MegEngine/imperative/python/megengine/autodiff/grad_manager.py

# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
import weakref
from collections import OrderedDict
from typing import Callable, Iterable, List, Union

from ..core._imperative_rt.core2 import pop_scope, push_scope, set_option
from ..core.autodiff.grad import Grad
from ..logger import get_logger
from ..tensor import Tensor
from ..utils.future import Future

logger = get_logger(__name__)

backwarding_grad_manager = None


def get_backwarding_grad_manager():
    return backwarding_grad_manager


class AttachSpec:
    __slots__ = "tensor", "callbacks"


_global_priority = 0


class GradManager:
    r"""
    GradManager computes gradients or more generally, vector-Jacobian product, by reverse mode
    automatic differentiation (a.k.a. back propagation).

    Reverse mode autodiff normally reuses many intermediate tensors for best computation efficiency.
    In a read-eval-print-loop (REPL) environment however, it is impossible to known how the user
    would take gradients later thus which tensors to keep. To solve this problem, the user must
    somehow declare beforehand which gradient could possibly be taken. With GradManager, users are
    required to call the :meth:`attach` method on a tensor if they want to take gradients with
    respect to it later. Furthermore, any computation on a tensor before it is attached is
    completely ignored from the autodiff perspective, so :meth:`attach` must be called before any
    computation that needs differentiation.

    For example, the following symbolic differentiation code

    .. code-block::

        x = get_x()
        y = f(x)
        dy = ones_like(y)
        dx = vjp(y, x, dy) # vector-Jacobian product

    can be rewriten using GradManager for REPL environment as

    .. code-block::

        with GradManager() as gm:
            x = get_x()
            gm.attach(x) # must be placed before any computation on x that needs differentiation
            y = f(x)
            dy = ones_like(y)
            gm.backward(y, dy) # doesn't need x, already known via attach()
            dx = x.grad # backward() saves result to .grad attribute

    A more realistic example of training a neural network would be like

    .. code-block::

        gm = GradManager()
        gm.attach(model.parameters())

        for data in dataset:
            with gm:
                loss = model(data)
                gm.backward(loss)
            # gradients w.r.t. parameters is accumulated into their .grad attributes

    You can also use ``record()`` and ``release()`` method instead of ``with`` context:

    .. code-block::

        gm = GradManager()
        gm.attach(model.parameters())

        for data in dataset:
            gm.record()
            loss = model(data)
            gm.backward(loss)
            # backward() will clear recorded history and free resources
            # call release() if backward() is not called
            # gm.release()

    For your convenience, GradManager may (not must) be reused. As shown in the examples, you
    only need to attach a tensor once and GradManager will remember it afterwards.
    However, a single GradManager can record only one computation history at a time. To run
    multiple differentiations simultaneously or perform high order differentiation, create
    as many GradManager as you need.

    .. note::

        Mutable tensors introduce ambiguities when doing symbolic differentiation: which version
        of the tensor are we referring to? For attached tensors, GradManager resolves this
        ambiguity by "snapshoting" them on first encounter, either on :meth:`record` (or entering
        with statement) if tensor is attached before :meth:`record`, or on :meth:`attach` if
        GradManager is already recording. Attached tensors will then be interpreted as their
        snapshotted version for differentiation purpose. The same ambiguity on the first parameter
        of :meth:`backward` is simply resolved by using the latest version.

    Typically, in data parallel, we would like to average the gradients across
    processes. Users will finally get the averaged gradients if an "AllReduce"
    callback is registered as follows:

    .. code-block::

        import megengine.distributed as dist

        gm = GradManager()
        gm.attach(model.parameters(), callback=dist.make_allreduce_cb("MEAN"))

    """

    def __init__(self):
        self._attach_specs = {}  # id(Tensor) -> AttachSpec
        self._recording = False
        self._grad = None
        self._after_backward_callback = []
        self._gradients = {}
        self._priority = None

    def attached_tensors(self):
        r"""Return attached tensor list from :meth:`attach`."""
        return [spec.tensor() for spec in self._attach_specs.values()]

    def attach(self, tensors: Iterable[Tensor], callbacks=None):
        r"""
        Instruct GradManager to track operations on tensors, so that gradients with respect
        to those tensors could be evaluated later.

