Module tf.convert.walker
Walker
You can convert a dataset to TF by writing a function that walks through it.
That function must trigger a sequence of actions when reading the data. These actions drive TF to build a valid TF dataset. Many checks will be performed.
to and from MQL
If your source is MQL, you are even better off: TF has a
module to import from MQL and to export to MQL.
See importMQL() and exportMQL().
Set up
Here is a schematic set up of such a conversion program.
from tf.fabric import Fabric
from tf.convert.walker import CV
TF = Fabric(locations=OUT_DIR)
cv = CV(TF)
def director(cv):
# your code to unwrap your source data and trigger
# the generation of TF nodes, edges and features
slotType = 'word' # or whatever you choose
otext = { # dictionary of config data for sections and text formats
...
}
generic = { # dictionary of metadata meant for all features
...
}
intFeatures = { # set of integer valued feature names
...
}
featureMeta = { # per feature dictionaries with metadata
...
}
good = cv.walk(
director,
slotType,
otext=otext,
generic=generic,
intFeatures=intFeatures,
featureMeta=featureMeta,
warn=True,
force=False,
)
if good:
... load the new TF data ...
See CV.walk().
Walking
When you walk through the input data source, you'll encounter things that have to become slots, non-slot nodes, edges and features in the new data set.
You issue these things by means of an action method from cv, such as
cv.slot() or cv.node(nodeType).
When your action creates slots or non slot nodes, TF will return you a reference to that node, that you can use later for more actions related to that node.
curPara = cv.node('para')
To add features to nodes, use a cv.feature() action.
It will apply to a node passed as argument.
To add features to edges, issue a cv.edge() action.
It will require two node arguments: the from node and the to node.
There is always a set of current embedder nodes. When you create a slot node
curWord = cv.slot()
then TF will link all current embedder nodes to the resulting slot.
There are actions to add nodes to the set of embedder nodes, to remove them from it, and to add them again.
If your data is organized in such a way that you see the slots in a different order than the intended order, you can pass a key value to the slot, like so
curWord = cv.slot(key=wordNumber)
After the walk is completed, the slots will be sorted by their keys, while keeping all feature assignments to them intact. All nodes will be linked to the same slots after sorting, and all edges that start from or arrive at slots will do that after the sorting.
For an example, see lowfat.py.
Dynamic Metadata
When the director runs, you may have already specified all your feature metadata, including the value types.
But if some of that information is dependent on what you encounter in the data, you can do two things:
(A) Run a preliminary pass over the data and gather the required information, before running the director.
(B) Update the metadata later on
by issuing cv.meta() actions from within your director, see below.
In doing this, you can remove the metadata from unused features, but you can also add metadata to features that have not got them yet.
If the set of possible features is not known on beforehand,
you can ask for the list of all feature names by means of
cv.features().
Action methods
The cv class contains methods that are responsible for particular actions
that steer the graph building:
CV.slot()CV.node()CV.terminate()CV.resume()CV.link()CV.linked()CV.feature()CV.features()CV.edge()CV.meta()CV.occurs()CV.active()CV.activeNodes()CV.activeTypes()CV.get()andcv.get(feature, nf, nt)CV.stop()
Expand source code Browse git
"""
# Walker
You can convert a dataset to TF by writing a function that walks through it.
That function must trigger a sequence of actions when reading the data.
These actions drive TF to build a valid TF dataset.
Many checks will be performed.
!!! hint "to and from MQL"
If your source is MQL, you are even better off: TF has a
module to import from MQL and to export to MQL.
See `tf.convert.mql.importMQL` and `tf.convert.mql.exportMQL`.
## Set up
Here is a schematic set up of such a conversion program.
``` python
from tf.fabric import Fabric
from tf.convert.walker import CV
TF = Fabric(locations=OUT_DIR)
cv = CV(TF)
def director(cv):
# your code to unwrap your source data and trigger
# the generation of TF nodes, edges and features
slotType = 'word' # or whatever you choose
otext = { # dictionary of config data for sections and text formats
...
}
generic = { # dictionary of metadata meant for all features
...
}
intFeatures = { # set of integer valued feature names
...
}
featureMeta = { # per feature dictionaries with metadata
...
}
good = cv.walk(
director,
slotType,
otext=otext,
generic=generic,
intFeatures=intFeatures,
featureMeta=featureMeta,
warn=True,
force=False,
)
if good:
... load the new TF data ...
```
See `tf.convert.walker.CV.walk`.
## Walking
When you walk through the input data source, you'll encounter things
that have to become slots, non-slot nodes, edges and features in the new data set.
You issue these things by means of an *action method* from `cv`, such as
`cv.slot()` or `cv.node(nodeType)`.
When your action creates slots or non slot nodes,
TF will return you a reference to that node,
that you can use later for more actions related to that node.
``` python
curPara = cv.node('para')
```
To add features to nodes, use a `cv.feature()` action.
It will apply to a node passed as argument.
To add features to edges, issue a `cv.edge()` action.
It will require two node arguments: the *from* node and the *to* node.
There is always a set of current *embedder nodes*.
When you create a slot node
``` python
curWord = cv.slot()
```
then TF will link all current embedder nodes to the resulting slot.
There are actions to add nodes to the set of embedder nodes,
to remove them from it,
and to add them again.
If your data is organized in such a way that you see the slots in a different
order than the intended order, you can pass a key value to the slot, like so
``` python
curWord = cv.slot(key=wordNumber)
```
After the walk is completed, the slots will be sorted by their keys, while keeping
all feature assignments to them intact. All nodes will be linked to the same slots
after sorting, and all edges that start from or arrive at slots will do that after
the sorting.
For an example, see
[lowfat.py](https://github.com/ETCBC/nestle1904/blob/master/programs/lowfat.py).
## Dynamic Metadata
When the director runs, you may have already specified all your feature
metadata, including the value types.
But if some of that information is dependent on what you encounter in the
data, you can do two things:
(A) Run a preliminary pass over the data and gather the required information,
before running the director.
(B) Update the metadata later on
by issuing `cv.meta()` actions from within your director, see below.
In doing this, you can remove the metadata from unused features,
but you can also add metadata to features that have not got them yet.
If the set of possible features is not known on beforehand,
you can ask for the list of all feature names by means of
`cv.features()`.
## Action methods
The `cv` class contains methods that are responsible for particular *actions*
that steer the graph building:
* `tf.convert.walker.CV.slot`
* `tf.convert.walker.CV.node`
* `tf.convert.walker.CV.terminate`
* `tf.convert.walker.CV.resume`
* `tf.convert.walker.CV.link`
* `tf.convert.walker.CV.linked`
* `tf.convert.walker.CV.feature`
* `tf.convert.walker.CV.features`
* `tf.convert.walker.CV.edge`
* `tf.convert.walker.CV.meta`
* `tf.convert.walker.CV.occurs`
* `tf.convert.walker.CV.active`
* `tf.convert.walker.CV.activeNodes`
* `tf.convert.walker.CV.activeTypes`
* `tf.convert.walker.CV.get` and `cv.get(feature, nf, nt)`
* `tf.convert.walker.CV.stop`
!!! hint "Example"
Follow the
[conversion tutorial](https://nbviewer.jupyter.org/github/annotation/banks/blob/master/programs/convert.ipynb)
Or study a more involved example:
[Old Babylonian](https://github.com/Nino-cunei/tfFromAtf/blob/master/programs/convert.py)
"""
import collections
import functools
import re
from ..parameters import WARP, OTYPE, OSLOTS, OTEXT
from ..core.helpers import itemize, isInt
from ..core.timestamp import SILENT_D, DEEP, silentConvert
class CV:
S = "slot"
N = "node"
T = "terminate"
R = "resume"
D = "delete"
F = "feature"
E = "edge"
def __init__(self, TF, silent=SILENT_D):
"""The object that contains the walker conversion machinery.
