ONNX exporter.

Open Neural Network eXchange (ONNX) is an open standard format for representing machine learning models. The torch.onnx module can export PyTorch models to ONNX. The model can then be consumed by any of the many runtimes that support ONNX.

Example: AlexNet from PyTorch to ONNX

Here is a simple script which exports a pretrained AlexNet to an ONNX file named alexnet.onnx. The call to torch.onnx.export runs the model once to trace its execution and then exports the traced model to the specified file:

import torch
import torchvision

dummy_input = torch.randn(10, 3, 224, 224, device="cuda")
model = torchvision.models.alexnet(pretrained=True).cuda()

# Providing input and output names sets the display names for values
# within the model's graph. Setting these does not change the semantics
# of the graph; it is only for readability.
# The inputs to the network consist of the flat list of inputs (i.e.
# the values you would pass to the forward() method) followed by the
# flat list of parameters. You can partially specify names, i.e. provide
# a list here shorter than the number of inputs to the model, and we will
# only set that subset of names, starting from the beginning.
input_names = [ "actual_input_1" ] + [ "learned_%d" % i for i in range(16) ]
output_names = [ "output1" ]

torch.onnx.export(model, dummy_input, "alexnet.onnx", verbose=True, input_names=input_names, output_names=output_names)

The resulting alexnet.onnx file contains a binary protocol buffer which contains both the network structure and parameters of the model you exported (in this case, AlexNet). The argument verbose=True causes the exporter to print out a human-readable representation of the model:

# These are the inputs and parameters to the network, which have taken on
# the names we specified earlier.
graph(%actual_input_1 : Float(10, 3, 224, 224)
      %learned_0 : Float(64, 3, 11, 11)
      %learned_1 : Float(64)
      %learned_2 : Float(192, 64, 5, 5)
      %learned_3 : Float(192)
      # ---- omitted for brevity ----
      %learned_14 : Float(1000, 4096)
      %learned_15 : Float(1000)) {
  # Every statement consists of some output tensors (and their types),
  # the operator to be run (with its attributes, e.g., kernels, strides,
  # etc.), its input tensors (%actual_input_1, %learned_0, %learned_1)
  %17 : Float(10, 64, 55, 55) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[11, 11], pads=[2, 2, 2, 2], strides=[4, 4]](%actual_input_1, %learned_0, %learned_1), scope: AlexNet/Sequential[features]/Conv2d[0]
  %18 : Float(10, 64, 55, 55) = onnx::Relu(%17), scope: AlexNet/Sequential[features]/ReLU[1]
  %19 : Float(10, 64, 27, 27) = onnx::MaxPool[kernel_shape=[3, 3], pads=[0, 0, 0, 0], strides=[2, 2]](%18), scope: AlexNet/Sequential[features]/MaxPool2d[2]
  # ---- omitted for brevity ----
  %29 : Float(10, 256, 6, 6) = onnx::MaxPool[kernel_shape=[3, 3], pads=[0, 0, 0, 0], strides=[2, 2]](%28), scope: AlexNet/Sequential[features]/MaxPool2d[12]
  # Dynamic means that the shape is not known. This may be because of a
  # limitation of our implementation (which we would like to fix in a
  # future release) or shapes which are truly dynamic.
  %30 : Dynamic = onnx::Shape(%29), scope: AlexNet
  %31 : Dynamic = onnx::Slice[axes=[0], ends=[1], starts=[0]](%30), scope: AlexNet
  %32 : Long() = onnx::Squeeze[axes=[0]](%31), scope: AlexNet
  %33 : Long() = onnx::Constant[value={9216}](), scope: AlexNet
  # ---- omitted for brevity ----
  %output1 : Float(10, 1000) = onnx::Gemm[alpha=1, beta=1, broadcast=1, transB=1](%45, %learned_14, %learned_15), scope: AlexNet/Sequential[classifier]/Linear[6]
  return (%output1);

You can also verify the output using the ONNX library, which you can install using pip:

pip install onnx

Then, you can run:

import onnx

# Load the ONNX model
model = onnx.load("alexnet.onnx")

# Check that the model is well formed

# Print a human readable representation of the graph

You can also run the exported model with one of the many runtimes that support ONNX. For example after installing ONNX Runtime, you can load and run the model:

import onnxruntime as ort

ort_session = ort.InferenceSession("alexnet.onnx")

outputs =
    {"actual_input_1": np.random.randn(10, 3, 224, 224).astype(np.float32)},

Here is a more involved tutorial on exporting a model and running it with ONNX Runtime.