        :meth:`attach` also accepts a list of callbacks, which will be called with the tensor and
        its gradient during :meth:`backward`. The signature of callbacks should look like:

            .. code-block::

                def callback(tensor: Tensor, grad: Tensor) -> Tensor:
                    ...
                    # returned grad is passed to subsequent callbacks
                    # and finally accumulated to the .grad attribute of tensor
                    return grad

        :meth:`attach` calls with overlapping tensors will result in their callbacks concatenated,
        independently for each tensor. For example,

            .. code-block::

                gm.attach([x, y], callbacks=[f])
                gm.attach([y], callbacks=[g])

        is equivalent to

            .. code-block::

                gm.attach([x], callbacks=[f])
                gm.attach([y], callbacks=[f, g])

        The effect of :meth:`attach` will persist across multiple uses of the GradManager. When
        reusing a GradManager, it is likely a mistake to call :meth:`attach` on the same set of
        tensors and callbacks repeatedly, which may grow the callback list indefinitely.

        .. note::

            When reusing a GradManager, it is sometimes desirable to attach temporary tensors each
            time, e.g. for computing gradients of inputs of a neural network. GradManager tries to
            accommodate such usages by holding weak references to attached tensors. Most of the
            times, this should be enough to prevent resource leak. Unfortunately, there are still
            some pitfalls left:

                - Callbacks should not hold strong references, directly or indirectly, to attached
                  tensors. Any strong reference, including those from callbacks, will prevent
                  garbage collection (even by the cycle collector!) of a attached tensor, until
                  the GradManager object is garbage collected.

            Please also note that GradManager might hold additional strong references to attached
            tensors when it is in use. This note only covers potential resource leaks across
            multiple uses of a GradManager, which is unrelated to whether resources is timely
            released within a single use.

        :param tensors: tensor or list of tensors to track
        :param callbacks: callback or list of callbacks
        """
        if callbacks is None:
            callbacks = []
        if isinstance(callbacks, Callable):
            callbacks = [callbacks]
        if isinstance(tensors, Tensor):
            tensors = [tensors]

        def make_spec(tensor):
            selfref = weakref.ref(self)
            key = id(tensor)

            def deleter(_):
                self = selfref()
                if self is not None:
                    del self._attach_specs[key]

            spec = AttachSpec()
            spec.tensor = weakref.ref(tensor, deleter)
            spec.callbacks = []
            return spec

        for x in tensors:
            assert isinstance(x, Tensor), "Object to be attached should be Tensor"
            spec = self._attach_specs.get(id(x))
            new_attach = spec is None
            if spec is None:
                spec = make_spec(x)
                self._attach_specs[id(x)] = spec
            spec.callbacks.extend(callbacks)
            if new_attach and self._recording:
                self._do_record(spec)

        return self

    def _register_after_backward_callback(self, callback):
        self._after_backward_callback.append(callback)
        return self

    def backward(
        self,
        y: Union[Tensor, List[Tensor]] = None,
        dy: Union[Tensor, List[Tensor]] = None,
    ):
        r"""
        Compute gradients (or vector-Jacobian product) for all attached tensors, accumulate to
        corresponding .grad attribute, and release resources along the way.

        :meth:`backward` computes the vector-Jacobian product :math:`dx_j = \sum_{i} dy_i J_{ij}`
        where :math:`J_{ij} = ∂y_i/∂x_j` is the Jacobian matrix between vector variables :math:`y`
        and :math:`x`, with all vectors involved represented as a list of tensors, in the sense of
        direct sums (or flatten-and-concatenate). :math:`y` and :math:`dy` are passed as the first
        and second parameter respectively, whereas :math:`x` is directly taken from the list of
        all attached tensors. The result :math:`dx` is also not returned. Instead, it is directly
        accumulated into the .grad attribute of matching attached tensors (a.k.a. :math:`x`). This
        can be done unambiguously since :math:`dx` as a list of tensors has the same structure as
        :math:`x`.

