"""
BiMDP Flow class for flexible (bidirectional) data flow.
The central class is a BiFlow, which implements all the flow handling options
offered by the BiNode class (see binode.py for a description).
"""
import itertools
import mdp
n = mdp.numx
from binode import BiNode
# this target value tells the flow to abort and return the current values
EXIT_TARGET = "exit"
class NoneIterable(object):
"""Iterable for an infinite sequence of Nones."""
def __iter__(self):
while True:
yield None
class BiFlowException(mdp.FlowException):
"""Exception for BiFlow problems."""
pass
class MessageResultContainer(object):
"""Store and combine msg output chunks from a BiNode.
It is for example used when the flow execution yields msg output, which has
to be joined for the end result.
"""
def __init__(self):
"""Initialize the internal storage variables."""
self._msg_results = dict() # all none array message results
self._msg_array_results = dict() # result dict for arrays
def add_message(self, msg):
"""Add a single msg result to the combined results.
msg must be either a dict of results or None. numpy arrays will be
transformed to a single numpy array in the end. For all other types the
addition operator will be used to combine results (i.e., lists will be
appended, single integers will be summed over).
"""
if msg:
for key in msg:
if type(msg[key]) is n.ndarray:
if key not in self._msg_array_results:
self._msg_array_results[key] = []
self._msg_array_results[key].append(msg[key])
else:
if key not in self._msg_results:
self._msg_results[key] = msg[key]
else:
try:
self._msg_results[key] += msg[key]
except:
err = ("Could not combine final msg results "
"in BiFlow.")
raise BiFlowException(err)
def get_message(self):
"""Return the msg which combines all the msg results."""
# move array results from _msg_array_results to _msg_results
for key in self._msg_array_results:
if key in self._msg_results:
err = ("A key in the msg results is used with "
"different data types.")
raise BiFlowException(err)
else:
self._msg_results[key] = n.concatenate(
self._msg_array_results[key])
if self._msg_results:
return self._msg_results
else:
return None
# TODO: make sure that isinstance(str, target) is never used, so that in
# principle any object could be used (the might have to overwrite the
# _node_with_id method?
class BiFlow(mdp.Flow):
"""BiMDP version of a flow, which supports jumps between nodes.
This capabilities can be used by classes derived from BiNode.
Normal nodes can also be used in this flow, the msg argument is skipped
for these. Normal nodes can be also jump targets, but only when a relative
target index is used (since they do not support node ids).
"""
def __init__(self, flow, verbose=False, **kwargs):
kwargs["crash_recovery"] = False
super(BiFlow, self).__init__(flow=flow, verbose=verbose, **kwargs)
### Basic Methods from Flow. ###
def train(self, data_iterables, msg_iterables=None,
stop_messages=None):
"""Train the nodes in the flow.
The nodes will be trained according to their place in the flow.
data_iterables -- Sequence of iterables with the training data for each
trainable node. Can also be a single array or None.
Note that iterables yielding tuples for additonal node arguments
(e.g. the class labels for an FDANode) are not supported in a
BiFlow. Instead use the BiNode version of the node and provide
the arguments in the message (via msg_iterables).
msg_iterables -- Sequence of iterables with the msg training data
for each trainable node.
stop_messages -- Sequence of messages for stop_training.
Note that the type and iterator length of the data iterables is taken
as reference, so the message iterables are assumed to have the
same length.
"""
# Note: When this method is updated BiCheckpointFlow should be updated
# as well.
self._bi_reset() # normaly not required, just for safety
data_iterables, msg_iterables = self._sanitize_training_iterables(
data_iterables=data_iterables,
msg_iterables=msg_iterables)
if stop_messages is None:
stop_messages = [None] * len(data_iterables)
# train each Node successively
for i_node in range(len(self.flow)):
if self.verbose:
print ("training node #%d (%s)" %
(i_node, str(self.flow[i_node])))
self._train_node(data_iterables[i_node], i_node,
msg_iterables[i_node], stop_messages[i_node])
if self.verbose:
print "training finished"
def _train_node(self, iterable, nodenr, msg_iterable=None,
stop_msg=None):
"""Train a particular node.
nodenr -- index of the node to be trained
msg_iterable -- optional msg data for the training
Note that the msg is only passed to the Node if it is an instance
of BiNode.
stop_msg -- optional msg data for stop_training
Note that the message is only passed to the Node if the msg is not
None, so for a normal node the msg has to be None.
