verify.py

import torch
import torch.jit
import torch.onnx

import onnx
import onnx.helper

import numpy as np

import difflib
import io


def colonize(msg, sep=": "):
    if not msg:
        return ""
    else:
        return msg + sep


class Errors(object):
    """
    An error-collecting object which supports error recovery.

    It is intended to be used like a context manager:

    >>> with Errors("Top-level error message") as errs:
    >>>     ...
    """

    def __init__(self, msg, rtol=1e-3, atol=1e-5):
        self.msg = msg
        self.errors = []
        self.context = []
        self.rtol = rtol
        self.atol = atol

        # Allocated upon instance creation so that multiple Errors
        # can be used
        class ShortCircuit(Exception):
            pass
        self.exc_class = ShortCircuit

    def requireAlmostEqual(self, x, y, msg=None):
        """
        Test that x and y are nearly equal (equal within self.rtol
        precision); aborts execution if they are not.
        """
        self.almostEqualAndThen(x, y, msg, self.failWith)

    def checkAlmostEqual(self, x, y, msg=None):
        """
        Test that x and y are nearly equal (equal within self.rtol
        precision), but continue execution even if they are not equal.

        To prevent error cascades, you should remember to call 'failIfErrs'
        at some later point in time.
        """
        self.almostEqualAndThen(x, y, msg, self.addErr)

    def almostEqualAndThen(self, x, y, msg, k):
        """
        Helper for implementing 'requireAlmostEqual' and 'checkAlmostEqual'.
        Upon failure, invokes continuation 'k' with the error message.

        At the moment, only tests on 'numpy.ndarray' are supported.
        """
        if isinstance(x, np.ndarray) and isinstance(y, np.ndarray):
            try:
                np.testing.assert_allclose(x, y, rtol=self.rtol, atol=self.atol,
                                           equal_nan=True, verbose=True)
            except AssertionError as e:
                raise
                k("{}{}".format(colonize(msg), str(e).lstrip()))
        else:
            raise RuntimeError("Unsupported almost equal test")

    def requireEqual(self, x, y, msg=None):
        """
        Test that x and y are equal; aborts execution if they are not.
        """
        self.equalAndThen(x, y, msg, self.failWith)

    def checkEqual(self, x, y, msg=None):
        """
        Test that x and y are equal, but continue execution even if they are not equal.

        To prevent error cascades, you should remember to call 'failIfErrs'
        at some later point in time.
        """
        self.equalAndThen(x, y, msg, self.addErr)

    # Bit-for-bit accuracy test
    def equalAndThen(self, x, y, msg, k):
        """
        Helper for implementing 'requireEqual' and 'checkEqual'.  Upon failure,
        invokes continuation 'k' with the error message.
        """
        if isinstance(x, onnx.TensorProto) and isinstance(y, onnx.TensorProto):
            self.equalAndThen(x.name, y.name, msg, k)
            # Use numpy for the comparison
            t1 = onnx.numpy_helper.to_array(x)
            t2 = onnx.numpy_helper.to_array(y)
            new_msg = "{}In embedded parameter '{}'".format(colonize(msg), x.name)
            self.equalAndThen(t1, t2, new_msg, k)
        elif isinstance(x, np.ndarray) and isinstance(y, np.ndarray):
            try:
                np.testing.assert_equal(x, y)
            except AssertionError as e:
                raise
                k("{}{}".format(colonize(msg, ": "), str(e).lstrip()))
        else:
            if x != y:
                # TODO: Better algorithm for lists
                sx = str(x)
                sy = str(y)
                if len(sx) > 40 or len(sy) > 40 or '\n' in sx or '\n' in sy:
                    # long form
                    l = "=" * 50
                    k("\n{}The value\n{}\n{}\n{}\n\ndoes not equal\n\n{}\n{}\n{}"
                        .format(colonize(msg, ":\n"), l, sx, l, l, sy, l))
                else:
                    k("{}{} != {}".format(colonize(msg), sx, sy))