Tracing vs Scripting

Internally, torch.onnx.export() requires a torch.jit.ScriptModule rather than a torch.nn.Module. If the passed-in model is not already a ScriptModule, export() will use tracing to convert it to one:

  • Tracing: If torch.onnx.export() is called with a Module that is not already a ScriptModule, it first does the equivalent of torch.jit.trace(), which executes the model once with the given args and records all operations that happen during that execution. This means that if your model is dynamic, e.g., changes behavior depending on input data, the exported model will not capture this dynamic behavior. We recommend examining the exported model and making sure the operators look reasonable. Tracing will unroll loops and if statements, exporting a static graph that is exactly the same as the traced run. If you want to export your model with dynamic control flow, you will need to use scripting.

  • Scripting: Compiling a model via scripting preserves dynamic control flow and is valid for inputs of different sizes. To use scripting:

    • Use torch.jit.script() to produce a ScriptModule.

    • Call torch.onnx.export() with the ScriptModule as the model. The args are still required, but they will be used internally only to produce example outputs, so that the types and shapes of the outputs can be captured. No tracing will be performed.

See Introduction to TorchScript and TorchScript for more details, including how to compose tracing and scripting to suit the particular requirements of different models.

Avoiding Pitfalls

Avoid NumPy and built-in Python types

PyTorch models can be written using NumPy or Python types and functions, but during tracing, any variables of NumPy or Python types (rather than torch.Tensor) are converted to constants, which will produce the wrong result if those values should change depending on the inputs.

For example, rather than using numpy functions on numpy.ndarrays:

# Bad! Will be replaced with constants during tracing.
x, y = np.random.rand(1, 2), np.random.rand(1, 2)
np.concatenate((x, y), axis=1)

Use torch operators on torch.Tensors:

# Good! Tensor operations will be captured during tracing.
x, y = torch.randn(1, 2), torch.randn(1, 2), y), dim=1)

And rather than use torch.Tensor.item() (which converts a Tensor to a Python built-in number):

# Bad! y.item() will be replaced with a constant during tracing.
def forward(self, x, y):
    return x.reshape(y.item(), -1)

Use torch’s support for implicit casting of single-element tensors:

# Good! y will be preserved as a variable during tracing.
def forward(self, x, y):
    return x.reshape(y, -1)


Using the field can produce an incorrect trace and therefore an incorrect ONNX graph. Use torch.Tensor.detach() instead. (Work is ongoing to remove entirely).

Avoid in-place operations when using tensor.shape in tracing mode

In tracing mode, shapes obtained from tensor.shape are traced as tensors, and share the same memory. This might cause a mismatch the final output values. As a workaround, avoid the use of inplace operations in these scenarios. For example, in the model:

class Model(torch.nn.Module):
  def forward(self, states):
      batch_size, seq_length = states.shape[:2]
      real_seq_length = seq_length
      real_seq_length += 2
      return real_seq_length + seq_length

real_seq_length and seq_length share the same memory in tracing mode. This could be avoided by rewriting the inplace operation:

real_seq_length = real_seq_length + 2



  • Only torch.Tensors, numeric types that can be trivially converted to torch.Tensors (e.g. float, int), and tuples and lists of those types are supported as model inputs or outputs. Dict and str inputs and outputs are accepted in tracing mode, but:

    • Any computation that depends on the value of a dict or a str input will be replaced with the constant value seen during the one traced execution.

    • Any output that is a dict will be silently replaced with a flattened sequence of its values (keys will be removed). E.g. {"foo": 1, "bar": 2} becomes (1, 2).

    • Any output that is a str will be silently removed.

  • Certain operations involving tuples and lists are not supported in scripting mode due to limited support in ONNX for nested sequences. In particular appending a tuple to a list is not supported. In tracing mode, the nested sequences will be flattened automatically during the tracing.

Differences in Operator Implementations

Due to differences in implementations of operators, running the exported model on different runtimes may produce different results from each other or from PyTorch. Normally these differences are numerically small, so this should only be a concern if your application is sensitive to these small differences.

Unsupported Tensor Indexing Patterns

Tensor indexing patterns that cannot be exported are listed below. If you are experiencing issues exporting a model that does not include any of the unsupported patterns below, please double check that you are exporting with the latest opset_version.

Reads / Gets

When indexing into a tensor for reading, the following patterns are not supported:

# Tensor indices that includes negative values.
data[torch.tensor([[1, 2], [2, -3]]), torch.tensor([-2, 3])]
# Workarounds: use positive index values.

Writes / Sets

When indexing into a Tensor for writing, the following patterns are not supported:

# Multiple tensor indices if any has rank >= 2
data[torch.tensor([[1, 2], [2, 3]]), torch.tensor([2, 3])] = new_data
# Workarounds: use single tensor index with rank >= 2,
#              or multiple consecutive tensor indices with rank == 1.

