node.attributeNames()
}
return graph_context.aten_op(
op_name,
*inputs,
overload_name=_get_aten_op_overload_name(node),
**attrs,
)
raise
except TypeError as e:
# Handle the specific case where we didn't successfully dispatch.
# Otherwise, the backtrace will have the clues you need.
e.args = (f"{e.args[0]} \n(Occurred when translating {op_name}).",)
raise
def _verify_custom_op_name(symbolic_name: str):
if not re.match(r"^[a-zA-Z0-9-_]+::[a-zA-Z-_]+[a-zA-Z0-9-_]*$", symbolic_name):
raise errors.OnnxExporterError(
f"Failed to register operator {symbolic_name}. "
"The symbolic name must match the format domain::name, "
"and should start with a letter and contain only "
"alphanumerical characters"
)
ns, _ = jit_utils.parse_node_kind(symbolic_name)
if ns == "onnx":
raise ValueError(
f"Failed to register operator {symbolic_name}. {ns} domain cannot be modified."
)
def register_custom_op_symbolic(
symbolic_name: str,
symbolic_fn: Callable,
opset_version: int,
):
"""Registers a symbolic function for a custom operator.
When the user registers symbolic for custom/contrib ops,
it is highly recommended to add shape inference for that operator via setType API,
otherwise the exported graph may have incorrect shape inference in some extreme cases.
An example of setType is `test_aten_embedding_2` in `test_operators.py`.
See "Custom Operators" in the module documentation for an example usage.
Args:
symbolic_name (str): The name of the custom operator in "<domain>::<op>"
format.
symbolic_fn (Callable): A function that takes in the ONNX graph and
the input arguments to the current operator, and returns new
operator nodes to add to the graph.
opset_version (int): The ONNX opset version in which to register.
"""
if symbolic_name.startswith("::"):
symbolic_name = f"aten{symbolic_name}"
_verify_custom_op_name(symbolic_name)
registration.custom_onnx_symbolic(symbolic_name, opset_version)(symbolic_fn)
def unregister_custom_op_symbolic(symbolic_name: str, opset_version: int):
"""Unregisters ``symbolic_name``.
See "Custom Operators" in the module documentation for an example usage.
Args:
symbolic_name (str): The name of the custom operator in "<domain>::<op>"
format.
opset_version (int): The ONNX opset version in which to unregister.
"""
if symbolic_name.startswith("::"):
symbolic_name = f"aten{symbolic_name}"
_verify_custom_op_name(symbolic_name)
registration.registry.unregister(symbolic_name, opset_version)
def _validate_dynamic_axes(dynamic_axes, model, input_names, output_names):
"""Ensures dynamic axes argument is follows the expected format."""
if len(dynamic_axes) == 0:
return
if hasattr(model, "graph"):
# Extracting set of valid input/output names that shall be used for dynamic_axes
if (input_names is None) or len(input_names) == 0:
input_names = [x.debugName() for x in model.graph.inputs()]
if (output_names is None) or len(output_names) == 0:
output_names = [y.debugName() for y in model.graph.outputs()]
valid_names = set((input_names or []) + (output_names or []))
# If dynamic axes are provided as a list rather than dictionary, they should
# first get converted to a dictionary in expected format. If desired axes names
# are not provided for dynamic axes, automatic names shall be generated for
# provided dynamic axes of specified input/output
for key, value in dynamic_axes.items():
if key not in valid_names:
warnings.warn(
f"Provided key {key} for dynamic axes is not a valid input/output name"
)
if isinstance(value, list):
warnings.warn(
"No names were found for specified dynamic axes of provided input."
f"Automatically generated names will be applied to each dynamic axes of input {key}"
)
value_dict = {}
for i, x in enumerate(value):
if not isinstance(x, int):
raise ValueError(
"The type of axis index is expected to be an integer"
)
if x in value_dict:
warnings.warn(
f"Duplicate dynamic axis index {x} was provided for input {key}."
)
else:
value_dict[x] = str(key) + "_dynamic_axes_" + str(i + 1)
dynamic_axes[key] = value_dict
def model_signature(model: torch.nn.Module | Callable) -> inspect.Signature:
return inspect.signature(
model.forward if isinstance(model, torch.nn.Module) else model
)