Note: unlike the normal mdp.Flow we do no exception handling here.
"""
if not self.flow[nodenr].is_trainable():
return
iterable, msg_iterable, _ = self._sanitize_iterables(iterable,
msg_iterable)
while True:
if not self.flow[nodenr].get_remaining_train_phase():
break
self._train_node_single_phase(iterable, nodenr,
msg_iterable, stop_msg)
self._bi_reset()
def _train_node_single_phase(self, iterable, nodenr,
msg_iterable, stop_msg=None):
"""Perform a single training phase for a given node.
This method should be only called internally in BiFlow.
"""
empty_iterator = True
for (x, msg) in itertools.izip(iterable, msg_iterable):
empty_iterator = False
## execute the flow until the nodes return value is right
i_node = 0
while True:
result = self._execute_seq(x, msg, i_node=i_node,
stop_at_node=nodenr)
## check the execution result, target should be True
if (not isinstance(result, tuple)) or (len(result) != 3):
err = ("The Node to be trained was not reached " +
"during training, last result: " + str(result))
raise BiFlowException(err)
elif result[2] is True:
x = result[0]
msg = result[1]
else:
err = ("Target node not found in flow during " +
"training, last target value: " + str(result[2]))
raise BiFlowException(err)
## perform node training
if isinstance(self.flow[nodenr], BiNode):
result = self.flow[nodenr].train(x, msg)
if result is None:
break # training is done for this chunk
else:
try:
self.flow[nodenr].train(x)
except TypeError:
# check if error is caused by additional node arguments
train_arg_keys = self._get_required_train_args(
self.flow[nodenr])
if len(train_arg_keys):
err = ("The node '%s' " % str(self.flow[nodenr]) +
"requires additional training " +
" arguments, which is not supported in a " +
"BiFlow. Instead use the BiNode version " +
"of the node and put the arguments in " +
"the msg.")
raise BiFlowException(err)
else:
raise
break
## training execution continues, interpret result
if not isinstance(result, tuple):
x = result
msg = None
target = None
elif len(result) == 2:
x, msg = result
target = None
elif len(result) == 3:
x, msg, target = result
else:
err = ("Node produced invalid return value " +
"during training: " + str(result))
raise BiFlowException(err)
i_node = self._target_to_index(target, nodenr)
self._bi_reset()
if empty_iterator:
if self.flow[nodenr].get_current_train_phase() == 1:
err_str = ("The training data iteration for node "
"no. %d could not be repeated for the "
"second training phase, you probably "
"provided an iterable instead of an "
"iterable." % (nodenr+1))
raise BiFlowException(err_str)
else:
err = ("The training data iterable for node "
"no. %d is empty." % (nodenr+1))
raise BiFlowException(err)
# unlike the normal mdp.Flow we always close the training
# otherwise stop_message propagation would be difficult
self._stop_training_hook()
if stop_msg is None:
result = self.flow[nodenr].stop_training()
else:
result = self.flow[nodenr].stop_training(stop_msg)
## deal with stop_message execution
if result is None:
return
elif not isinstance(result, tuple) and nodenr+1 < len(self.flow):
result = self._stop_message_seq(msg=result, i_node=nodenr+1)
elif len(result) == 2:
msg, target = result
if isinstance(target, int):
i_node = i_node + target
# values of +1 and -1 beyond this flow are tolerated
if i_node == len(self.flow) or i_node == -1:
return
else:
i_node = self._target_to_index(target, nodenr)
if isinstance(i_node, int):
result = self._stop_message_seq(msg=msg, i_node=i_node)
else:
err = ("Node produced invalid return value for " +
"stop_training: " + str(result))
raise BiFlowException(err)
# results without an explicit target are thrown away
if isinstance(result, tuple) and not isinstance(result[1], int):
err = ("Stop message target node not found in flow " +
"during training, last result: " +
str(result))
raise BiFlowException(err)
def execute(self, iterable, msg_iterable=None, target_iterable=None):
"""Execute the flow and return y or (y, msg).