# Multiple tensor indices that are not consecutive
data[torch.tensor([2, 3]), :, torch.tensor([1, 2])] = new_data
# Workarounds: transpose `data` such that tensor indices are consecutive.

# Tensor indices that includes negative values.
data[torch.tensor([1, -2]), torch.tensor([-2, 3])] = new_data
# Workarounds: use positive index values.

# Implicit broadcasting required for new_data.
data[torch.tensor([[0, 2], [1, 1]]), 1:3] = new_data
# Workarounds: expand new_data explicitly.
# Example:
#   data shape: [3, 4, 5]
#   new_data shape: [5]
#   expected new_data shape after broadcasting: [2, 2, 2, 5]

Adding support for operators

When exporting a model that includes unsupported operators, you’ll see an error message like:

RuntimeError: ONNX export failed: Couldn't export operator foo

When that happens, there are a few things you can do:

  1. Change the model to not use that operator.

  2. Create a symbolic function to convert the operator and register it as a custom symbolic function.

  3. Contribute to PyTorch to add the same symbolic function to torch.onnx itself.

If you decided to implement a symbolic function (we hope you will contribute it back to PyTorch!), here is how you can get started:

ONNX exporter internals

A “symbolic function” is a function that decomposes a PyTorch operator into a composition of a series of ONNX operators.

During export, each node (which contains a PyTorch operator) in the TorchScript graph is visited by the exporter in topological order. Upon visiting a node, the exporter looks for a registered symbolic functions for that operator. Symbolic functions are implemented in Python. A symbolic function for an op named foo would look something like:

def foo(
  input_0: torch._C.Value,
  input_1: torch._C.Value) -> Union[None, torch._C.Value, List[torch._C.Value]]:
  Adds the ONNX operations representing this PyTorch function by updating the
  graph g with `g.op()` calls.

    g (Graph): graph to write the ONNX representation into.
    input_0 (Value): value representing the variables which contain
        the first input for this operator.
    input_1 (Value): value representing the variables which contain
        the second input for this operator.

    A Value or List of Values specifying the ONNX nodes that compute something
    equivalent to the original PyTorch operator with the given inputs.

    None if it cannot be converted to ONNX.

The torch._C types are Python wrappers around the types defined in C++ in ir.h.

The process for adding a symbolic function depends on the type of operator.

ATen operators

ATen is PyTorch’s built-in tensor library. If the operator is an ATen operator (shows up in the TorchScript graph with the prefix aten::), make sure it is not supported already.

List of supported operators

Visit the auto generated list of supported TorchScript operators for details on which operator are supported in each opset_version.

Adding support for an aten or quantized operator

If the operator is not in the list above:

  • Define the symbolic function in torch/onnx/symbolic_opset<version>.py, for example torch/onnx/ Make sure the function has the same name as the ATen function, which may be declared in torch/_C/_VariableFunctions.pyi or torch/nn/functional.pyi (these files are generated at build time, so will not appear in your checkout until you build PyTorch).

  • By default, the first arg is the ONNX graph. Other arg names must EXACTLY match the names in the .pyi file, because dispatch is done with keyword arguments.

  • In the symbolic function, if the operator is in the ONNX standard operator set, we only need to create a node to represent the ONNX operator in the graph. If not, we can compose several standard operators that have the equivalent semantics to the ATen operator.

Here is an example of handling missing symbolic function for the ELU operator.

If we run the following code:

        torch.nn.ELU(), # module
        torch.ones(1)   # example input

We see something like:

graph(%self : __torch__.torch.nn.modules.activation.___torch_mangle_0.ELU,
      %input : Float(1, strides=[1], requires_grad=0, device=cpu)):
  %4 : float = prim::Constant[value=1.]()
  %5 : int = prim::Constant[value=1]()
  %6 : int = prim::Constant[value=1]()
  %7 : Float(1, strides=[1], requires_grad=0, device=cpu) = aten::elu(%input, %4, %5, %6)
  return (%7)

Since we see aten::elu in the graph, we know this is an ATen operator.

We check the ONNX operator list, and confirm that Elu is standardized in ONNX.

We find a signature for elu in torch/nn/functional.pyi:

def elu(input: Tensor, alpha: float = ..., inplace: bool = ...) -> Tensor: ...

We add the following lines to

def elu(g, input: torch.Value, alpha: torch.Value, inplace: bool = False):
    return g.op("Elu", input, alpha_f=alpha)

Now PyTorch is able to export models containing the aten::elu operator!