silent: string, optional tf.core.timestamp.SILENT_D
See `tf.core.timestamp.Timestamp`
"""
self.TF = TF
self.silent = silent
tmObj = TF.tmObj
isSilent = tmObj.isSilent
setSilent = tmObj.setSilent
silent = silentConvert(silent)
self.wasSilent = isSilent()
setSilent(silent)
def _showWarnings(self):
silent = self.silent
tmObj = self.TF.tmObj
error = tmObj.error
info = tmObj.info
indent = tmObj.indent
warnings = self.warnings
warn = self.warn
if warn is None:
if warnings:
info("use `cv.walk(..., warn=False)` to make warnings visible")
info("use `cv.walk(..., warn=True)` to stop on warnings")
else:
method = error if warn else info
if warnings:
for (kind, msgs) in sorted(warnings.items()):
method(f"WARNING {kind} ({len(msgs)} x):", force=silent != DEEP)
indent(level=2)
for msg in sorted(set(msgs))[0:20]:
if msg:
method(f"{msg}", tm=False, force=silent != DEEP)
self.warnings = {}
if warn:
info("use `cv.walk(..., warn=False)` to continue after warnings")
info("use `cv.walk(..., warn=None)` to suppress warnings")
self.good = False
else:
info("use `cv.walk(..., warn=True)` to stop after warnings")
info("use `cv.walk(..., warn=None)` to suppress warnings")
def _showErrors(self):
silent = self.silent
tmObj = self.TF.tmObj
error = tmObj.error
info = tmObj.info
indent = tmObj.indent
forcedStop = self.forcedStop
errors = self.errors
if errors:
for (kind, msgs) in sorted(errors.items()):
error(f"ERROR {kind} ({len(msgs)} x):")
indent(level=2)
for msg in sorted(set(msgs))[0:20]:
if msg:
error(f"{msg}", tm=False)
self.errors = {}
self.good = False
if forcedStop:
error("STOPPED by the stop() instruction")
elif not errors:
if self.good:
info("OK", force=silent != DEEP)
else:
error("STOPPED because of warnings")
def walk(
self,
director,
slotType,
otext={},
generic={},
intFeatures=set(),
featureMeta={},
warn=True,
generateTf=True,
force=False,
):
"""Asks a director function to walk through source data and receives its actions.
The `director` function should unravel the source.
You have to program the `director`, which takes one argument: `cv`.
From the `cv` you can use a few standard actions that instruct TF
to build a graph.
This function will check whether the metadata makes sense and is minimally
complete.
During node creation the section structure will be watched,
and you will be warned if irregularities occur.
After the creation of the feature data, some extra checks will be performed
to see whether the metadata matches the data and vice versa.
If the slots need to be sorted by their keys, it will happen at this point,
and the generated features will be adapted to the sorted slots.
The new feature data will be written to the output directory of the
underlying TF object. In fact, the rules are exactly the same as for
`tf.fabric.Fabric.save`.
Parameters
----------
slotType: string
The node type that acts as the type of the slots in the data set.
oText: dict
The configuration information to be stored in the `otext` feature
(see `tf.core.text`):
* section types
* section features
* structure types
* structure features
* text formats
generic: dict
Metadata that will be written into the header of all generated TF features.
You can make changes to this later on, dynamically in your director.
intFeatures: iterable
The set of features that have integer values only.
You can make changes to this later on, dynamically in your director.
featureMeta: dict of dict
For each node or edge feature descriptive metadata can be supplied.
You can make changes to this later on, dynamically in your director.
warn: boolean, optional True
This regulates the response to warnings:
`True` (default): stop after warnings (as if they are errors);
`False` continue after warnings but do show them;
`None` suppress all warnings.
force: boolean, optional False
This forces the process to continue after errors.
Your TF set might not be valid.
Yet this can be useful during testing, when you know
that not everything is OK, but you want to check some results.
Especially when dealing with large datasets, you might want to test
with little pieces. But then you get a kind of non-fatal errors that
stand in the way of testing. For those cases: `force=True`.
generateTf: boolean, optional True
You can pass `False` here to suppress the actual writing of TF data.
In that way you can dry-run the director to check for errors and warnings
director: function
An ordinary function that takes one argument, the `cv` object, and
should not deliver anything.
Writing this function is the main job to do when you want to convert a data source
to TF.
See `tf.convert.walker` for more details.
Returns
-------
boolean
Whether the operation was successful
"""
tmObj = self.TF.tmObj
info = tmObj.info
indent = tmObj.indent
setSilent = tmObj.setSilent
silent = self.silent
indent(level=0, reset=True)
info("Importing data from walking through the source ...", force=silent != DEEP)
self.force = force
self.good = True
self.forcedStop = False
self.errors = collections.defaultdict(list)
self.warnings = collections.defaultdict(list)
self.warn = warn
self.slotType = slotType
self.intFeatures = set(intFeatures)
self.featureMeta = featureMeta
self.metaData = {}
self.nodeFeatures = {}
self.edgeFeatures = {}
self.slotKeys = {}
self.discardables = set()
indent(level=1, reset=True)
self._prepareMeta(otext, generic)
indent(level=1, reset=True)
self._follow(director)
indent(level=1, reset=True)
self._removeUnlinked()
indent(level=1, reset=True)
self._checkGraph()
indent(level=1, reset=True)
self._checkFeatures()
indent(level=1, reset=True)
self._reorderNodes()
indent(level=1, reset=True)
self._reassignFeatures()
indent(level=0)
info("Features ready to write")
if generateTf:
if self.good or self.force:
self.good = self.TF.save(
metaData=self.metaData,
nodeFeatures=self.nodeFeatures,
edgeFeatures=self.edgeFeatures,
silent=silent,
)
self._showWarnings()
setSilent(self.wasSilent)
return self.good
def _prepareMeta(self, otext, generic):
silent = self.silent
varRe = re.compile(r"\{([^}]+)\}")
tmObj = self.TF.tmObj
info = tmObj.info
indent = tmObj.indent
if not self.good and not self.force:
return
info("Preparing metadata... ", force=silent != DEEP)
intFeatures = self.intFeatures
featureMeta = self.featureMeta
errors = self.errors
self.metaData = {
"": generic,
OTYPE: {"valueType": "str"},
OSLOTS: {"valueType": "str"},
OTEXT: otext,
}
metaData = self.metaData
self.intFeatures = intFeatures
self.sectionTypes = []
self.sectionFeatures = []
self.sectionFromLevel = {}
self.levelFromSection = {}
self.structureTypes = []
self.structureFeatures = []
self.structureLevel = {}
self.textFormats = {}
self.textFeatures = set()
if not generic:
errors['Missing feature meta data in "generic"'].append(
"Consider adding provenance metadata to all features"
)
if not otext:
errors['Missing "otext" configuration'].append(
"Consider adding configuration for text representation and section levels"
)
else:
sectionInfo = {}
for f in ("sectionTypes", "sectionFeatures"):
if f not in otext:
errors['Incomplete section specs in "otext"'].append(
f'no key "{f}"'
)
sectionInfo[f] = []
else:
sFields = itemize(otext[f], sep=",")
sectionInfo[f] = sFields
if f == "sectionTypes":
for (i, s) in enumerate(sFields):
self.levelFromSection[s] = i + 1
self.sectionFromLevel[i + 1] = s
sLevels = {f: len(sectionInfo[f]) for f in sectionInfo}
if min(sLevels.values()) != max(sLevels.values()):
errors["Inconsistent section info"].append(
" but ".join(f'"{f}" has {sLevels[f]} levels' for f in sLevels)
)
self.sectionFeatures = sectionInfo["sectionFeatures"]
self.sectionTypes = sectionInfo["sectionTypes"]
self.featFromSectionType = {
typ: feat
for (typ, feat) in zip(self.sectionTypes, self.sectionFeatures)
