Source code for torch.onnx

The torch.onnx module contains 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.

import torch
import torch.jit
import torch.autograd
import torch.serialization
import re
import collections
import contextlib
import numbers
import warnings
import functools
import types
from torch._six import string_classes
from torch.autograd import Function, function
from torch.jit import _unique_state_dict

def set_training(model, mode):
    A context manager to temporarily set the training mode of 'model'
    to 'mode', resetting it when we exit the with-block.  A no-op if
    mode is None.
    if mode is None:
    old_mode =
    if old_mode != mode:
        if old_mode != mode:

[docs]def export(model, args, f, export_params=True, verbose=False, training=False, input_names=None, output_names=None, aten=False): """ Export a model into ONNX format. This exporter runs your model once in order to get a trace of its execution to be exported; at the moment, it supports a limited set of dynamic models (e.g., RNNs.) See also: :ref:`onnx-export` Arguments: model (torch.nn.Module): the model to be exported. args (tuple of arguments): the inputs to the model, e.g., such that ``model(*args)`` is a valid invocation of the model. Any non-Variable arguments will be hard-coded into the exported model; any Variable arguments will become inputs of the exported model, in the order they occur in args. If args is a Variable, this is equivalent to having called it with a 1-ary tuple of that Variable. (Note: passing keyword arguments to the model is not currently supported. Give us a shout if you need it.) f: a file-like object (has to implement fileno that returns a file descriptor) or a string containing a file name. A binary Protobuf will be written to this file. export_params (bool, default True): if specified, 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, the ordering as specified by ``model.state_dict().values()`` verbose (bool, default False): if specified, we will print out a debug description of the trace being exported. training (bool, default False): export the model in training mode. At the moment, ONNX is oriented towards exporting models for inference only, so you will generally not need to set this to True. input_names(list of strings, default empty list): names to assign to the input nodes of the graph, in order output_names(list of strings, default empty list): names to assign to the output nodes of the graph, in order aten (bool, default False): export the model in aten mode. If using aten mode, all the ops original exported by the functions in are exported as ATen ops. """ _export(model, args, f, export_params, verbose, training, input_names, output_names)
def _optimize_trace(trace, aten): # run dce first to eliminate dead parts of the graph that might have been # left behind by things like symbolic_override torch._C._jit_pass_dce(trace) torch._C._jit_pass_lint(trace) torch._C._jit_pass_peephole(trace) torch._C._jit_pass_lint(trace) torch._C._jit_pass_onnx(trace, aten) torch._C._jit_pass_lint(trace) torch._C._jit_pass_onnx_peephole(trace) torch._C._jit_pass_lint(trace) torch._C._jit_pass_dce(trace) torch._C._jit_pass_lint(trace) torch._C._jit_pass_canonicalize(trace) torch._C._jit_pass_lint(trace) def _trace(func, args, return_outs=False, aten=False): # Special case for common case of passing a single Variable if isinstance(args, torch.autograd.Variable): args = (args, ) trace, torch_out = torch.jit.trace(func, args) _optimize_trace(trace, aten) if return_outs: return trace, torch_out return trace def _export(model, args, f, export_params=True, verbose=False, training=False, input_names=None, output_names=None, aten=False): # Special case for common case of passing a single Variable if isinstance(args, torch.autograd.Variable): args = (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 = _unique_state_dict(model).keys() # By default, training=False, 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=True (or None, to preserve whatever the original # training mode was.) with set_training(model, training): trace, torch_out = torch.jit.trace(model, args) if orig_state_dict_keys != _unique_state_dict(model).keys(): raise RuntimeError("state_dict changed after running the tracer; " "something weird is happening in your model!") _optimize_trace(trace, aten) _set_input_and_output_names(trace.graph(), input_names, output_names) if verbose: print(trace) # TODO: Don't allocate a