See the torch/onnx/symbolic_opset*.py files for more examples.


If the operator is a sub-class of torch.autograd.Function, there are three ways to export it.

Static Symbolic Method

You can add a static method named symbolic to your function class. It should return ONNX operators that represent the function’s behavior in ONNX. For example:

class MyRelu(torch.autograd.Function):
    def forward(ctx, input: torch.Tensor) -> torch.Tensor:
        return input.clamp(min=0)

    def symbolic(g: torch.Graph, input: torch.Value) -> torch.Value:
        return g.op("Clip", input, g.op("Constant", value_t=torch.tensor(0, dtype=torch.float)))

Inline Autograd Function

In cases where a static symbolic method is not provided for its subsequent torch.autograd.Function or where a function to register prim::PythonOp as custom symbolic functions is not provided, torch.onnx.export() tries to inline the graph that corresponds to that torch.autograd.Function such that this function is broken down into individual operators that were used within the function. The export should be successful as long as these individual operators are supported. For example:

class MyLogExp(torch.autograd.Function):
    def forward(ctx, input: torch.Tensor) -> torch.Tensor:
        h = input.exp()
        return h.log().log()

There is no static symbolic method present for this model, yet it is exported as follows:

graph(%input : Float(1, strides=[1], requires_grad=0, device=cpu)):
    %1 : float = onnx::Exp[](%input)
    %2 : float = onnx::Log[](%1)
    %3 : float = onnx::Log[](%2)
    return (%3)

If you need to avoid inlining of torch.autograd.Function, you should export models with operator_export_type set to ONNX_FALLTHROUGH or ONNX_ATEN_FALLBACK.

Custom operators

If a model uses a custom operator implemented in C++ as described in Extending TorchScript with Custom C++ Operators, you can export it by following this example:

from torch.onnx import symbolic_helper

# Define custom symbolic function
@symbolic_helper.parse_args("v", "v", "f", "i")
def symbolic_foo_forward(g, input1, input2, attr1, attr2):
    return g.op("custom_domain::Foo", input1, input2, attr1_f=attr1, attr2_i=attr2)

# Register custom symbolic function
torch.onnx.register_custom_op_symbolic("custom_ops::foo_forward", symbolic_foo_forward, 9)

class FooModel(torch.nn.Module):
    def __init__(self, attr1, attr2):
        self.attr1 = attr1
        self.attr2 = attr2

    def forward(self, input1, input2):
        # Calling custom op
        return torch.ops.custom_ops.foo_forward(input1, input2, self.attr1, self.attr2)

model = FooModel(attr1, attr2)
    (example_input1, example_input1),
    # only needed if you want to specify an opset version > 1.
    custom_opsets={"custom_domain": 2}

You can export your model as one or a combination of many standard ONNX ops, or as a custom ONNX operator.

The example above exports it as a custom operator in the “custom_domain” opset. When exporting a custom operator, you can specify the custom domain version using the custom_opsets dictionary at export. If not specified, the custom opset version defaults to 1.

The runtime that consumes the model needs to support the custom op. See Caffe2 custom ops, ONNX Runtime custom ops, or your runtime of choice’s documentation.

Discovering all unconvertible ATen ops at once

When export fails due to an unconvertible ATen op, there may in fact be more than one such op but the error message only mentions the first. To discover all of the unconvertible ops in one go you can:

# prepare model, args, opset_version

torch_script_graph, unconvertible_ops = torch.onnx.utils.unconvertible_ops(
    model, args, opset_version=opset_version


The set 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, but this should give you a general idea of what ops are not supported. Please feel free to open Github Issues for op support requests.

Frequently Asked Questions

Q: I have exported my LSTM model, but its input size seems to be fixed?

The tracer records the shapes of the example inputs. If the model should accept inputs of dynamic shapes, set dynamic_axes when calling torch.onnx.export().

Q: How to export models containing loops?

Q: How to export models with primitive type inputs (e.g. int, float)?

Support for primitive numeric type inputs was added in PyTorch 1.9. However, the exporter does not support models with str inputs.

Q: Does ONNX support implicit scalar datatype casting?

The ONNX standard does not, but the exporter will try to handle that part. Scalars are exported as constant tensors. The exporter will figure out the right data type for scalars. In rare cases when it is unable to do so, you will need to manually specify the datatype with e.g. dtype=torch.float32. If you see any errors, please [create a GitHub issue](

Q: Are lists of Tensors exportable to ONNX?

Yes, for opset_version >= 11, since ONNX introduced the Sequence type in opset 11.