}
self.sectionSet = set(self.sectionTypes)
structureInfo = {}
for f in ("structureTypes", "structureFeatures"):
if f not in otext:
structureInfo[f] = []
continue
sFields = itemize(otext[f], sep=",")
structureInfo[f] = sFields
if not structureInfo:
info("No structure definition found in otext")
sLevels = {f: len(structureInfo[f]) for f in structureInfo}
if min(sLevels.values()) != max(sLevels.values()):
errors["Inconsistent structure info"].append(
" but ".join(f'"{f}" has {sLevels[f]} levels' for f in sLevels)
)
structureInfo = {}
if not structureInfo or all(
len(info) == 0 for (s, info) in structureInfo.items()
):
info("No structure nodes will be set up")
self.structureFeatures = []
self.structureTypes = []
self.structureFeatures = structureInfo["structureFeatures"]
self.structureTypes = structureInfo["structureTypes"]
self.featFromStructureType = {
typ: feat
for (typ, feat) in zip(self.structureTypes, self.structureFeatures)
}
self.structureSet = set(self.structureTypes)
textFormats = {}
textFeatures = set()
for (k, v) in sorted(otext.items()):
if k.startswith("fmt:"):
featureSet = set()
features = varRe.findall(v)
for ff in features:
fr = ff.rsplit(":", maxsplit=1)[0]
for f in fr.split("/"):
featureSet.add(f)
textFormats[k[4:]] = featureSet
textFeatures |= featureSet
if not textFormats:
errors['No text formats in "otext"'].append('add "fmt:text-orig-full"')
elif "text-orig-full" not in textFormats:
errors["No default text format in otext"].append(
'add "fmt:text-orig-full"'
)
self.textFormats = textFormats
self.textFeatures = textFeatures
info(f'SECTION TYPES: {", ".join(self.sectionTypes)}', tm=False)
info(f'SECTION FEATURES: {", ".join(self.sectionFeatures)}', tm=False)
info(f'STRUCTURE TYPES: {", ".join(self.structureTypes)}', tm=False)
info(f'STRUCTURE FEATURES: {", ".join(self.structureFeatures)}', tm=False)
info("TEXT FEATURES:", tm=False)
indent(level=2)
for (fmt, feats) in sorted(textFormats.items()):
info(f'{fmt:<20} {", ".join(sorted(feats))}', tm=False)
indent(level=1)
for feat in WARP + ("",):
if feat in intFeatures:
if feat == "":
errors["intFeatures"].append(
'Do not declare the "valueType" for all features'
)
else:
errors["intFeatures"].append(
f'Do not mark the "{feat}" feature as integer valued'
)
self.good = False
for (feat, featMeta) in sorted(featureMeta.items()):
good = self._checkFeatMeta(
feat,
featMeta,
checkRegular=True,
valueTypeAllowed=False,
showErrors=False,
)
if not good:
self.good = False
metaData.setdefault(feat, {}).update(featMeta)
metaData[feat]["valueType"] = "int" if feat in intFeatures else "str"
self._showErrors()
def _checkFeatMeta(
self,
feat,
featMeta,
checkRegular=False,
valueTypeAllowed=True,
showErrors=True,
):
errors = collections.defaultdict(list)
good = True
if checkRegular:
if feat in WARP + ("",):
if feat == "":
errors["featureMeta"].append(
'Specify the generic feature meta data in "generic"'
)
good = False
elif feat == OTEXT:
errors["featureMeta"].append(
f'Specify the "{OTEXT}" feature in "otext"'
)
good = False
else:
errors["featureMeta"].append(
f'Do not pass metaData for the "{feat}" feature in "featureMeta"'
)
good = False
if "valueType" in featMeta:
if not valueTypeAllowed:
errors["featureMeta"].append(
f'Do not specify "valueType" for the "{feat}" feature in "featureMeta"'
)
good = False
elif featMeta["valueType"] not in {"int", "str"}:
errors["featureMeta"].append('valueType must be "int" or "str"')
good = False
for (e, eData) in errors.items():
self.errors[e].extend(eData)
if showErrors:
self._showErrors
return good
def stop(self, msg):
"""Stops the director. No further input will be read.
```
cv.stop(msg)
```
The director will exit with a non-good status and issue the message `msg`.
If you have called `walk()` with `force=True`, indicating that the
director must proceed after errors, then this stop command will cause
termination nevertheless.
Parameters
----------
msg: string
A message to display upon stopping.
Returns
-------
None
"""
tmObj = self.TF.tmObj
error = tmObj.error
error(f"Forced stop: {msg}")
self.good = False
self.force = False
self.forcedStop = True
def slot(self, key=None):
"""Makes a slot node and return the handle to it in `n`.
```
n = cv.slot()
```
No further information is needed.
Remember that you can add features to the node by later
```
cv.feature(n, key=value, ...)
```
calls.
Parameters
----------
key: string, optional None
If passed, it acts as a sort key on the slot.
At the end of the walk, all slots will be sorted by their key and then
by their original order. Care will be taken that slots retain their
features and linkages.
!!! note "Keys are strings"
Note that the key must be a string. If you want to sort on numbers,
make sure to pad all numbers with leading zeros.
Returns
-------
node reference: tuple
The node reference consists of a node type and a sequence number,
but normally you do not have to dig these out.
Just pass the tuple as a whole to actions that require a node argument.
"""
curSeq = self.curSeq
curEmbedders = self.curEmbedders
oslots = self.oslots
slotKeys = self.slotKeys
levelFromSection = self.levelFromSection
warnings = self.warnings
self.stats[self.S] += 1
nType = self.slotType
curSeq[nType] += 1
seq = curSeq[nType]
if key is not None:
slotKeys[seq] = key
inSection = False
for eNode in curEmbedders:
if eNode[0] in levelFromSection:
inSection = True
oslots[eNode].add(seq)
if levelFromSection and not inSection:
warnings["slot outside sections"].append(f"{seq}")
return (nType, seq)
def node(self, nType, slots=None):
"""Makes a non-slot node and return the handle to it in `n`.
```
n = cv.node(nodeType)
```
You have to pass its *node type*, i.e. a string.
Think of `sentence`, `paragraph`, `phrase`, `word`, `sign`, whatever.
There are two modes for this function:
* Auto: (`slots=None`):
Non slot nodes will be automatically added to the set of embedders.
* Explicit: (`slots=iterable`):
The slots in iterable will be assigned to this node and nothing else.
The node will not be added to the set of embedders.
Put otherwise: the node will be terminated after construction.
However: you could resume it later to add other slots.
Remember that you can add features to the node by later
```
cv.feature(n, key=value, ...)
```
calls.
Parameters
----------
nType: string
A node type, not the slot type
slots: iterable of integer, optional None
The slots to assign to this node.
If left out, the node is left as an embedding node and
subsequent slots will be added to it automatically.
All slots in the iterable must have been generated before
by means of the `cv.slot()` action.
Returns
-------
node reference or None
If an error occurred, `None` is returned.
The node reference consists of a node type and a sequence number,
but normally you do not have to dig these out.
Just pass the tuple as a whole to actions that require a node argument.
"""
slotType = self.slotType
errors = self.errors
if nType == slotType:
errors[f'use `cv.slot()` instead of `cv.node("{nType}")`'].append(None)
return
curSeq = self.curSeq
curEmbedders = self.curEmbedders
self.stats[self.N] += 1
curSeq[nType] += 1
seq = curSeq[nType]
node = (nType, seq)
self._checkSecLevel(node, before=True)
if slots:
maxSlot = curSeq[slotType]
for s in slots:
if not 1 <= s <= maxSlot:
errors[f"slot out of range in `cv.node(({nType}, {seq}))`"].append(
f"{s}"
)
else:
oslots = self.oslots
oslots[node].add(s)
self.stats[self.T] += 1
else:
curEmbedders.add(node)
return node
def terminate(self, node):
"""**terminates** a node.
```
cv.terminate(n)
```
The node `n` will be removed from the set of current embedders.
This `n` must be the result of a previous `cv.slot()` or `cv.node()` action.
