autograd.py

# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.

# coding: utf-8
"""Autograd for NDArray."""

from array import array
from threading import Lock
import traceback
import ctypes
from ctypes import c_int, c_void_p, CFUNCTYPE, POINTER, cast
from .base import _LIB, check_call, string_types, mx_uint
from .base import NDArrayHandle, c_array, c_handle_array, c_array_buf, MXCallbackList, SymbolHandle
from .ndarray import NDArray, _ndarray_cls
from .ndarray import _GRAD_REQ_MAP
from .symbol import Symbol
from .util import is_np_array


def set_recording(is_recording): #pylint: disable=redefined-outer-name
    """Set status to recording/not recording. When recording, graph will be constructed
    for gradient computation.

    Parameters
    ----------
    is_recording: bool

    Returns
    -------
    previous state before this set.
    """
    prev = ctypes.c_int()
    check_call(_LIB.MXAutogradSetIsRecording(
        ctypes.c_int(is_recording), ctypes.byref(prev)))
    return bool(prev.value)

def set_training(train_mode): #pylint: disable=redefined-outer-name
    """Set status to training/predicting. This affects ctx.is_train in operator
    running context. For example, Dropout will drop inputs randomly when
    train_mode=True while simply passing through if train_mode=False.

    Parameters
    ----------
    train_mode: bool

    Returns
    -------
    previous state before this set.
    """
    prev = ctypes.c_int()
    check_call(_LIB.MXAutogradSetIsTraining(
        ctypes.c_int(train_mode), ctypes.byref(prev)))
    return bool(prev.value)

def is_recording():
    """Get status on recording/not recording.

    Returns
    -------
    Current state of recording.
    """
    curr = ctypes.c_bool()
    check_call(_LIB.MXAutogradIsRecording(ctypes.byref(curr)))
    return curr.value

def is_training():
    """Get status on training/predicting.

    Returns
    -------
    Current state of training/predicting.
    """
    curr = ctypes.c_bool()
    check_call(_LIB.MXAutogradIsTraining(ctypes.byref(curr)))
    return curr.value


class _RecordingStateScope(object):
    """Scope for managing training state.

    Example::

        with _RecordingStateScope(True, True):
            y = model(x)
            backward([y])

    """
    def __init__(self, is_record, train_mode): #pylint: disable=redefined-outer-name
        self._enter_is_record = is_record
        self._enter_train_mode = train_mode
        self._prev_is_record = None
        self._prev_train_mode = None

    def __enter__(self):
        if self._enter_is_record is not None:
            self._prev_is_record = set_recording(self._enter_is_record)
        if self._enter_train_mode is not None:
            self._prev_train_mode = set_training(self._enter_train_mode)

    def __exit__(self, ptype, value, trace):
        if self._enter_is_record is not None and self._prev_is_record != self._enter_is_record:
            set_recording(self._prev_is_record)
        if self._enter_train_mode is not None and self._prev_train_mode != self._enter_train_mode:
            set_training(self._prev_train_mode)


def record(train_mode=True): #pylint: disable=redefined-outer-name
    """Returns an autograd recording scope context to be used in 'with' statement
    and captures code that needs gradients to be calculated.

    .. note:: When forwarding with train_mode=False, the corresponding backward
              should also use train_mode=False, otherwise gradient is undefined.

    Example::

        with autograd.record():
            y = model(x)
            backward([y])
        metric.update(...)
        optim.step(...)

    Parameters
    ----------
    train_mode: bool, default True
        Whether the forward pass is in training or predicting mode. This controls the behavior
        of some layers such as Dropout, BatchNorm.
    """
    return _RecordingStateScope(True, train_mode)


def pause(train_mode=False): #pylint: disable=redefined-outer-name
    """Returns a scope context to be used in 'with' statement for codes that do not need
    gradients to be calculated.

    Example::

        with autograd.record():
            y = model(x)
            backward([y])
            with autograd.pause():
                # testing, IO, gradient updates...

    Parameters
    ----------
    train_mode: bool, default False
        Whether to do forward for training or predicting.
    """
    return _RecordingStateScope(False, train_mode)


def train_mode():
    """Returns a scope context to be used in 'with' statement
    in which forward pass behavior is set to training mode,
    without changing the recording states.

    Example::

        y = model(x)
        with autograd.train_mode():
            y = dropout(y)

    """
    return _RecordingStateScope(None, True)


def predict_mode():
    """Returns a scope context to be used in 'with' statement
    in which forward pass behavior is set to inference mode,
    without changing the recording states.

    Example::

        with autograd.record():
            y = model(x)
            with autograd.predict_mode():
                y = sampling(y)
            backward([y])
    """
    return _RecordingStateScope(None, False)


def mark_variables(variables, gradients, grad_reqs='write'):
    """Mark NDArrays as variables to compute gradient for autograd.

    This is equivalent to the function .attach_grad() in a variable, but with this
    call we can set the gradient to any value.