iterable -- Can be an iterable or iterator for arrays, a single array
or None. In the last two cases it is assumed that msg is a single
message as well.
msg_iterable -- Can be an iterable or iterator or a single message
(but only if iterable is a single array or None).
target_iterable -- Like msg_iterable, but for target.
Note that the type and iteration length of iterable is taken as
reference, so msg is assumed to have the same length.
If msg results are found and if iteration is used then the BiFlow
tries to join the msg results (and concatenate in the case of arrays).
"""
self._bi_reset() # normaly not required, just for safety
iterable, msg_iterable, target_iterable = \
self._sanitize_iterables(iterable, msg_iterable, target_iterable)
y_results = None
msg_results = MessageResultContainer()
empty_iterator = True
for (x, msg, target) in itertools.izip(iterable, msg_iterable,
target_iterable):
empty_iterator = False
if not target:
i_node = 0
else:
i_node = self._target_to_index(target)
result = self._execute_seq(x=x, msg=msg, i_node=i_node)
if not isinstance(result, tuple):
y = result
msg = None
elif (len(result) == 2):
y, msg = result
elif (len(result) == 3) and (result[2] in [EXIT_TARGET, -1]):
# target -1 is allowed for easier inverse handling
y, msg = result[:2]
elif len(result) == 3:
err = ("Target node not found in flow during execute," +
" last result: " + str(result))
raise BiFlowException(err)
else:
err = ("BiNode execution returned invalid result type: " +
result)
raise BiFlowException(err)
self._bi_reset()
if msg:
msg_results.add_message(msg)
# check if all y have the same type and store it
# note that the checks for msg are less restrictive
if y is not None:
if y_results is None:
y_results = [y]
elif y_results is False:
err = "Some but not all y return values were None."
raise BiFlowException(err)
else:
y_results.append(y)
else:
if y_results is None:
y_results = False
else:
err = "Some but not all y return values were None."
raise BiFlowException(err)
if empty_iterator:
err = ("The execute data iterable is empty.")
raise BiFlowException(err)
# consolidate results
if y_results:
y_results = n.concatenate(y_results)
result_msg = msg_results.get_message()
if result_msg:
return (y_results, result_msg)
else:
return y_results
def __call__(self, iterable, msg_iterable=None):
"""Calling an instance is equivalent to call its 'execute' method."""
return self.execute(iterable, msg_iterable=msg_iterable)
### New Methods for BiMDP. ###
def _bi_reset(self):
"""Reset the nodes and internal flow variables."""
self._global_message_emitter = None
for node in self.flow:
if isinstance(node, BiNode):
node.bi_reset()
def _request_node_id(self, node_id):
"""Return first hit of _request_node_id on internal nodes.
So _request_node_id is called for all nodes in the flow until a return
value is not None. If no such node is found the return value is None.