Parameters
----------
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
Returns
-------
None
"""
self.stats[self.T] += 1
if node is not None:
self.curEmbedders.discard(node)
self._checkSecLevel(node, before=False)
def delete(self, node):
"""**deletes** a node.
```
cv.delete(n)
```
The node `n` will be deleted from the set of nodes that will be created.
This `n` must be the result of a previous `cv.node()` action.
slots cannot be deleted.
Parameters
----------
node: tuple
A node reference, obtained by the actions `node`.
Returns
-------
None
"""
self.stats[self.D] += 1
if node is not None:
slotType = self.slotType
nType = node[0]
if nType == slotType:
errors = self.errors
errors["cannot delete a slot"].append(node[1])
return
self.discardables.add(node)
def resume(self, node):
"""**resumes** a node.
```
cv.resume(n)
```
If you resume a non-slot node, you add it again to the set of embedders.
No new node will be created.
If you resume a slot node, it will be added to the set of current embedders.
No new slot will be created.
Parameters
----------
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
Returns
-------
None
"""
curEmbedders = self.curEmbedders
oslots = self.oslots
self.stats[self.R] += 1
(nType, seq) = node
if nType == self.slotType:
for eNode in curEmbedders:
oslots[eNode].add(seq)
else:
self._checkSecLevel(node, before=None)
curEmbedders.add(node)
def link(self, node, slots):
"""Links the given, existing slots to a node.
```
cv.link(n, [s1, s2])
```
Sometimes the automatic linking of slots to nodes is not sufficient.
This happens when you feel the need to construct a node retro-actively,
when the slots that need to be linked to it have already been created.
This action is precisely meant for that.
Parameters
----------
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
slots: iterable of integer
Returns
-------
boolean
"""
oslots = self.oslots
good = True
for seq in slots:
oslots[node].add(seq)
return good
def linked(self, node):
"""Returns the slots `ss` to which a node is currently linked.
```
ss = cv.linked(n)
```
If you construct non-slot nodes without linking them to slots,
they will be removed when TF validates the collective result
of the action methods.
If you want to prevent that, you can insert an extra slot, but in order
to do so, you have to detect that a node is still unlinked.
This action is precisely meant for that.
Parameters
----------
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
Returns
-------
tuple of integer
The slots are returned as a tuple of integers, sorted.
"""
oslots = self.oslots
return tuple(sorted(oslots.get(node, [])))
def feature(self, node, **features):
"""Adds **node features**.
```
cv.feature(n, name=value, ... , name=value)
```
Parameters
----------
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
features: keyword arguments
The names and values of features to assign to this node.
Returns
-------
None
!!! caution "None values"
If a feature value is `None` it will not be added!
"""
nodeFeatures = self.nodeFeatures
self.stats[self.F] += 1
for (k, v) in features.items():
if v is None:
continue
# self._checkType(k, v, self.N)
nodeFeatures[k][node] = v
def edge(self, nodeFrom, nodeTo, **features):
"""Adds **edge features**.
```
cv.edge(nf, nt, name=value, ... , name=value)
```
Parameters
----------
nodeFrom, nodeTo: tuple
Two node references, obtained by one of the actions `slot` or `node`.
features: keyword arguments
The names and values of features to assign to this edge
(i.e. pair of nodes).
Returns
-------
None
!!! note "None values"
You may pass values that are `None`,
and a corresponding edge will be created.
If for all edges the value is `None`,
an edge without values will be created.
For every `nodeFrom`, such a feature
essentially specifies a set of nodes `{ nodeTo }`.
"""
edgeFeatures = self.edgeFeatures
self.stats[self.E] += 1
for (k, v) in features.items():
# self._checkType(k, v, self.E)
edgeFeatures[k][nodeFrom][nodeTo] = v
def occurs(self, feat):
"""Whether the feature `featureName` occurs in the resulting data so far.
```
occurs = cv.occurs(featureName)
```
If you have assigned None values to a feature, that will count, i.e.
that feature occurs in the data.
If you add feature values conditionally, it might happen that some
features will not be used at all.
For example, if your conversion produces errors, you might
add the error information to the result in the form of error features.
Later on, when the errors have been weeded out, these features will
not occur any more in the result, but then TF will complain that
such is feature is declared but not used.
At the end of your director you can remove unused features
conditionally, using this function.
Parameters
----------
feat: string
The name of a feature
Returns
-------
boolean
"""
nodeFeatures = self.nodeFeatures
edgeFeatures = self.edgeFeatures
if feat in nodeFeatures or feat in edgeFeatures:
return True
return False
def meta(self, feat, **metadata):
"""Adds, modifies, deletes metadata fields of features.
```
cv.meta(feature, name=value, ... , name=value)
```
Parameters
----------
feat: string
The name of a feature
metadata: dict
If a `value` is `None`, that `name` will be deleted from the
metadata fields of the feature.
A bare `cv.meta(feature)` will remove the all metadata from the feature.
If you modify the field `valueType` of a feature, that feature will be
added or removed from the set of `intFeatures`.
It will be checked whether you specify either `int` or `str`.
Returns
-------
None
"""
errors = self.errors
intFeatures = self.intFeatures
metaData = self.metaData
featMeta = metaData.get(feat, {})
good = True
if not metadata:
if feat in metaData:
del metaData[feat]
intFeatures.discard(feat)
for (field, text) in metadata.items():
if text is None:
if field == "valueType":
errors['did not delete metadata field "valueType"'].append(feat)
good = False
else:
if feat in metaData and field in metaData[feat]:
del metaData[feat][field]
else:
metaData.setdefault(feat, {})[field] = text
if field == "valueType":
if text == "int":
intFeatures.add(feat)
else:
intFeatures.discard(feat)
self.good = self._checkFeatMeta(feat, featMeta) and good and self.good
def features(self):
"""Gets the list of all features.
```
featureNames = cv.features()
```
Returns
-------
list
"""
nodeFeatures = self.nodeFeatures
edgeFeatures = self.edgeFeatures
return sorted(list(nodeFeatures) + list(edgeFeatures))
def active(self, node):
"""Returns whether a node is currently active.
Active nodes are the nodes in the set of current embedders.
```
isActive = cv.active(n)
```
If you construct your nodes in a very dynamic way, it might be
hard to keep track for each node whether it has been created, terminated,
or resumed, in other words, whether it is active or not.
This action is provides a direct and precise way to know
whether a node is active.
Parameters
----------
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
Returns
-------
boolean
"""
return node in self.curEmbedders
def activeNodes(self, nTypes=None):
"""The currently active nodes, i.e. the embedders.
```
nodes = cv.activeNodes()
nodes = cv.activeNodes(nTypes=("sentence", "clause"))
```
Parameters
----------
nTypes: iterable optional None
If None, all active nodes are returned.
Else the iterable lists a few node types,
and only active nodes in these types are returned.
Returns
-------
set
"""
if nTypes is None:
return set(self.curEmbedders)
nTypes = set(nTypes)
return {n for n in self.curEmbedders if n[0] in nTypes}
def activeTypes(self):
"""The node types of the currently active nodes, i.e. the embedders.
```
nTypes = cv.activeTypes()
```
Parameters
----------
None
Returns
-------
set
"""
return {node[0] for node in self.curEmbedders}
def get(self, feature, *args):
"""Retrieves feature values.
```
cv.get(feature, n) and cv.get(feature, nf, nt)
```
`feature` is the name of the feature.
For node features, `n` is the node which carries the value.
For edge features, `nf, nt` is the pair of from-to nodes which carries the value.
Parameters
----------
feature: string
The name of a feature
node: tuple
A node reference, obtained by one of the actions `slot` or `node`.
The node in question when retrieving the value of a node feature.
nodeFrom, nodeTo: tuple
Two node references, obtained by one of the actions `slot` or `node`.
The nodes in question when retrieving the value of an edge feature.
