python-mdp 3.3-2 / usr / lib / python2.7 / dist-packages / bimdp / binode.py

"""
Special BiNode class derived from Node to allow complicated flow patterns.

Messages:
=========

The message argument 'msg' of the outer method 'execute' or 'train' is either a
dict or None (which is treated equivalently to an empty dict and is the default
value). The message is automatically parsed against the method signature of
_train or _execute (or any other specified method) in the following way:

    normal key string -- Is copied if in signature and passed as a named
        argument.

    node_id->key -- Is extracted (i.e. removed in original message) and passed
        as a named argument. The separator '->' is also stored available
        as the constant MSG_ID_SEP. If the key is not an argument of the
        message then the whole key is simply erased.

The msg returned from the inner part of the method (e.g. _execute) is then used
to update the original message (so values can be overwritten).

If args without default value are missing in the message, this will result in
the standard Python missing-arguments-exception (this is not checked by BiNode
itself).


BiNode Return Value Options:
============================

 result for execute:
     x, (x, msg), (x, msg, target)

 result for train:
    None -- terminates training
    x, (x, msg), (x, msg, target) -- Execution is continued and
        this node will be reached at a later time to terminate training.
        If the result has the form (None, msg) then the msg is dropped (so
        it is not required to 'clear' the message manually).

 result for stop_training:
    None -- Simply terminates the training, like for a normal node.
    x, (x, msg), (x, msg, target) -- Causes an execute like phase, which
        terminates when the end of the flow is reached or when EXIT_TARGET
        is given as target value (just like during a normal execute phase).


Magic message keys:
===================

When the incoming message is parsed by the BiNode base class, some argument
keywords are treated in a special way:

 'msg' -- If any method like _execute accept a 'msg' keyword then the complete
     remaining message (after parsing the other keywords) is supplied. The
     message in the return value then completely replaces the original message
     (instead of only updating it). This way a node can completely control the
     message and for example remove keys.

 'target' -- If any template method like execute finds a 'target' keyword in
     the message then this is used as the target value in the return value.
     However, if _execute then also returns a target value this overwrites the
     target value.

  'method' -- Specify the name of the method that should be used instead of
      the standard one (e.g. in execute the standard method is _execute).
      An underscore is automatically added in front, so to select _execute
      one would have to provide 'execute'.
      If 'inverse' is given then the inverse dimension check will be performed
      and if no target is provided it will be set to -1.

"""

import inspect

import mdp

# separator for node_id in message keys
MSG_ID_SEP = "->"


class BiNodeException(mdp.NodeException):
    """Exception for BiNode problems."""
    pass


class BiNode(mdp.Node):
    """Abstract base class for nodes that use bimdp features.

    This class itself is not non-functional.

    Derived class should, if necessary, overwrite the _bi_reset method
    (in addition to the normal mdp.Node methods). 

    Note hat this class can also be used as an Adapter / Mixin for normal nodes.
    This can for example be useful for nodes which require additional data
    arguments during training or execution. These can then be encapsulated in a
    messsage. Note that BiNode has to come first in the MRO to make all this
    work.
    """

    def __init__(self, node_id=None, stop_result=None, **kwargs):
        """Initialize BiNode.

        node_id -- None or string which identifies the node.
        stop_result -- A (msg, target) tupple which is used by stop_training.
            If _stop_training returns a result as well then is updates /
            overwrites the stop_result, otherwise simply stop_result is
            returned (with x set to None).
            If the node has multiple training phases then stop_result must be
            None or an iterable with one entry for each training phase.

        kwargs are forwarded via super to the next __init__ method
        in the MRO.
        """
        self._node_id = node_id
        self._stop_result = stop_result
        self._coroutine_instances = None
        super(BiNode, self).__init__(**kwargs)

    ### Modified template methods from mdp.Node. ###

    def execute(self, x, msg=None):
        """Return single value y or a result tuple.

        x can be None, then the usual checks are omitted.

        The possible return types are y, (y, msg), (y, msg, target)
        The outgoing msg carries forward the incoming message content.
        The last entry in a result tuple must not be None.
        y can be None if the result is a tuple.