"""
for node in self.flow:
if isinstance(node, BiNode):
found_node = node._request_node_id(node_id)
if found_node:
return found_node
return None
## container special methods to support node_id
def __getitem__(self, key):
if isinstance(key, str):
item = self._request_node_id(key)
if item is None:
err = ("This biflow contains no node with with the id " +
str(key))
raise KeyError(err)
return item
else:
return super(BiFlow, self).__getitem__(key)
def __setitem__(self, key, value):
if isinstance(key, str):
err = "Setting nodes by node_id is not supported."
raise BiFlowException(err)
else:
super(BiFlow, self).__setitem__(key, value)
def __delitem__(self, key):
if isinstance(key, str):
err = "Deleting nodes by node_id is not supported."
raise BiFlowException(err)
else:
super(BiFlow, self).__delitem__(key)
def __contains__(self, key):
if isinstance(key, str):
if self._request_node_id(key) is not None:
return True
else:
return False
else:
return super(BiFlow, self).__contains__(key)
### Flow Implementation Methods ###
def _sanitize_training_iterables(self, data_iterables, msg_iterables):
"""Check and adjust the training iterable list."""
if data_iterables is None:
if msg_iterables is None:
err = ("Both the training data and the training messages are "
"None.")
raise BiFlowException(err)
else:
data_iterables = [None] * len(self.flow)
elif isinstance(data_iterables, n.ndarray):
data_iterables = [[data_iterables]] * len(self.flow)
# the form of msg_iterables follows that of data_iterables
msg_iterables = [[msg_iterables]] * len(data_iterables)
else:
data_iterables = self._train_check_iterables(data_iterables)
if msg_iterables is None:
msg_iterables = [None] * len(self.flow)
else:
msg_iterables = self._train_check_iterables(msg_iterables)
return data_iterables, msg_iterables
def _sanitize_iterables(self, iterable, msg_iterable, target_iterable=None):
"""Check and adjust a data, message and target iterable."""
# TODO: maybe add additional checks
if isinstance(iterable, n.ndarray):
iterable = [iterable]
msg_iterable = [msg_iterable]
target_iterable = [target_iterable]
elif iterable is None:
if msg_iterable is None:
err = "Both the data and the message iterable is None."
raise BiFlowException(err)
else:
iterable = NoneIterable()
if isinstance(msg_iterable, dict):
msg_iterable = [msg_iterable]
target_iterable = [target_iterable]
else:
if msg_iterable is None:
msg_iterable = NoneIterable()
if target_iterable is None:
target_iterable = NoneIterable()
return iterable, msg_iterable, target_iterable
def _target_to_index(self, target, current_node=0):
"""Return the absolute node index of the target code.
If the string id target node is not found in this flow then the string
is returned without alteration.
When a relative index is given it is translated to the absolute index
and it is checked if it is in the allowed range.
target -- Can be a string node id, a relative index or None (which
is interpreted as 1).
current_node -- If target is specified as a relative index then this
node index is used to translate the target to the absolute node
index (otherwise it has no effect).
check_bounds -- If False then it is not checked wether the node index
is in range(len(flow)).
"""
if target == EXIT_TARGET:
return EXIT_TARGET
if target is None:
target = 1
if not isinstance(target, int):
for i_node, node in enumerate(self.flow):
if isinstance(node, BiNode) and node._request_node_id(target):
return i_node
# no matching node was found
return target
else:
absolute_index = current_node + target
if absolute_index < 0:
err = "Target int value references node at position < 0."
raise BiFlowException(err)
elif absolute_index >= len(self.flow):
err = ("Target int value references a node"
" beyond the flow length (target " + str(target) +
", current node " + str(current_node) + ").")
raise BiFlowException(err)
return absolute_index
# TODO: update docstring for case when target is not found
def _execute_seq(self, x, msg=None, i_node=0, stop_at_node=None):
"""Execute the whole flow as far as possible.
i_node -- Can specify a node index where the excecution is supposed to
start.
stop_at_node -- Node index where the execution should stop. The input
values for this node are returned in this case in the form
(x, msg, target) with target being set to True.
If the end of the flow is reached then the return value is y
or (y, msg).
If the an execution target node is not found then (x, msg, target) is
returned (target values of 1 and -1 are also possible).
If a normal Node (not derived from BiNode) is encountered then the
current msg is simply carried forward around it.