Returns
-------
string or integer
"""
errors = self.errors
nodeFeatures = self.nodeFeatures
edgeFeatures = self.edgeFeatures
nArgs = len(args)
if nArgs == 0 or nArgs > 2:
errors["use `cv.get(ft, n)` or `cv.get(ft, nf, nt)`"].append(None)
return None
return (
nodeFeatures.get(feature, {}).get(args[0], None)
if len(args) == 1
else edgeFeatures.get(feature, {}).get(args[0], {}).get(args[1], None)
)
def _checkSecLevel(self, node, before=True):
levelFromSection = self.levelFromSection
sectionFeatures = self.sectionFeatures
nodeFeatures = self.nodeFeatures
warnings = self.warnings
curEmbedders = self.curEmbedders
(nType, seq) = node
msg = "starts" if before is True else "ends" if before is False else "resumes"
if levelFromSection:
level = levelFromSection.get(nType, None)
if level is None:
return
headingFeature = sectionFeatures[level - 1]
nHeading = nodeFeatures.get(headingFeature, {}).get(node, "??")
for em in curEmbedders:
eType = em[0]
if eType in levelFromSection:
eLevel = levelFromSection.get(eType, None)
eHeadingFeature = sectionFeatures[eLevel - 1]
eHeading = nodeFeatures.get(eHeadingFeature, {}).get(em, "??")
if eType == nType:
warnings[
f'section {nType} "{nHeading}" of level {level}'
f" enclosed in another {nType}: {eHeading}"
].append(None)
elif eType in levelFromSection:
eLevel = levelFromSection[eType]
if eLevel > level:
warnings[
f'section {nType} "{nHeading}" of level {level} {msg}'
f' inside a {eType} "{eHeading}" of level {eLevel}'
].append(None)
def _follow(self, director):
# after node = yield ('N', nodeType) all slot nodes that are yielded
# will be linked to node, until a ('T', node) is yielded.
# If needed, you can resume this node again, after which new slot nodes
# continue to be linked to this node.
# If you resume a slot node, it all non slot nodes in the current context
# will be linked to it.
silent = self.silent
tmObj = self.TF.tmObj
info = tmObj.info
if not self.good:
return
info("Following director... ", force=silent != DEEP)
slotType = self.slotType
errors = self.errors
self.oslots = collections.defaultdict(set)
self.nodeFeatures = collections.defaultdict(dict)
self.edgeFeatures = collections.defaultdict(
lambda: collections.defaultdict(dict)
)
self.nodes = collections.defaultdict(set)
nodes = self.nodes
self.curSeq = collections.Counter()
self.curEmbedders = set()
curEmbedders = self.curEmbedders
self.stats = {
actionType: 0
for actionType in (self.S, self.N, self.T, self.R, self.D, self.F, self.E)
}
director(self)
if not self.stats:
self.good = False
return
for (actionType, amount) in sorted(self.stats.items()):
info(f'"{actionType}" actions: {amount}', force=silent != DEEP)
totalNodes = 0
for (nType, lastSeq) in sorted(self.curSeq.items()):
for seq in range(1, lastSeq + 1):
nodes[nType].add(seq)
slotRep = " = slot type" if nType == slotType else ""
info(f'{lastSeq:>8} x "{nType}" node {slotRep}', tm=False)
totalNodes += lastSeq
info(f"{totalNodes:>8} nodes of all types", tm=False, force=silent != DEEP)
self.totalNodes = totalNodes
if curEmbedders:
embedCount = collections.Counter()
for (nType, seq) in curEmbedders:
embedCount[nType] += 1
for (nType, amount) in sorted(
embedCount.items(),
key=lambda x: (-x[1], x[0]),
):
errors["Unterminated nodes"].append(f"{nType}: {amount} x")
self._showErrors()
def _removeUnlinked(self):
silent = self.silent
tmObj = self.TF.tmObj
info = tmObj.info
indent = tmObj.indent
if not self.good and not self.force:
return
nodeTypes = self.curSeq
nodes = self.nodes
slotType = self.slotType
oslots = self.oslots
discardables = self.discardables
nodeFeatures = self.nodeFeatures
edgeFeatures = self.edgeFeatures
unlinked = {}
nDeleted = 0
for (nType, seq) in discardables:
unlinked.setdefault(nType, set()).add(seq)
nDeleted += 1
if (nType, seq) in oslots:
del oslots[(nType, seq)]
for nType in nodeTypes:
if nType == slotType:
continue
for seq in range(1, nodeTypes[nType] + 1):
if (nType, seq) not in oslots:
unlinked.setdefault(nType, set()).add(seq)
if unlinked:
info(
f"Removing unlinked (of which {nDeleted} deleted) nodes ... ",
force=silent != DEEP,
)
indent(level=2)
totalRemoved = 0
for (nType, seqs) in unlinked.items():
seqs = sorted(seqs)
theseNodes = nodes[nType]
lSeqs = len(seqs)
totalRemoved += lSeqs
rep = " ..." if lSeqs > 5 else ""
pl = "" if lSeqs == 1 else "s"
info(
f'{lSeqs:>6} unlinked "{nType}" node{pl}: {seqs[0:5]}{rep}',
force=silent != DEEP,
)
for seq in seqs:
node = (nType, seq)
theseNodes.discard(seq)
for (f, fData) in nodeFeatures.items():
if node in fData:
del fData[node]
for (f, fData) in edgeFeatures.items():
if node in fData:
del fData[node]
for (fNode, toValues) in fData:
if node in toValues:
del toValues[node]
pl = "" if totalRemoved == 1 else "s"
info(f"{totalRemoved:>6} unlinked node{pl}", force=silent != DEEP)
self.totalNodes -= totalRemoved
info(f"Leaving {self.totalNodes:>6} nodes", force=silent != DEEP)
indent(level=1)
def _checkGraph(self):
silent = self.silent
tmObj = self.TF.tmObj
info = tmObj.info
if not self.good and not self.force:
return
info("checking for nodes and edges ... ", force=silent != DEEP)
nodes = self.nodes
errors = self.errors
edgeFeatures = self.edgeFeatures
# edges refer to nodes
for (k, featureData) in edgeFeatures.items():
for nFrom in featureData:
(nType, seq) = nFrom
if nType not in nodes or seq not in nodes[nType]:
errors["Edge feature: illegal node"].append(
f'"{k}": from-node {nFrom} not in node set'
)
continue
for nTo in featureData[nFrom]:
(nType, seq) = nTo
if nType not in nodes or seq not in nodes[nType]:
errors["Edge feature: illegal node"].append(
f'"{k}": to-node {nTo} not in node set'
)
self._showErrors()
def _checkFeatures(self):
silent = self.silent
tmObj = self.TF.tmObj
info = tmObj.info
if not self.good and not self.force:
return
info("checking (section) features ... ", force=silent != DEEP)
intFeatures = self.intFeatures
metaData = self.metaData
nodes = self.nodes
nodeFeatures = self.nodeFeatures
edgeFeatures = self.edgeFeatures
errors = self.errors
for feat in intFeatures:
if (
feat not in WARP
and feat not in nodeFeatures
and feat not in edgeFeatures
):
errors["intFeatures"].append(
f'"{feat}" is declared as integer valued, '
"but this feature does not occur"
)
for nType in self.sectionTypes:
if nType not in nodes:
errors["sections"].append(
f'node type "{nType}" is declared as a section type, '
"but this node type does not occur"
)
for feat in self.sectionFeatures:
if feat not in nodeFeatures:
errors["sections"].append(
f'"{feat}" is declared as a section feature, '
"but this node feature does not occur"
)
for nType in self.structureTypes:
if nType not in nodes:
errors["structure"].append(
f'node type "{nType}" is declared as a structure type,'
f" but this node type does not occur"
)
for feat in self.structureFeatures:
if feat not in nodeFeatures:
errors["structure"].append(
f'"{feat}" is declared as a structure feature, '