Contributing / developing

Developer docs.


torch.onnx.export(model, args, f, export_params=True, verbose=False, training=<TrainingMode.EVAL: 0>, input_names=None, output_names=None, operator_export_type=<OperatorExportTypes.ONNX: 0>, opset_version=None, do_constant_folding=True, dynamic_axes=None, keep_initializers_as_inputs=None, custom_opsets=None, export_modules_as_functions=False)[source]

Exports a model into ONNX format.

If model is not a torch.jit.ScriptModule nor a torch.jit.ScriptFunction, this runs model once in order to convert it to a TorchScript graph to be exported (the equivalent of torch.jit.trace()). Thus this has the same limited support for dynamic control flow as torch.jit.trace().

  • model (torch.nn.Module, torch.jit.ScriptModule or torch.jit.ScriptFunction) – the model to be exported.

  • args (tuple or torch.Tensor) –

    args can be structured either as:


      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.

    1. A TENSOR:

      args = torch.Tensor([1])

    This is equivalent to a 1-ary tuple of that Tensor.


      args = (
              "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.


    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:

            # WRONG: will be interpreted as named arguments
            {y: z}


            {y: z},

  • f (Union[str, BytesIO]) – a file-like object (such that f.fileno() returns a file descriptor) or a string containing a file name. A binary protocol buffer will be written to this file.

  • export_params (bool, default True) – 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 (bool, default False) – 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 (enum, default TrainingMode.EVAL) –

    • TrainingMode.EVAL: export the model in inference mode.

    • TrainingMode.PRESERVE: export the model in inference mode if is

      False and in training mode if 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) –

    • 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 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 is PyTorch’s built-in tensor library, so this instructs the runtime to use PyTorch’s implementation of these ops.


      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)

      If PyTorch was built with Caffe2 (i.e. with BUILD_CAFFE2=1), then Caffe2-specific behavior will be enabled, including special support for ops are produced by the modules described in Quantization.


      Models exported this way are probably runnable only by Caffe2.

  • opset_version (int, default 14) – The version of the default (ai.onnx) opset to target. Must be >= 7 and <= 16.

  • do_constant_folding (bool, default True) – 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 (dict[string, dict[int, string]] or dict[string, list(int)], default empty dict) –

    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


    • 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.ones(2, 2),),


    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


        (torch.ones(2, 2),),
            # dict value: manually named axes
            "x": {0: "my_custom_axis_name"},
            # list value: automatic names
            "sum": [0],


    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 (bool, default None) –

    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 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.

    If None, then the behavior is chosen automatically as follows:

    • If operator_export_type=OperatorExportTypes.ONNX, the behavior is equivalent

      to setting this argument to False.

    • Else, the behavior is 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 (bool or set of python:type of nn.Module, default False) –

    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 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.

  • torch.onnx.errors.CheckerError – If the ONNX checker detects an invalid ONNX graph.

  • torch.onnx.errors.UnsupportedOperatorError – If the ONNX graph cannot be exported because it uses an operator that is not supported by the exporter.

  • torch.onnx.errors.OnnxExporterError – Other errors that can occur during export. All errors are subclasses of errors.OnnxExporterError.

torch.onnx.export_to_pretty_string(model, args, export_params=True, verbose=False, training=<TrainingMode.EVAL: 0>, input_names=None, output_names=None, operator_export_type=<OperatorExportTypes.ONNX: 0>, 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)[source]

Similar to export(), but returns a text representation of the ONNX model. Only differences in args listed below. All other args are the same as export().

  • add_node_names (bool, default True) – Whether or not to set 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.


A UTF-8 str containing a human-readable representation of the ONNX model.

torch.onnx.register_custom_op_symbolic(symbolic_name, symbolic_fn, opset_version)[source]

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

See “Custom Operators” in the module documentation for an example usage.

  • 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.

torch.onnx.unregister_custom_op_symbolic(symbolic_name, opset_version)[source]

Unregisters symbolic_name.

See “Custom Operators” in the module documentation for an example usage.

  • symbolic_name (str) – The name of the custom operator in “<domain>::<op>” format.

  • opset_version (int) – The ONNX opset version in which to unregister.

torch.onnx.select_model_mode_for_export(model, mode)[source]

A context manager to temporarily set the training mode of model to mode, resetting it when we exit the with-block.

  • model – Same type and meaning as model arg to export().

  • mode (TrainingMode) – Same type and meaning as training arg to export().


Returns whether it is in the middle of ONNX export.

Return type:



Enables ONNX logging.


Disables ONNX logging.



Scalar types defined in torch.


Access comprehensive developer documentation for PyTorch

View Docs


Get in-depth tutorials for beginners and advanced developers

View Tutorials


Find development resources and get your questions answered

View Resources