    Parameters
    ----------
    variables: NDArray or list of NDArray
    gradients: NDArray or list of NDArray
    grad_reqs: str or list of str
    """
    if isinstance(variables, NDArray):
        assert isinstance(gradients, NDArray)
        variables = [variables]
        gradients = [gradients]

    if isinstance(grad_reqs, string_types):
        grad_reqs = [_GRAD_REQ_MAP[grad_reqs]]*len(variables)
    else:
        grad_reqs = [_GRAD_REQ_MAP[i] for i in grad_reqs]

    check_call(_LIB.MXAutogradMarkVariables(
        len(variables),
        c_handle_array(variables),
        c_array_buf(mx_uint, array('I', grad_reqs)),
        c_handle_array(gradients)))


def _parse_head(heads, head_grads):
    """parse head gradient for backward and grad."""
    if isinstance(heads, NDArray):
        heads = [heads]
    if isinstance(head_grads, NDArray):
        head_grads = [head_grads]

    head_handles = c_handle_array(heads)

    if head_grads is None:
        hgrad_handles = ctypes.c_void_p(0)
    else:
        msg = "heads and head_grads must be lists of the same length: {} vs. {}"
        assert len(heads) == len(head_grads), msg.format(len(heads), len(head_grads))
        hgrad_handles = c_array(NDArrayHandle,
                                [i.handle if i is not None else NDArrayHandle(0)
                                 for i in head_grads])
    return head_handles, hgrad_handles


def backward(heads, head_grads=None, retain_graph=False, train_mode=True): #pylint: disable=redefined-outer-name
    """Compute the gradients of heads w.r.t previously marked variables.

    Parameters
    ----------
    heads: NDArray or list of NDArray
        Output NDArray(s)
    head_grads: NDArray or list of NDArray or None
        Gradients with respect to heads.
    train_mode: bool, optional
        Whether to do backward for training or predicting.
    """
    head_handles, hgrad_handles = _parse_head(heads, head_grads)

    check_call(_LIB.MXAutogradBackwardEx(
        len(head_handles),
        head_handles,
        hgrad_handles,
        0,
        ctypes.c_void_p(0),
        ctypes.c_int(retain_graph),
        ctypes.c_int(0),
        ctypes.c_int(train_mode),
        ctypes.c_void_p(0),
        ctypes.c_void_p(0)))


def grad(heads, variables, head_grads=None, retain_graph=None, create_graph=False,
         train_mode=True):  #pylint: disable=redefined-outer-name
    """Compute the gradients of heads w.r.t variables. Gradients will be
    returned as new NDArrays instead of stored into `variable.grad`.
    Supports recording gradient graph for computing higher order gradients.

    .. note::

      Currently only a very limited set of operators support higher order \
      gradients.

    Parameters
    ----------
    heads: NDArray or list of NDArray
        Output NDArray(s)
    variables: NDArray or list of NDArray
        Input variables to compute gradients for.
    head_grads: NDArray or list of NDArray or None
        Gradients with respect to heads.
    retain_graph: bool
        Whether to keep computation graph to differentiate again, instead
        of clearing history and release memory. Defaults to the same value
        as create_graph.
    create_graph: bool
        Whether to record gradient graph for computing higher order
    train_mode: bool, optional
        Whether to do backward for training or prediction.

    Returns
    -------
    NDArray or list of NDArray:
        Gradients with respect to variables.

    Examples
    --------
    >>> x = mx.nd.ones((1,))
    >>> x.attach_grad()
    >>> with mx.autograd.record():
    ...     z = mx.nd.elemwise_add(mx.nd.exp(x), x)
    >>> dx = mx.autograd.grad(z, [x], create_graph=True)
    >>> print(dx)
    [
    [ 3.71828175]
    <NDArray 1 @cpu(0)>]
    """
    head_handles, hgrad_handles = _parse_head(heads, head_grads)

    if isinstance(variables, NDArray):
        variables = [variables]
    else:
        assert len(variables), "variables cannot be an empty list."
    var_handles = c_handle_array(variables)

    retain_graph = retain_graph if retain_graph is not None else create_graph
    grad_vars = ctypes.POINTER(NDArrayHandle)()
    grad_stypes = ctypes.POINTER(ctypes.c_int)()

    check_call(_LIB.MXAutogradBackwardEx(
        len(head_handles),
        head_handles,
        hgrad_handles,
        len(var_handles),
        var_handles,
        ctypes.c_int(retain_graph),
        ctypes.c_int(create_graph),
        ctypes.c_int(train_mode),
        ctypes.byref(grad_vars),
        ctypes.byref(grad_stypes)))

    ret = [_ndarray_cls(ctypes.cast(grad_vars[i], NDArrayHandle),
                        stype=grad_stypes[i])
           for i in range(len(var_handles))]
    if isinstance(variables, NDArray):
        return ret[0]
    return ret


def get_symbol(x):
    """Retrieve recorded computation history as `Symbol`.

    Parameters
    ----------
    x : NDArray
        Array representing the head of computation graph.

    Returns
    -------
    Symbol
        The retrieved Symbol.
    """
    assert isinstance(x, NDArray), \
       f"get_symbol: Invalid argument type, expecting {NDArray}, got {type(x)}"
    hdl = SymbolHandle()
    check_call(_LIB.MXAutogradGetSymbol(x.handle, ctypes.byref(hdl)))
    return Symbol(hdl)


class Function(object):
    """Customize differentiation in autograd.