"but this node feature does not occur"
)
nodeFeatures[feat] = {}
sectionSet = self.sectionSet
structureSet = self.structureSet
featFromSectionType = self.featFromSectionType
featFromStructureType = self.featFromStructureType
for nType in nodes:
if nType in structureSet:
feat = featFromStructureType[nType]
for seq in nodes[nType]:
if (nType, seq) not in nodeFeatures[feat]:
errors["structure features"].append(
f'"structure element "{nType}" {seq} '
f'has no value for "{feat}"'
)
if nType in sectionSet:
feat = featFromSectionType[nType]
for seq in nodes[nType]:
if (nType, seq) not in nodeFeatures[feat]:
errors["section features"].append(
f'"section element "{nType}" {seq} '
f'has no value for "{feat}"'
)
for feat in self.textFeatures:
if feat not in nodeFeatures:
errors["text formats"].append(
f'"{feat}" is used in a text format, '
"but this node feature does not occur"
)
for feat in WARP:
if feat in nodeFeatures or feat in edgeFeatures:
errors[feat].append(f'Do not construct the "{feat}" feature yourself')
for feat in sorted(nodeFeatures) + sorted(edgeFeatures):
if feat not in self.metaData:
errors["feature metadata"].append(
f'node feature "{feat}" has no metadata'
)
for feat in sorted(metaData):
if (
feat
and feat not in WARP
and feat not in nodeFeatures
and feat not in edgeFeatures
):
errors["feature metadata"].append(
f'node feature "{feat}" has metadata but does not occur'
)
for (feat, featData) in sorted(nodeFeatures.items()):
if None in featData:
errors["feature values assigned to None"].append(
f'node feature "{feat}" has a node None'
)
for (feat, featData) in sorted(edgeFeatures.items()):
if None in featData:
errors["feature values assigned to None"].append(
f'edge feature "{feat}" has a from-node None'
)
for toValues in featData.values():
if None in toValues:
errors["feature values assigned to None"].append(
f'edge feature "{feat}" has a to-node None'
)
for (feat, featData) in sorted(edgeFeatures.items()):
if feat in WARP:
continue
hasValues = False
for (nodeTo, toValues) in featData.items():
if any(v is not None for v in toValues.values()):
hasValues = True
break
if not hasValues:
edgeFeatures[feat] = {
nodeTo: set(toValues) for (nodeTo, toValues) in featData.items()
}
metaData.setdefault(feat, {})["edgeValues"] = hasValues
for feat in intFeatures:
if feat in WARP:
continue
if feat in nodeFeatures:
featData = nodeFeatures[feat]
for (k, v) in featData.items():
if not isInt(v):
(nType, node) = k
errors["Not a number"].append(
f'"node feature "{feat}": {nType} {node} => "{v}"'
)
if feat in edgeFeatures and metaData[feat]["edgeValues"]:
featData = edgeFeatures[feat]
for (fromNode, toValues) in featData.items():
(fType, fNode) = fromNode
for (toNode, v) in toValues.items():
(tType, tNode) = toNode
if not isInt(v):
errors["Not a number"].append(
f'"edge feature "{feat}":'
f' {fType} {fNode} ="{v}"=> {tType} {tNode}'
)
self._showErrors()
def _reorderNodes(self):
silent = self.silent
tmObj = self.TF.tmObj
info = tmObj.info
if not self.good and not self.force:
return
info("reordering nodes ...", force=silent != DEEP)
nodeTypes = self.curSeq
nodes = self.nodes
oslots = self.oslots
slotType = self.slotType
slotKeys = self.slotKeys
nTypes = (slotType,) + tuple(
sorted(nType for nType in nodes if nType != slotType)
)
self.nodeMap = {}
self.maxSlot = nodeTypes[slotType]
nodeMap = self.nodeMap
maxSlot = self.maxSlot
newN = 0
# we build a node map that we use later on to remap the features
# However, we do not have to remap the otype feature.
# And we only have to remap the oslots feature if the order of the slots
# is changed.
# first we reorder the slots, if needed. As a consequence, we have to
# 1. map the old slot nodes to the new slot nodes
# 2. adapt the oslots: each node must be linked to the new slots
# Because of the node map, _reassign features will adjust the other
# node and edge features for the slots
sortedSeqs = range(1, maxSlot + 1)
if len(slotKeys) > 0:
info(f'Sorting {maxSlot} slots (node type "{slotType}")')
sortedSeqs = sorted(sortedSeqs, key=lambda x: (slotKeys.get(x, ""), x))
else:
info("No slot sorting needed")
for seq in sortedSeqs:
newN += 1
nodeMap[(slotType, seq)] = newN
# now we adapt the oslots feature
if len(slotKeys) > 0:
self.oslots = {
n: {nodeMap[(slotType, s)] for s in slots}
for (n, slots) in oslots.items()
}
oslots = self.oslots
for nType in nTypes:
canonical = self._canonical(nType)
if nType == slotType:
continue
seqs = nodes[nType]
info(f'Sorting {len(seqs)} nodes of type "{nType}"')
sortedSeqs = sorted(seqs, key=canonical)
for seq in sortedSeqs:
newN += 1
nodeMap[(nType, seq)] = newN
self.maxNode = newN
info(f"Max node = {newN}", force=silent != DEEP)
self._showErrors()
def _canonical(self, nType):
oslots = self.oslots
def before(nodeA, nodeB):
slotsA = oslots[(nType, nodeA)]
slotsB = oslots[(nType, nodeB)]
if slotsA == slotsB:
return 0
aWithoutB = slotsA - slotsB
if not aWithoutB:
return 1
bWithoutA = slotsB - slotsA
if not bWithoutA:
return -1
aMin = min(aWithoutB)
bMin = min(bWithoutA)
return -1 if aMin < bMin else 1
return functools.cmp_to_key(before)
def _reassignFeatures(self):
silent = self.silent
tmObj = self.TF.tmObj
info = tmObj.info
indent = tmObj.indent
if not self.good and not self.force:
return
info("reassigning feature values ...", force=silent != DEEP)
nodeMap = self.nodeMap
oslots = self.oslots
nodeFeatures = self.nodeFeatures
edgeFeatures = self.edgeFeatures
otype = {n: nType for ((nType, seq), n) in nodeMap.items()}
oslots = {nodeMap[node]: slots for (node, slots) in oslots.items()}
nodeFeaturesProto = self.nodeFeatures
edgeFeaturesProto = self.edgeFeatures
nodeFeatures = collections.defaultdict(dict)
edgeFeatures = collections.defaultdict(lambda: collections.defaultdict(dict))
indent(level=2)
for k in sorted(nodeFeaturesProto):
featureDataProto = nodeFeaturesProto[k]
ln = len(featureDataProto)
pl = "" if ln == 1 else "s"
info(f'node feature "{k}" with {ln} node{pl}', tm=False)
featureData = {}
for (node, value) in featureDataProto.items():
featureData[nodeMap[node]] = value
nodeFeatures[k] = featureData
for k in sorted(edgeFeaturesProto):
featureDataProto = edgeFeaturesProto[k]
ln = len(featureDataProto)
pl = "" if ln == 1 else "s"
info(f'edge feature "{k}" with {ln} start node{pl}', tm=False)
featureData = {}
for (nodeFrom, toValues) in featureDataProto.items():
if type(toValues) is set:
toData = {nodeMap[nodeTo] for nodeTo in toValues}
else:
toData = {}
for (nodeTo, value) in toValues.items():
toData[nodeMap[nodeTo]] = value
featureData[nodeMap[nodeFrom]] = toData
edgeFeatures[k] = featureData
nodeFeatures["otype"] = otype
edgeFeatures["oslots"] = oslots
indent(level=1)
self.oslots = None
self.otype = None
self.nodeFeatures = nodeFeatures
self.edgeFeatures = edgeFeatures
self._showErrors()Classes
class CV (TF, silent='auto')-
The object that contains the walker conversion machinery.