    def requireMultiLineEqual(self, x, y, msg=None):
        """
        Test that long, multi-line strings x and y are equal;
        aborts execution if they are not.
        """
        self.multiLineEqualAndThen(x, y, msg, self.failWith)

    def multiLineEqualAndThen(self, x, y, msg, k):
        """
        Helper for implementing 'requireMultiLineEqual'.  Upon failure,
        invokes continuation 'k' with the error message.
        """
        if msg is None:
            msg = "Strings are not equal"
        if x != y:
            diff = difflib.ndiff(x.splitlines(True), y.splitlines(True))
            k("{}{}".format(colonize(msg, ":\n\n"), "".join(diff)))

    def addErr(self, msg):
        """
        Add an error to the error context, but continue executing.
        """
        # TODO: instead of immediately concatenating the context in the msg,
        # attach it as metadata and make a decision how to format it later.
        msg_w_ctx = msg
        for c in reversed(self.context):
            msg += "\n\n  * " + "\n    ".join(c.splitlines())
        self.errors.append(msg)

    def fail(self):
        """
        Immediately fail and short-circuit to the next recovery context.

        NB: It is an error to 'fail' without having added any errors to
        the error context.
        """
        raise self.exc_class()

    def failWith(self, msg):
        """
        Add an error to the error context, and then short-circuit.
        """
        self.addErr(msg)
        self.fail()

    def failIfErrs(self):
        """
        If there are any errors in the error context, short-circuit.

        This is used to prevent error cascades.
        """
        if self.errors:
            self.fail()

    def recover(self):
        """
        Returns a context manager which can be used to recover in case of
        an error.  Example usage:

        >>> with errs.recover():
        >>>     ...
        """
        parent_self = self

        class Recover(object):
            def __enter__(self):
                pass

            def __exit__(self, exc_type, exc_value, traceback):
                if exc_type == parent_self.exc_class:
                    return True
        return Recover()

    def addErrCtxt(self, msg):
        """
        Returns a context manager which encloses a fragment of code with
        an extra contextual message, e.g., where an error occurred, or a hint
        applicable to all errors in the area.  Example usage:

        >>> with errs.addErrCtx("Some text"):
        >>>     ...
        """
        parent_self = self

        class AddContext(object):
            def __enter__(self):
                parent_self.context.append(msg)

            def __exit__(self, exc_type, exc_value, traceback):
                parent_self.context.pop()
        return AddContext()

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_value, traceback):
        if self.errors:
            errors_msg = "\n\n".join("ERROR: " + x for x in self.errors)
            final_msg = "{}\n{}\n{}".format(self.msg, '-' * 70, errors_msg)
            raise AssertionError(final_msg)
        if exc_type == self.exc_class:
            raise RuntimeError("ShortCircuit was raised, but no errors were recorded")

def verify(model, args, backend, verbose=False, training=torch.onnx.TrainingMode.EVAL, rtol=1e-3, atol=1e-7,
           test_args=2, do_constant_folding=True, example_outputs=None, opset_version=None,
           keep_initializers_as_inputs=True, add_node_names=False,
           operator_export_type=torch.onnx.OperatorExportTypes.ONNX):
    """
    Export a model into ONNX, import it into a specified ONNX backend, and then
    on a few random inputs verify that PyTorch and the backend produced the same
    results.  Requires onnx to be installed.

    This function may spuriously fail: some operators are implemented with
    different numerical precision in an ONNX backend, in which case an unstable
    network (e.g., Inception) may blow up these numerical instabilities.  This
    situation is less likely to happen if your model has been trained.  However,
    if this is not the case, you may have found a bug!  Please report it to the
    PyTorch developers.  You can also debug the issue yourself by removing
    suffixes of operators from your model until verification passes.

    For reproducibility, we recommend explicitly setting PyTorch's seed before
    invoking this function.