"""
## this method is also used by other classes, like BiFlowNode
while i_node != stop_at_node:
if isinstance(self.flow[i_node], BiNode):
result = self.flow[i_node].execute(x, msg)
# check the type of the result
if type(result) is not tuple:
x = result
msg = None
target = 1
elif len(result) == 2:
x, msg = result
target = 1
elif len(result) == 3:
x, msg, target = result
else:
err = ("BiNode execution returned invalid result type: " +
result)
raise BiFlowException(err)
else:
# just a normal MDP node
x = self.flow[i_node].execute(x)
# note that the message is carried forward unchanged
target = 1
## check if the target is in this flow, return otherwise
if isinstance(target, int):
i_node = i_node + target
# values of +1 and -1 beyond this flow are allowed
if i_node == len(self.flow):
if not msg:
return x
else:
return (x, msg)
elif i_node == -1:
return x, msg, -1
else:
i_node = self._target_to_index(target, i_node)
if not isinstance(i_node, int):
# target not found in this flow
# this is also the exit point when EXIT_TARGET is given
return x, msg, target
# reached stop_at_node, signal this by returning target value True
return x, msg, True
def _stop_message_seq(self, msg, i_node):
"""Propagate a stop_message through the flow.
i_node -- Index of the node that is called first.
If the bi_stop_message terminated there is no return value. If a target
node was not found along the way the return value is msg or
(msg, target).
"""
## note that the code is almost identical to _message_seq
while True:
if not isinstance(self.flow[i_node], BiNode):
err = ("Called stop_message on a normal node, target was: " +
str(self.flow[i_node]))
raise BiFlowException(err)
result = self.flow[i_node].stop_message(msg)
# check the type of the result
if result is None:
# reached end of message sequence
return None
elif not isinstance(result, tuple):
return None
elif len(result) == 2:
msg, target = result
else:
err = ("Node stop_message produced invalid return value " +
"during training: " + str(result))
raise BiFlowException(err)
## check if the target is in this flow, return otherwise
if isinstance(target, int):
i_node = i_node + target
# values of +1 and -1 beyond this flow are allowed
if i_node == len(self.flow):
return msg, 1
elif i_node == -1:
return msg, -1
else:
i_node = self._target_to_index(target, i_node)
if not isinstance(i_node, int):
# target not found in this flow
# this is also the exit point when EXIT_TARGET is given
return msg, target
### Some useful flow classes. ###
class BiCheckpointFlow(BiFlow, mdp.CheckpointFlow):
"""Similar to normal checkpoint flow.
The main difference is that even the last training phase of a
node is already closed before the checkpoint function is called (otherwise
it would be difficult to propagate bi_stop_message).
"""
def train(self, data_iterables, checkpoints, msg_iterables=None,
stop_messages=None):
"""Train the nodes in the flow.
The nodes will be trained according to their place in the flow.
Additionally calls the checkpoint function 'checkpoint[i]'
when the training phase of node #i is over.
A checkpoint function takes as its only argument the trained node.
If the checkpoint function returns a dictionary, its content is
added to the instance's dictionary.
The class CheckpointFunction can be used to define user-supplied
checkpoint functions.
"""
self._bi_reset() # normaly not required, just for safety
data_iterables, msg_iterables = self._sanitize_training_iterables(
data_iterables=data_iterables,
msg_iterables=msg_iterables)
if stop_messages is None:
stop_messages = [None] * len(data_iterables)
checkpoints = self._train_check_checkpoints(checkpoints)
# train each Node successively
for i_node in range(len(self.flow)):
if self.verbose:
print ("training node #%d (%s)" %
(i_node, str(self.flow[i_node])))
self._train_node(data_iterables[i_node], i_node,
msg_iterables[i_node], stop_messages[i_node])
if i_node <= len(checkpoints) and checkpoints[i_node] is not None:
checkpoint_dict = checkpoints[i_node](self.flow[i_node])
if dict:
self.__dict__.update(checkpoint_dict)
if self.verbose:
print "training finished"
|