silent: string, optional tf.core.timestamp.SILENT_D See
TimestampExpand source code Browse git
class CV: S = "slot" N = "node" T = "terminate" R = "resume" D = "delete" F = "feature" E = "edge" def __init__(self, TF, silent=SILENT_D): """The object that contains the walker conversion machinery. silent: string, optional tf.core.timestamp.SILENT_D See `tf.core.timestamp.Timestamp` """ self.TF = TF self.silent = silent tmObj = TF.tmObj isSilent = tmObj.isSilent setSilent = tmObj.setSilent silent = silentConvert(silent) self.wasSilent = isSilent() setSilent(silent) def _showWarnings(self): silent = self.silent tmObj = self.TF.tmObj error = tmObj.error info = tmObj.info indent = tmObj.indent warnings = self.warnings warn = self.warn if warn is None: if warnings: info("use `cv.walk(..., warn=False)` to make warnings visible") info("use `cv.walk(..., warn=True)` to stop on warnings") else: method = error if warn else info if warnings: for (kind, msgs) in sorted(warnings.items()): method(f"WARNING {kind} ({len(msgs)} x):", force=silent != DEEP) indent(level=2) for msg in sorted(set(msgs))[0:20]: if msg: method(f"{msg}", tm=False, force=silent != DEEP) self.warnings = {} if warn: info("use `cv.walk(..., warn=False)` to continue after warnings") info("use `cv.walk(..., warn=None)` to suppress warnings") self.good = False else: info("use `cv.walk(..., warn=True)` to stop after warnings") info("use `cv.walk(..., warn=None)` to suppress warnings") def _showErrors(self): silent = self.silent tmObj = self.TF.tmObj error = tmObj.error info = tmObj.info indent = tmObj.indent forcedStop = self.forcedStop errors = self.errors if errors: for (kind, msgs) in sorted(errors.items()): error(f"ERROR {kind} ({len(msgs)} x):") indent(level=2) for msg in sorted(set(msgs))[0:20]: if msg: error(f"{msg}", tm=False) self.errors = {} self.good = False if forcedStop: error("STOPPED by the stop() instruction") elif not errors: if self.good: info("OK", force=silent != DEEP) else: error("STOPPED because of warnings") def walk( self, director, slotType, otext={}, generic={}, intFeatures=set(), featureMeta={}, warn=True, generateTf=True, force=False, ): """Asks a director function to walk through source data and receives its actions. The `director` function should unravel the source. You have to program the `director`, which takes one argument: `cv`. From the `cv` you can use a few standard actions that instruct TF to build a graph. This function will check whether the metadata makes sense and is minimally complete. During node creation the section structure will be watched, and you will be warned if irregularities occur. After the creation of the feature data, some extra checks will be performed to see whether the metadata matches the data and vice versa. If the slots need to be sorted by their keys, it will happen at this point, and the generated features will be adapted to the sorted slots. The new feature data will be written to the output directory of the underlying TF object. In fact, the rules are exactly the same as for `tf.fabric.Fabric.save`. Parameters ---------- slotType: string The node type that acts as the type of the slots in the data set. oText: dict The configuration information to be stored in the `otext` feature (see `tf.core.text`): * section types * section features * structure types * structure features * text formats generic: dict Metadata that will be written into the header of all generated TF features. You can make changes to this later on, dynamically in your director. intFeatures: iterable The set of features that have integer values only. You can make changes to this later on, dynamically in your director. featureMeta: dict of dict For each node or edge feature descriptive metadata can be supplied. You can make changes to this later on, dynamically in your director. warn: boolean, optional True This regulates the response to warnings: `True` (default): stop after warnings (as if they are errors); `False` continue after warnings but do show them; `None` suppress all warnings. force: boolean, optional False This forces the process to continue after errors. Your TF set might not be valid. Yet this can be useful during testing, when you know that not everything is OK, but you want to check some results. Especially when dealing with large datasets, you might want to test with little pieces. But then you get a kind of non-fatal errors that stand in the way of testing. For those cases: `force=True`. generateTf: boolean, optional True You can pass `False` here to suppress the actual writing of TF data. In that way you can dry-run the director to check for errors and warnings director: function An ordinary function that takes one argument, the `cv` object, and should not deliver anything. Writing this function is the main job to do when you want to convert a data source to TF. See `tf.convert.walker` for more details. Returns ------- boolean Whether the operation was successful """ tmObj = self.TF.tmObj info = tmObj.info indent = tmObj.indent setSilent = tmObj.setSilent silent = self.silent indent(level=0, reset=True) info("Importing data from walking through the source ...", force=silent != DEEP) self.force = force self.good = True self.forcedStop = False self.errors = collections.defaultdict(list) self.warnings = collections.defaultdict(list) self.warn = warn self.slotType = slotType self.intFeatures = set(intFeatures) self.featureMeta = featureMeta self.metaData = {} self.nodeFeatures = {} self.edgeFeatures = {} self.slotKeys = {} self.discardables = set() indent(level=1, reset=True) self._prepareMeta(otext, generic) indent(level=1, reset=True) self._follow(director) indent(level=1, reset=True) self._removeUnlinked() indent(level=1, reset=True) self._checkGraph() indent(level=1, reset=True) self._checkFeatures() indent(level=1, reset=True) self._reorderNodes() indent(level=1, reset=True) self._reassignFeatures() indent(level=0) info("Features ready to write") if generateTf: if self.good or self.force: self.good = self.TF.save( metaData=self.metaData, nodeFeatures=self.nodeFeatures, edgeFeatures=self.edgeFeatures, silent=silent, ) self._showWarnings() setSilent(self.wasSilent) return self.good def _prepareMeta(self, otext, generic): silent = self.silent varRe = re.compile(r"\{([^}]+)\}") tmObj = self.TF.tmObj info = tmObj.info indent = tmObj.indent if not self.good and not self.force: return info("Preparing metadata... ", force=silent != DEEP) intFeatures = self.intFeatures featureMeta = self.featureMeta errors = self.errors self.metaData = { "": generic, OTYPE: {"valueType": "str"}, OSLOTS: {"valueType": "str"}, OTEXT: otext, } metaData = self.metaData self.intFeatures = intFeatures self.sectionTypes = [] self.sectionFeatures = [] self.sectionFromLevel = {} self.levelFromSection = {} self.structureTypes = [] self.structureFeatures = [] self.structureLevel = {} self.textFormats = {} self.textFeatures = set() if not generic: errors['Missing feature meta data in "generic"'].append( "Consider adding provenance metadata to all features" ) if not otext: errors['Missing "otext" configuration'].append( "Consider adding configuration for text representation and section levels" ) else: sectionInfo = {} for f in ("sectionTypes", "sectionFeatures"): if f not in otext: errors['Incomplete section specs in "otext"'].append( f'no key "{f}"' ) sectionInfo[f] = [] else: sFields = itemize(otext[f], sep=",") sectionInfo[f] = sFields if f == "sectionTypes": for (i, s) in enumerate(sFields): self.levelFromSection[s] = i + 1 self.sectionFromLevel[i + 1] = s sLevels = {f: len(sectionInfo[f]) for f in sectionInfo} if min(sLevels.values()) != max(sLevels.values()): errors["Inconsistent section info"].append( " but ".join(f'"{f}" has {sLevels[f]} levels' for f in sLevels) ) self.sectionFeatures = sectionInfo["sectionFeatures"] self.sectionTypes = sectionInfo["sectionTypes"] self.featFromSectionType = { typ: feat for (typ, feat) in zip(self.sectionTypes, self.sectionFeatures) } self.sectionSet = set(self.sectionTypes) structureInfo = {} for f in ("structureTypes", "structureFeatures"): if f not in otext: structureInfo[f] = [] continue sFields = itemize(otext[f], sep=",") structureInfo[f] = sFields if not structureInfo: info("No structure definition found