    Args:
        model (torch.nn.Module): the model to be exported and verified
        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.)
        backend (onnx.backend module): ONNX backend to verify with
        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.
        rtol (float, default 1e-3): relative precision required
        test_args (int or iterable of args, default 2):
            either an integer specifying the number
            of random arguments to generate, or an iterable producing arguments
            to test under.
        opset_version (int, default None): the opset version of the model to
            export. If not specified, the default value in symboli_helper will
            be used in utils._export().
        operator_export_type (enum, default OperatorExportTypes.ONNX): the operator
            export type to use when exporting the model. The default value converts
            all operators to ONNX ops.
    """
    def _nested_map(condition, fn, condition_msg=None):
        def _map(obj):
            if condition(obj):
                return fn(obj)
            elif obj is None:
                return None
            elif isinstance(obj, (list, tuple)):
                return type(obj)(_map(x) for x in obj)
            else:
                raise ValueError("Auto nesting doesn't know how to process "
                                 "an input object of type " + torch.typename(obj) +
                                 (". Accepted types: " + condition_msg +
                                  ", or lists/tuples of them"
                                  if condition_msg else ""))

        return _map

    def _iter_filter(condition, allow_unknown=False, condition_msg=None):
        def _iter(obj):
            if condition(obj):
                yield obj
            elif obj is None:
                return
            elif isinstance(obj, (list, tuple)):
                for o in obj:
                    for var in _iter(o):
                        yield var
            elif allow_unknown:
                yield obj
            else:
                raise ValueError("Auto nesting doesn't know how to process "
                                 "an input object of type " + torch.typename(obj) +
                                 (". Accepted types: " + condition_msg +
                                  ", or lists/tuples of them"
                                  if condition_msg else ""))

        return _iter

    def is_tensor(o):
        return isinstance(o, torch.Tensor)

    _iter_tensors = _iter_filter(is_tensor, condition_msg="Tensors")

    def randomize_arg(arg):
        new_data = arg.data.clone()
        # For now, don't try randomizing non-float tensors; these
        # are likely to be things like indices, where just randomly
        # spattering some longs is unlikely to work.  One way we could
        # make this work is to apply a random permutation or something.
        if arg.is_floating_point():
            new_data.uniform_()
        return torch.autograd.Variable(new_data, requires_grad=arg.requires_grad)

    randomize_args = _nested_map(is_tensor, randomize_arg)

    def backend_args(args):
        # TODO: onnx should accept iterables
        return tuple(v.data.cpu().numpy() for v in _iter_tensors(args))

    def load_bytes(b):
        b.seek(0)
        x = onnx.load(b)
        # doc_string has stack traces - let's remove them to make comparison
        # sane
        onnx.helper.strip_doc_string(x)
        return x

    # Special case for common case of passing a single Tensor
    if isinstance(args, torch.Tensor):
        args = (args,)

    with torch.onnx.select_model_mode_for_export(model, training):
        proto_bytes = io.BytesIO()
        torch_out = torch.onnx._export(model, args, proto_bytes, verbose=verbose,
                                       do_constant_folding=do_constant_folding,
                                       example_outputs=example_outputs,
                                       opset_version=opset_version,
                                       keep_initializers_as_inputs=keep_initializers_as_inputs,
                                       add_node_names=add_node_names,
                                       operator_export_type=operator_export_type)
        if isinstance(model, torch.jit.ScriptModule):
            torch_out = model(*args)
        proto = load_bytes(proto_bytes)
        prepared = backend.prepare(proto)

        def run(args):
            alt_proto_bytes = io.BytesIO()
            torch_out = torch.onnx._export(model, args, alt_proto_bytes, verbose=verbose,
                                           do_constant_folding=do_constant_folding,
                                           example_outputs=example_outputs,
                                           opset_version=opset_version,
                                           keep_initializers_as_inputs=keep_initializers_as_inputs,
                                           add_node_names=add_node_names,
                                           operator_export_type=operator_export_type)
            if isinstance(model, torch.jit.ScriptModule):
                torch_out = model(*args)
            alt_proto = load_bytes(alt_proto_bytes)
            if proto.SerializeToString() != alt_proto.SerializeToString():
                # OK, let's try to figure out what happened.
                msg = "When I exported your model with different inputs, the result was different."
                if not verbose:
                    msg += "\n(To get more information, run torch.onnx.verify(..., verbose=True))"
                with Errors(msg, rtol=rtol, atol=atol) as errs:
                    # First, check if we have the same number of parameters, and
                    # that they're the same order.  If they don't, something has *really* gone wrong.
                    initializer_order_hint = ("This is really strange! The second time I exported your model,\n"
                                              "it had a different set of parameters.  Are you assigning Parameters\n"
                                              "in the forward() of your model definition?")
                    with errs.addErrCtxt(initializer_order_hint):
                        errs.requireEqual([x.name for x in proto.graph.initializer],
                                          [x.name for x in alt_proto.graph.initializer],
                                          msg="Parameters list differs")