        If :math:`y` is a scalar and :math:`dy` is chosen to be 1, the vector-Jacobian product
        yield gradient of :math:`y` with repect to :math:`x` as a special case. In that case,
        you will be able to omit the :math:`dy` parameter and :meth:`backward` will automatically
        use 1 for it and compute the gradient.

        :meth:`backward` consumes all resources held by this GradManager and releases them in the
        process of this call. When the call successfully finishes, the GradManager will be put back
        to an inactive state.

        :param y: tensor or list of tensors
        :param dy: tensor or list of tensors. Defaults to 1 if y is scalar
        """
        push_scope("backward")
        set_option("record_computing_path", 0)
        from ..functional import ones_like

        global backwarding_grad_manager
        cache = backwarding_grad_manager
        backwarding_grad_manager = self
        if not self._recording:
            raise RuntimeError(
                "no computation history. "
                "did you forget record() or "
                "call a method that clears the history?"
            )
        assert self._grad is not None
        # These checks should be consistent with GradScaler's
        if y is None:
            ys = []
        elif isinstance(y, (tuple, list)):
            ys = y
        else:
            ys = [y]
        if dy is None:
            dys = [ones_like(y) for y in ys]
        elif isinstance(dy, (tuple, list)):
            dys = dy
        else:
            dys = [dy]
        try:
            self._grad(ys, dys)
            for callback in self._after_backward_callback:
                callback()
            for id_, grad in self._gradients.items():
                if isinstance(grad, Future):
                    grad = grad.get()
                spec = self._attach_specs.get(id_)
                tensor = spec and spec.tensor()
                if tensor is not None:
                    if tensor.grad is None:
                        tensor.grad = grad
                    else:
                        tensor.grad += grad
                    if tensor._isscalar() and tensor.grad is not None:
                        tensor.grad._setscalar()
        finally:
            self.release()
            backwarding_grad_manager = cache
        set_option("record_computing_path", 1)
        pop_scope("backward")

    def record(self):
        r"""
        Start recording operations

        After this call, you will be able to call :meth:`backward`.
        """
        global _global_priority
        if self._recording:
            raise RuntimeError("already recording")
        grad = Grad()
        self._recording = True
        self._grad = grad
        for spec in self._attach_specs.values():
            self._do_record(spec)
        if self._priority is None:
            grad._priority = _global_priority
            _global_priority -= 1
        grad.__enter__()

    def _do_record(self, spec):
        tensor = spec.tensor()
        if tensor is None:
            return

        def callback(grad, callbacks=spec.callbacks):
            for cb in callbacks:
                grad = cb(tensor, grad)
            self._gradients[id(tensor)] = grad

        # NOTE: override prev callback wrt when called serval times
        self._grad.wrt(tensor, callback=callback)

    def release(self):
        r"""
        Stop recording operations and release resources kept for gradient computation

        After this call, you will not be able to call :meth:`backward`.
        """
        global _global_priority
        if self._grad is not None:
            self._grad.__exit__(None, None, None)
            self._grad = None
        self._recording = False
        self._gradients = dict()
        if self._priority is None:
            _global_priority += 1

    def __enter__(self):
        self.record()
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.release()

    def __or__(self, other):
        if isinstance(other, GradManager):
            return GradManagerGroup([self, other])
        return NotImplemented

    __ror__ = __or__


class GradManagerGroup:
    def __init__(self, gms) -> None:
        self._gms = list(gms)

    def merge_with(self, other):
        if isinstance(other, GradManager):
            other = GradManagerGroup([other])
        elif not isinstance(other, GradManagerGroup):
            return NotImplemented
        return GradManagerGroup([*self._gms, *other._gms])

    __or__ = merge_with
    __ror__ = merge_with

    def __enter__(self):
        global _global_priority
        _global_priority += 1
        for gm in self._gms:
            gm._priority = _global_priority
            gm.record()

    def __exit__(self, exc_type, exc_val, exc_tb):
        global _global_priority
        _global_priority -= 1
        for gm in self._gms:
            gm.release()
            gm._priority = None