        This template method normally calls the corresponding _execute method
        or another method as specified in the message (using the magic 'method'
        key.
        """
        if msg is None:
            if x is None:
                err = "Both x and msg are None."
                raise BiNodeException(err)
            return super(BiNode, self).execute(x)
        msg_id_keys = self._get_msg_id_keys(msg)
        target = self._extract_message_key("target", msg, msg_id_keys)
        method_name = self._extract_message_key("method", msg, msg_id_keys)
        method, target = self._get_method(method_name, self._execute, target)
        msg, arg_dict = self._extract_method_args(method, msg, msg_id_keys)
        # perform specific checks
        if x is not None:
            if (not method_name) or (method_name == "execute"):
                self._pre_execution_checks(x)
                x = self._refcast(x)
            # testing for the actual method allows nodes to delegate method
            # resolution to internal nodes by manipulating _get_method
            elif method == self._inverse:
                self._pre_inversion_checks(x)
        result = method(x, **arg_dict)
        return self._combine_result(result, msg, target)

    def train(self, x, msg=None):
        """Train and return None or more if the execution should continue.

        The possible return types are None, y, (y, msg), (y, msg, target).
        The last entry in a result tuple must not be None.
        y can be None if the result is a tuple.

        This template method normally calls the corresponding _train method
        or another method as specified in the message (using the magic 'method'
        key.

        Note that the remaining msg and taret values are only used if _train
        (or the requested method) returns something different from None
        (so an empty dict can be used to trigger continued execution).
        """
        # perform checks, adapted from Node.train
        if not self.is_trainable():
            raise mdp.IsNotTrainableException("This node is not trainable.")
        if not self.is_training():
            err = "The training phase has already finished."
            raise mdp.TrainingFinishedException(err)
        if msg is None:
            if x is None:
                err = "Both x and msg are None."
                raise BiNodeException(err)
            # no fall-back on Node.train because we might have a return value
            self._check_input(x)
            try:
                self._check_train_args(x)
            except TypeError:
                err = ("%s training seems to require " % str(self) +
                       "additional arguments, but none were given.")
                raise BiNodeException(err)
            self._train_phase_started = True
            x = self._refcast(x)
            return self._train_seq[self._train_phase][0](x)
        msg_id_keys = self._get_msg_id_keys(msg)
        target = self._extract_message_key("target", msg, msg_id_keys)
        method_name = self._extract_message_key("method", msg, msg_id_keys)
        default_method = self._train_seq[self._train_phase][0]
        method, target = self._get_method(method_name, default_method, target)
        msg, arg_dict = self._extract_method_args(method, msg, msg_id_keys)
        # perform specific checks
        if x is not None:
            if (not method_name) or (method_name == "train"):
                self._check_input(x)
                try:
                    self._check_train_args(x, **arg_dict)
                except TypeError:
                    err = ("The given additional arguments %s " %
                                str(arg_dict.keys()) +
                           "are not compatible with training %s." % str(self))
                    raise BiNodeException(err)
                self._train_phase_started = True
                x = self._refcast(x)
            elif method == self._inverse:
                self._pre_inversion_checks(x)
        result = method(x, **arg_dict)
        if result is None:
            return None
        result = self._combine_result(result, msg, target)
        if (isinstance(result, tuple) and len(result) == 2 and
            result[0] is None):
            # drop the remaining msg, so that no maual clearing is required
            return None
        return result

    def stop_training(self, msg=None):
        """Stop training phase and start an execute phase with a target.

        The possible return types are None, y, (y, msg), (y, msg, target).
        For None nothing more happens, the training phase ends like for a
        standard MDP node.
        If a return value is given then an excute phase is started.

        This template method normally calls a _stop_training method from
        self._train_seq.

        If a stop_result was given in __init__ then it is used but can be
        overwritten by the returned _stop_training result or by the 
        msg argument provided by the BiFlow.
        """
        # basic checks
        if self.is_training() and self._train_phase_started == False:
            raise mdp.TrainingException("The node has not been trained.")
        if not self.is_training():
            err = "The training phase has already finished."
            raise mdp.TrainingFinishedException(err)
        # call stop_training
        if not msg:
            result = self._train_seq[self._train_phase][1]()
            target = None
        else:
            msg_id_keys = self._get_msg_id_keys(msg)
            target = self._extract_message_key("target", msg, msg_id_keys)
            method_name = self._extract_message_key("method", msg, msg_id_keys)
            default_method = self._train_seq[self._train_phase][1]
            method, target = self._get_method(method_name,
                                              default_method, target)
            msg, arg_dict = self._extract_method_args(method, msg, msg_id_keys)
            result = method(**arg_dict)
        # close the current phase
        self._train_phase += 1
        self._train_phase_started = False
        # check if we have some training phase left
        if self.get_remaining_train_phase() == 0:
            self._training = False
        # use stored stop message and update it with the result
        if self._stop_result:
            if self.has_multiple_training_phases():
                stored_stop_result = self._stop_result[self._train_phase - 1]
            else:
                stored_stop_result = self._stop_result
            # make sure that the original dict in stored_stop_result is not
            # modified (this could have unexpected consequences in some cases)
            stored_msg = stored_stop_result[0].copy()
            if msg:
                stored_msg.update(msg)
            msg = stored_msg
            if target is None:
                target = stored_stop_result[1]
        return self._combine_result(result, msg, target)