                    # Now check if the embedded parameters are actually the same
                    initializer_hint = ("A difference in embedded parameters usually means that\n"
                                        "your model is updating parameters/buffers even in inference\n"
                                        "mode.  Look for a buggy nn.Module which isn't respecting train().\n")
                    with errs.recover(), errs.addErrCtxt(initializer_hint):
                        for x, y in zip(proto.graph.initializer, alt_proto.graph.initializer):
                            errs.checkEqual(x, y)

                    # Next, check if the model structure lines up.
                    structure_hint = ("A difference in model structure usually means that\n"
                                      "your model has dynamic control flow.  These models are not\n"
                                      "currently supported by the exporter.")
                    with errs.recover(), errs.addErrCtxt(structure_hint):
                        # Delete initializers since we already tested them
                        stripped_proto = onnx.ModelProto()
                        stripped_proto.CopyFrom(proto)
                        del stripped_proto.graph.initializer[:]

                        stripped_alt_proto = onnx.ModelProto()
                        stripped_alt_proto.CopyFrom(alt_proto)
                        del stripped_alt_proto.graph.initializer[:]

                        # Compare the printable graph representations first
                        errs.requireMultiLineEqual(onnx.helper.printable_graph(stripped_proto.graph),
                                                   onnx.helper.printable_graph(stripped_alt_proto.graph))

                        # Compare the actual protobuf text formats now (not
                        # very user-friendly!)
                        errs.requireMultiLineEqual(str(stripped_proto), str(stripped_alt_proto))

                        # One last ditch effort, using built-in equality on
                        # protobufs
                        errs.requireEqual(stripped_proto, stripped_alt_proto)

                    errs.failIfErrs()

                    # At this point, we should have figured out why the binary
                    # protobufs differed, and short-circuited out of this code
                    # with a helpful error message.  But what if we didn't?
                    # We better still try to give a good error message in this
                    # case.  We EXPECT these requires to fail.  If they don't,
                    # that is a bug in verify
                    errs.requireEqual(proto, alt_proto)
                    errs.requireEqual(proto_bytes.getvalue(), alt_proto_bytes.getvalue())
                    raise AssertionError()

            # TODO: test that the traced model also returns the same thing...
            run_helper(torch_out, args)

        # Factored out so we can avoid one run of the model
        def run_helper(torch_out, args):
            backend_out = prepared.run(backend_args(args))
            if isinstance(torch_out, torch.Tensor):
                torch_out = (torch_out,)
            torch_out, _ = torch._C._jit_flatten(torch_out)
            # NB: onnx backend NEVER returns bare numpy array
            msg = "ONNX backend returned different results from PyTorch"
            result_hint = ("If you are not using trained parameters, a difference in results\n"
                           "could mean that your network is numerically unstable.  Otherwise\n"
                           "it indicates a bug in PyTorch/ONNX; please file a bug report.")
            with Errors(msg, rtol=rtol, atol=atol) as errs, errs.addErrCtxt(result_hint):
                for i, (x, y) in enumerate(zip(torch_out, backend_out)):
                    errs.checkAlmostEqual(x.data.cpu().numpy(), y, "In output {}".format(i))

        run_helper(torch_out, args)

        if isinstance(test_args, int):
            for i in range(test_args):
                run(randomize_args(args))
        else:
            for test_arg in test_args:
                run(test_arg)