    If you don't want to use the gradients computed by the default
    chain-rule, you can use Function to customize differentiation for
    computation. You define your computation in
    the forward method and provide the customized differentiation
    in the backward method. During gradient computation, autograd will
    use the user-defined backward function instead of the default chain-rule.
    You can also cast to numpy array and back for some operations in
    forward and backward.

    For example, a stable sigmoid function can be defined as::

        class sigmoid(mx.autograd.Function):
            def forward(self, x):
                y = 1 / (1 + mx.nd.exp(-x))
                self.save_for_backward(y)
                return y

            def backward(self, dy):
                # backward takes as many inputs as forward's return value,
                # and returns as many NDArrays as forward's arguments.
                y, = self.saved_tensors
                return dy * y * (1-y)

    Then, the function can be used in the following way::

        func = sigmoid()
        x = mx.nd.random.uniform(shape=(10,))
        x.attach_grad()

        with mx.autograd.record():
            m = func(x)
            m.backward()
        dx = x.grad.asnumpy()

    """
    _bwd_functype = CFUNCTYPE(c_int, c_int, c_int, POINTER(c_void_p),
                              POINTER(c_int), c_int, c_void_p)
    _del_functype = CFUNCTYPE(c_int, c_void_p)
    class _Registry(object):
        """CustomOp registry."""
        def __init__(self):
            self.ref_holder = {}
            self.counter = 0
            self.lock = Lock()

        def inc(self):
            """Get index for new entry."""
            self.lock.acquire()
            cur = self.counter
            self.counter += 1
            self.lock.release()
            return cur

    _registry = _Registry()

    def __init__(self):
        self._used = False
        self.saved_tensors = ()

    def save_for_backward(self, *args):
        self.saved_tensors = args

    def __call__(self, *inputs):
        assert not self._used, \
            "Each Function instance can only be called once. "\
            "Please create another instance."
        self._used = True

        prev_recording = set_recording(False)
        outputs = self.forward(*inputs)
        set_recording(prev_recording)

        if not prev_recording:
            return outputs

        ret_outputs = outputs
        if isinstance(outputs, NDArray):
            outputs = (outputs,)

        key = Function._registry.inc()
        if is_np_array():
            from .numpy import ndarray
            array_cls = ndarray
        else:
            array_cls = NDArray

        def backward_entry(num_ograds, num_igrads, ptrs, reqs, is_train, _):
            """entry point for backward."""
            # pylint: disable=W0613
            try:
                output_grads = [array_cls(ctypes.cast(i, NDArrayHandle), writable=False) \
                                for i in ptrs[:num_ograds]]
                input_grads = [array_cls(ctypes.cast(i, NDArrayHandle), writable=True) \
                               for i in ptrs[num_ograds:num_ograds+num_igrads]]
                reqs = [reqs[i] for i in range(num_igrads)]
                rets = self.backward(*output_grads)
                if isinstance(rets, array_cls):
                    rets = (rets,)
                assert len(rets) == len(input_grads), \
                    f"{self.__class__.name}.backward must return exactly the same number " \
                    "of NDArrays as the number of NDArrays arguments to forward." \
                    f"Expecting {len(input_grads)} got {len(rets)}"
                for igrad, ret, req in zip(input_grads, rets, reqs):
                    assert isinstance(ret, array_cls), \
                        f"autograd.Function.backward must return NDArrays, not {type(ret)}"
                    if req == 0:  # null
                        return True
                    elif req in (1, 2):  # write or inplace
                        igrad[:] = ret
                    elif req == 'add':
                        igrad[:] += ret
            except Exception:  # pylint: disable=broad-except
                print(f'Error in Function.backward: {traceback.format_exc()}')
                return False
            return True

        def delete_entry(_):
            """C Callback for CustomFunction::delete"""
            try:
                del Function._registry.ref_holder[key]
            except Exception:  # pylint: disable=broad-except
                print(f'Error in autograd.Function.delete: {traceback.format_exc()}')
                return False
            return True

        callbacks = [Function._bwd_functype(backward_entry),
                     Function._del_functype(delete_entry)]
        callbacks = [cast(i, CFUNCTYPE(c_int)) for i in callbacks]
        context = MXCallbackList(c_int(len(callbacks)),
                                 cast(c_array(CFUNCTYPE(c_int), callbacks),
                                      POINTER(CFUNCTYPE(c_int))),
                                 cast(c_array(c_void_p, [None]*len(callbacks)),
                                      POINTER(c_void_p)))
        Function._registry.ref_holder[key] = context
        check_call(_LIB.MXCustomFunctionRecord(
            c_int(len(inputs)),
            c_handle_array(inputs),
            c_int(len(outputs)),
            c_handle_array(outputs),
            ctypes.byref(context)))

        return ret_outputs

    def forward(self, *inputs):
        """Forward computation."""
        raise NotImplementedError

    def backward(self, *output_grads):
        """Backward computation.

        Takes as many inputs as forward's outputs,
        and returns as many NDArrays as forward's inputs.
        """
        raise NotImplementedError