    ## Additional new methods. ##

    @property
    def node_id(self):
        """Return the node id (should be string) or None."""
        return self._node_id

    def bi_reset(self):
        """Reset the node for the next data chunck.

        This template method calls the _bi_reset method.

        This method is automatically called by BiFlow after the processing of
        a data chunk is completed (during both training and execution).

        All temporary data should be deleted. The internal node structure can
        be reset for the next data chunk. This is especially important if this
        node is called multiple times for a single chunk and an internal state
        keeps track of the actions to be performed for each call.
        """
        if self._coroutine_instances is not None:
            # delete the instance attributes to unshadow the coroutine
            # initialization methods
            for key in self._coroutine_instances:
                delattr(self, key)
            self._coroutine_instances = None
        self._bi_reset()

    def _bi_reset(self):
        """Hook method, overwrite when needed."""
        pass

    def _request_node_id(self, node_id):
        """Return the node if it matches the provided node id.

        Otherwise the return value is None. In this default implementation
        self is returned if node_id == self._node_id.

        Use this method instead of directly accessing self._node_id.
        This allows a node to be associated with multiple node_ids. Otherwise
        node_ids would not work for container nodes like BiFlowNode.
        """
        if self._node_id == node_id:
            return self
        else:
            return None

    ### Helper methods for msg handling. ###

    def _get_msg_id_keys(self, msg):
        """Return the id specific message keys for this node.

        The format is [(key, fullkey),...].
        """
        msg_id_keys = []
        for fullkey in msg:
            if fullkey.find(MSG_ID_SEP) > 0:
                node_id, key = fullkey.split(MSG_ID_SEP)
                if node_id == self._node_id:
                    msg_id_keys.append((key, fullkey))
        return msg_id_keys

    @staticmethod
    def _extract_message_key(key, msg, msg_id_keys):
        """Extract and return the requested key from the message.

        Note that msg and msg_id_keys are modfied if the found key was
        node_id specific.
        """
        value = None
        if key in msg:
            value = msg[key]
        # check for node_id specific key and remove it from the msg
        for i, (_key, _fullkey) in enumerate(msg_id_keys):
            if key == _key:
                value = msg.pop(_fullkey)
                msg_id_keys.pop(i)
                break
        return value

    @staticmethod
    def _extract_method_args(method, msg, msg_id_keys):
        """Extract the method arguments form the message.

        Return the new message and a dict with the keyword arguments (the
        return of the message is done because it can be set to None).
        """
        arg_keys = inspect.getargspec(method)[0]
        arg_dict = dict((key, msg[key]) for key in msg if key in arg_keys)
        for key, fullkey in msg_id_keys:
            if key in arg_keys:
                arg_dict[key] = msg.pop(fullkey)
            else:
                del msg[fullkey]
        if "msg" in arg_keys:
            arg_dict["msg"] = msg
            msg = None
        return msg, arg_dict

    def _get_method(self, method_name, default_method, target=None):
        """Return the method to be called and the target return value.

        method_name -- as provided in msg (without underscore)
        default_method -- bound method object
        target -- return target value as provided in message or None

        If the chosen method is _inverse then the default target is -1.
        """
        if not method_name:
            method = default_method
        elif method_name == "inverse":
            method = self._inverse
            if target is None:
                target = -1
        else:
            method_name = "_" + method_name
            try:
                method = getattr(self, method_name)
            except AttributeError:
                err = ("The message requested a method named '%s', but "
                       "there is no such method." % method_name)
                raise BiNodeException(err)
        return method, target

    @staticmethod
    def _combine_result(result, msg, target):
        """Combine the execution result with the provided values.

        result -- x, (x, msg) or (x, msg, target)

        The values in result always has priority.
        """
        # overwrite result values if necessary and return
        if isinstance(result, tuple):
            if msg:
                if result[1]:
                    # combine outgoing msg and remaining msg values
                    msg.update(result[1])
                result = (result[0], msg) + result[2:]
            if (target is not None) and (len(result) == 2):
                # use given target if no target value was returned
                result += (target,)
            return result
        else:
            # result is only single array
            if (not msg) and (target is None):
                return result
            elif target is None:
                return result, msg
            else:
                return result, msg, target