in-memory string for the protobuf from torch.onnx.symbolic import _onnx_opset_version if export_params: # NB: OrderedDict values is not actually a list, but trace.export is # not duck-typed and expects an actual list. proto = trace.export(list(_unique_state_dict(model).values()), _onnx_opset_version) else: proto = trace.export([], _onnx_opset_version) torch.serialization._with_file_like(f, "wb", lambda f: f.write(proto)) return torch_out 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) did not match number of %ss (%d)" % (descriptor, len(name_list), descriptor, len(node_list))) for name, node in zip(name_list, node_list): node.setUniqueName(name) set_names(list(graph.inputs()), input_names, 'input') set_names(list(graph.outputs()), output_names, 'output') attr_pattern = re.compile("^(.+)_([ifstgz])$") def _run_symbolic_method(op_name, symbolic_fn, args): """ This trampoline function gets invoked for every symbolic method call from C++. """ try: return symbolic_fn(*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 = ("{} (occurred when translating {})".format(e.args[0], op_name), ) raise def _is_onnx_list(value): if not isinstance(value, string_classes) and not torch.is_tensor(value) and isinstance(value, collections.Iterable): return True return False def _add_attribute(node, key, value, aten): """ initializes the right attribute based on type of value """ m = attr_pattern.match(key) if m is None: raise IndexError(( "Invalid attribute specifier '{}' names " + " must be suffixed with type, e.g. 'dim_i' or 'dims_i'").format(key)) name, kind =, if _is_onnx_list(value): kind += "s" if aten: if torch.is_tensor(value): # Caffe2 proto does not support tensor attribute. if value.numel() > 1: raise ValueError("Should not pass tensor attribute") value = _scalar(value) if isinstance(value, float): kind = "f" else: kind = "i" return getattr(node, kind + "_")(name, value) def _scalar(x): """Convert a scalar tensor into a Python value.""" assert x.numel() == 1 return x[0] def _newNode(g, opname, outputs, *args, **kwargs): aten = kwargs.pop("aten", False) n = g.create(opname, args, outputs) for k, v in sorted(kwargs.items()): # TODO: enable inplace in aten exporting mode. if k == "inplace": continue _add_attribute(n, k, v, aten=aten) return n def _graph_op(g, opname, *raw_args, **kwargs): """ Create an ONNX operator 'opname', taking 'args' as inputs and attributes 'kwargs'; returning the node representing the single output of this operator (see the `outputs` keyword argument for multi-return nodes). The set of operators and the inputs/attributes they take is documented at This function is monkey-patched onto Graph. Arguments: opname (string): The ONNX operator name, e.g., `Abs` or `Add`. args (Node...): The inputs to the operator; usually provided as arguments to the `symbolic` definition. kwargs: The attributes of the ONNX operator, with keys named according to the following convention: `alpha_f` indicates the `alpha` attribute with type `f`. The valid type specifiers are `f` (float), `i` (int), `s` (string) or `t` (Tensor). An attribute specified with type float accepts either a single float, or a list of floats (e.g., you would say `dims_i` for a `dims` attribute that takes a list of integers). outputs (int, optional): The number of outputs this operator returns; by default an operator is assumed to return a single output. If `outputs` is greater than one, this functions returns a tuple of output `Node`, representing each output of the ONNX operator in positional. """ outputs = kwargs.pop('outputs', 1) # Filter out None attributes, this can be convenient client side because # now they can pass through None attributes, and have them not show up kwargs = dict((k, v) for k, v in kwargs.items() if v is not None) def const_if_tensor(arg): if arg is None: return arg elif isinstance(arg, torch._C.Value): return arg else: return g.op("Constant", value_z=arg) args = list(const_if_tensor(arg) for arg in raw_args) n = g.appendNode(_newNode(g, opname, outputs, *args, **kwargs)) if outputs == 1: return n.output() return tuple(o for o in n.outputs()) # Note [Export inplace] # ~~~~~~~~~~~~~~~~~~~~~ # In abstract, it would be better for us to export inplace annotations, # than to not export them, since it is useful information that can # help the target of an ONNX export export more efficiently. However, # ONNX doesn't currently formalize inplace. Fortunately, it's sound to drop # inplace annotations, but we are losing information this way. def _run_symbolic_function(g, n, inputs, aten=False): import torch.onnx.symbolic try: # See Note [Export inplace] if n.kind().endswith('_'): op_name = n.kind()[:-1] elif n.kind().endswith('_forward'): # NB: it seems the tracing generates duplicate ops, # such as both threshold and threshold_forward, introduced in #4395. return inputs else: op_name = n.kind() # Export ops in aten mode. if aten: attrs = {k + "_" + n.kindOf(k)[0]: n[k] for k in n.attributeNames()} outputs = n.outputsSize() attrs["outputs"] = outputs return _graph_at(g, op_name, *inputs, aten=True, **attrs) # Export ONNX regular ops. attrs = {k: n[k] for k in n.attributeNames()} if not hasattr(torch.onnx.symbolic, op_name): warnings.warn("ONNX export failed on {} because torch.onnx.symbolic.{} does not exist" .format(op_name, op_name)) return None fn = getattr(torch.onnx.symbolic, op_name) return fn(g, *inputs, **attrs) 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 = ("{} (occurred when translating {})".format(e.args[0], op_name), ) raise # Generate an ONNX ATen op node. def _graph_at(g, opname, *args, **kwargs): return g.op("ATen", *args, operator_s=opname, **kwargs) # This helper function can create either constant tensor or constant scalar. # If dims is None or 0 or [0], generate a 0-d tensor (scalar). # # TODO: We might not need this anymore, since most scalars now show up # as tensors def _graph_constant(g, value, dims, type, *args, **kwargs): assert isinstance(value, numbers.Number) assert type is not None isscalar = False if dims is None or dims == 0 or set(dims) == set([0]): dims = [1] isscalar = True type = type.lower() if type == "char": tensor = torch.CharTensor(*dims) elif type == "short": tensor = torch.ShortTensor(*dims) elif type == "int": tensor = torch.IntTensor(*dims) elif type == "long": tensor = torch.LongTensor(*dims) elif type == "half": tensor = torch.HalfTensor(*dims) elif type == "float": tensor = torch.FloatTensor(*dims) elif type == "double": tensor = torch.DoubleTensor(*dims) else: raise ValueError("Unknown type, type should be one of the following strings: " "char, short, int, long, half, float, double") tensor.fill_(value) if isscalar: return g.op("Constant", *args, value_z=tensor, **kwargs) return g.op("Constant", *args, value_t=tensor, **kwargs) def _node_getitem(self, k): """ Accessor for attributes of a node which is polymorphic over return type. NB: This is monkey-patched onto Node. """ sel = self.kindOf(k) return getattr(self, sel)(k) def symbolic_override(symbolic_fn): """ Decorator to override ONNX export of the a function with specified subgraph. Effectively allows to attach symbolic() implementation to an arbitrary python function or autograd.Function. Requirements for the decorated function: - being non-member function or autograd.Function - positional inputs are Variables/Tensors or (nested) lists or tuples of them (similar requirement to NestedIOFunction) - outputs are similarly Variables/Tensors or (nested) lists or tuples of them - keyword arguments are of non-tensor type Example usage: ``` def symb(g, x, y): return g.op('Sum', x, y[0], y[1]) @symbolic_override(symb) def foo(x, y): return x + y[0] + y[1] ``` """ def wrapper_maker(fn): def wrapper(*args, **kwargs): output = fn(*args, **kwargs) flat_args = tuple(function._iter_variables(args)) if not any(map(torch._C._jit_is_tracing, flat_args)): return output flat_output_tensors = tuple( for v in function._iter_variables(output)) assert len(list(function._iter_variables_permissive( list(kwargs.values())))) == 0, \ "Passing Variable through kwargs is not supported" class ExportProxy(Function): @staticmethod def symbolic(g, *flat_args): symbolic_args = function._unflatten(flat_args, args) symbolic_output = symbolic_fn(g, *symbolic_args, **kwargs) return tuple(function._iter_jit_values(symbolic_output)) @staticmethod def forward(ctx, *unused_args): return flat_output_tensors @staticmethod def backward(ctx, *unused_args, **unused_kwargs): raise RuntimeError( "symbolic_override is meant for inference export only") flat_proxy_output = ExportProxy.apply(*flat_args) return function._unflatten(flat_proxy_output, output) # fn might be autograd.Function too, in this case wrapping doesn't work if isinstance(fn, types.FunctionType): wrapper = functools.wraps(fn)(wrapper) return wrapper return wrapper_maker torch._C.Graph.op = _graph_op = _graph_at torch._C.Graph.constant = _graph_constant torch._C.Node.__getitem__ = _node_getitem