    ### Overwrite Special Methods ###

    def __repr__(self):
        """BiNode version of the Node representation, adding the node_id."""
        name = type(self).__name__
        inp = "input_dim=%s" % str(self.input_dim)
        out = "output_dim=%s" % str(self.output_dim)
        if self.dtype is None:
            typ = 'dtype=None'
        else:
            typ = "dtype='%s'" % self.dtype.name
        node_id = self.node_id
        if node_id is None:
            nid = 'node_id=None'
        else:
            nid = 'node_id="%s"' % node_id
        args = ', '.join((inp, out, typ, nid))
        return name + '(' + args + ')'

    def __add__(self, other):
        """Adding binodes returns a BiFlow.

        If a normal Node or Flow is added to a BiNode then a BiFlow is
        returned.
        Note that if a flow is added then a deep copy is used (deep
        copies of the nodes are used).
        """
        # unfortunately the inline imports are required to avoid
        # a cyclic import (unless one adds a helper function somewhere else)
        if isinstance(other, mdp.Node):
            import bimdp
            return bimdp.BiFlow([self, other])
        elif isinstance(other, mdp.Flow):
            flow_copy = other.copy()
            import bimdp
            biflow = bimdp.BiFlow([self.copy()] + flow_copy.flow)
            return biflow
        else:
            # can delegate old cases
            return super(BiNode, self).__add__(other)


class PreserveDimBiNode(BiNode, mdp.PreserveDimNode):
    """BiNode version of the PreserveDimNode."""
    pass


### Helper Functions / Decorators ###

def binode_coroutine(args=None, defaults=()):
    """Decorator for the convenient definition of BiNode couroutines.
    
    This decorator takes care of all the boilerplate code to use a coroutine
    as a BiNode method for continuations (which is more elegant and convenient
    than using a a state machine implementation). 

    args -- List of string names of the additional arguments. Note that the
        standard 'x' array is always given as the first value. So if n args
        are requested the yield will return n+1 values.
    defaults -- Tuple of default values for the arguments. If this tuple has
        n elements, they correspond to the last n elements in 'args'
        (following the convention of inspect.getargspec).
    
    Internally there are three methods/functions:
        - The user defined function containing the original coroutine code.
          This is only stored in the decorator closure.
        - A new method ('_coroutine_initialization') with the name and
          signature  of the decorated coroutine, which internally handles the
          first initialization of the coroutine instance.
          This method is returned by the decorator.
        - A method with the signature specified by the 'args' for the
          decorator.
          After the coroutine has been initialized this
          method shadows the initialization method in the class instance
          (using an instance attribute to shadow the class attribute).
    """
    if args is None:
        args = ["self", "x"]
    else:
        args = ["self", "x"] + args
    def _binode_coroutine(coroutine):
        # the original coroutine is only stored in this closure
        infodict = mdp.NodeMetaclass._function_infodict(coroutine)
        original_name = infodict["name"]
        ## create the coroutine interface method
        def _coroutine_interface(self, *args):
            try:
                return self._coroutine_instances[original_name].send(args)
            except StopIteration, exception:
                delattr(self, original_name)
                del self._coroutine_instances[original_name]
                if len(exception.args):
                    return exception.args
                else:
                    return None
        # turn the signature into the one specified by the args
        interface_infodict = infodict.copy()
        interface_infodict["signature"] = ", ".join(args)
        interface_infodict["defaults"] = defaults
        coroutine_interface = mdp.NodeMetaclass._wrap_function(
                                    _coroutine_interface, interface_infodict)
        ## create the initialization method
        def _coroutine_initialization(self, *args):
            coroutine_instance = coroutine(self, *args)
            bound_coroutine_interface = coroutine_interface.__get__(
                                                        self, self.__class__)
            if self._coroutine_instances is None:
                self._coroutine_instances = dict()
            self._coroutine_instances[original_name] = coroutine_instance
            setattr(self, original_name, bound_coroutine_interface)
            try:
                return coroutine_instance.next()
            except StopIteration, exception:
                delattr(self, original_name)
                del self._coroutine_instances[original_name]
                if len(exception.args):
                    return exception.args
                else:
                    return None
        coroutine_initialization = mdp.NodeMetaclass._wrap_function(
                                    _coroutine_initialization, infodict)
        return coroutine_initialization
    return _binode_coroutine