in otext") sLevels = {f: len(structureInfo[f]) for f in structureInfo} if min(sLevels.values()) != max(sLevels.values()): errors["Inconsistent structure info"].append( " but ".join(f'"{f}" has {sLevels[f]} levels' for f in sLevels) ) structureInfo = {} if not structureInfo or all( len(info) == 0 for (s, info) in structureInfo.items() ): info("No structure nodes will be set up") self.structureFeatures = [] self.structureTypes = [] self.structureFeatures = structureInfo["structureFeatures"] self.structureTypes = structureInfo["structureTypes"] self.featFromStructureType = { typ: feat for (typ, feat) in zip(self.structureTypes, self.structureFeatures) } self.structureSet = set(self.structureTypes) textFormats = {} textFeatures = set() for (k, v) in sorted(otext.items()): if k.startswith("fmt:"): featureSet = set() features = varRe.findall(v) for ff in features: fr = ff.rsplit(":", maxsplit=1)[0] for f in fr.split("/"): featureSet.add(f) textFormats[k[4:]] = featureSet textFeatures |= featureSet if not textFormats: errors['No text formats in "otext"'].append('add "fmt:text-orig-full"') elif "text-orig-full" not in textFormats: errors["No default text format in otext"].append( 'add "fmt:text-orig-full"' ) self.textFormats = textFormats self.textFeatures = textFeatures info(f'SECTION TYPES: {", ".join(self.sectionTypes)}', tm=False) info(f'SECTION FEATURES: {", ".join(self.sectionFeatures)}', tm=False) info(f'STRUCTURE TYPES: {", ".join(self.structureTypes)}', tm=False) info(f'STRUCTURE FEATURES: {", ".join(self.structureFeatures)}', tm=False) info("TEXT FEATURES:", tm=False) indent(level=2) for (fmt, feats) in sorted(textFormats.items()): info(f'{fmt:<20} {", ".join(sorted(feats))}', tm=False) indent(level=1) for feat in WARP + ("",): if feat in intFeatures: if feat == "": errors["intFeatures"].append( 'Do not declare the "valueType" for all features' ) else: errors["intFeatures"].append( f'Do not mark the "{feat}" feature as integer valued' ) self.good = False for (feat, featMeta) in sorted(featureMeta.items()): good = self._checkFeatMeta( feat, featMeta, checkRegular=True, valueTypeAllowed=False, showErrors=False, ) if not good: self.good = False metaData.setdefault(feat, {}).update(featMeta) metaData[feat]["valueType"] = "int" if feat in intFeatures else "str" self._showErrors() def _checkFeatMeta( self, feat, featMeta, checkRegular=False, valueTypeAllowed=True, showErrors=True, ): errors = collections.defaultdict(list) good = True if checkRegular: if feat in WARP + ("",): if feat == "": errors["featureMeta"].append( 'Specify the generic feature meta data in "generic"' ) good = False elif feat == OTEXT: errors["featureMeta"].append( f'Specify the "{OTEXT}" feature in "otext"' ) good = False else: errors["featureMeta"].append( f'Do not pass metaData for the "{feat}" feature in "featureMeta"' ) good = False if "valueType" in featMeta: if not valueTypeAllowed: errors["featureMeta"].append( f'Do not specify "valueType" for the "{feat}" feature in "featureMeta"' ) good = False elif featMeta["valueType"] not in {"int", "str"}: errors["featureMeta"].append('valueType must be "int" or "str"') good = False for (e, eData) in errors.items(): self.errors[e].extend(eData) if showErrors: self._showErrors return good def stop(self, msg): """Stops the director. No further input will be read. ``` cv.stop(msg) ``` The director will exit with a non-good status and issue the message `msg`. If you have called `walk()` with `force=True`, indicating that the director must proceed after errors, then this stop command will cause termination nevertheless. Parameters ---------- msg: string A message to display upon stopping. Returns ------- None """ tmObj = self.TF.tmObj error = tmObj.error error(f"Forced stop: {msg}") self.good = False self.force = False self.forcedStop = True def slot(self, key=None): """Makes a slot node and return the handle to it in `n`. ``` n = cv.slot() ``` No further information is needed. Remember that you can add features to the node by later ``` cv.feature(n, key=value, ...) ``` calls. Parameters ---------- key: string, optional None If passed, it acts as a sort key on the slot. At the end of the walk, all slots will be sorted by their key and then by their original order. Care will be taken that slots retain their features and linkages. !!! note "Keys are strings" Note that the key must be a string. If you want to sort on numbers, make sure to pad all numbers with leading zeros. Returns ------- node reference: tuple The node reference consists of a node type and a sequence number, but normally you do not have to dig these out. Just pass the tuple as a whole to actions that require a node argument. """ curSeq = self.curSeq curEmbedders = self.curEmbedders oslots = self.oslots slotKeys = self.slotKeys levelFromSection = self.levelFromSection warnings = self.warnings self.stats[self.S] += 1 nType = self.slotType curSeq[nType] += 1 seq = curSeq[nType] if key is not None: slotKeys[seq] = key inSection = False for eNode in curEmbedders: if eNode[0] in levelFromSection: inSection = True oslots[eNode].add(seq) if levelFromSection and not inSection: warnings["slot outside sections"].append(f"{seq}") return (nType, seq) def node(self, nType, slots=None): """Makes a non-slot node and return the handle to it in `n`. ``` n = cv.node(nodeType) ``` You have to pass its *node type*, i.e. a string. Think of `sentence`, `paragraph`, `phrase`, `word`, `sign`, whatever. There are two modes for this function: * Auto: (`slots=None`): Non slot nodes will be automatically added to the set of embedders. * Explicit: (`slots=iterable`): The slots in iterable will be assigned to this node and nothing else. The node will not be added to the set of embedders. Put otherwise: the node will be terminated after construction. However: you could resume it later to add other slots. Remember that you can add features to the node by later ``` cv.feature(n, key=value, ...) ``` calls. Parameters ---------- nType: string A node type, not the slot type slots: iterable of integer, optional None The slots to assign to this node. If left out, the node is left as an embedding node and subsequent slots will be added to it automatically. All slots in the iterable must have been generated before by means of the `cv.slot()` action. Returns ------- node reference or None If an error occurred, `None` is returned. The node reference consists of a node type and a sequence number, but normally you do not have to dig these out. Just pass the tuple as a whole to actions that require a node argument. """ slotType = self.slotType errors = self.errors if nType == slotType: errors[f'use `cv.slot()` instead of `cv.node("{nType}")`'].append(None) return curSeq = self.curSeq curEmbedders = self.curEmbedders self.stats[self.N] += 1 curSeq[nType] += 1 seq = curSeq[nType] node = (nType, seq) self._checkSecLevel(node, before=True) if slots: maxSlot = curSeq[slotType] for s in slots: if not 1 <= s <= maxSlot: errors[f"slot out of range in `cv.node(({nType}, {seq}))`"].append( f"{s}" ) else: oslots = self.oslots oslots[node].add(s) self.stats[self.T] += 1 else: curEmbedders.add(node) return node def terminate(self, node): """**terminates** a node. ``` cv.terminate(n) ``` The node `n` will be removed from the set of current embedders. This `n` must be the result of a previous `cv.slot()` or `cv.node()` action. Parameters ---------- node: tuple A node reference, obtained by one of the actions `slot` or `node`. Returns ------- None """ self.stats[self.T] += 1 if node is not None: self.curEmbedders.discard(node) self._checkSecLevel(node, before=False) def delete(self, node): """**deletes** a node. ``` cv.delete(n) ``` The node `n` will be deleted from the set of nodes that will be created. This `n` must be the result of a previous `cv.node()` action. slots cannot be deleted. Parameters ---------- node: tuple A node reference, obtained by the actions `node`. Returns ------- None """ self.stats[self.D] += 1 if node is not None: slotType = self.slotType nType = node[0] if nType == slotType: errors = self.errors errors["cannot delete a slot"].append(node[1]) return self.discardables.add(node) def resume(self, node): """**resumes** a node. ``` cv.resume(n) ``` If you resume a non-slot node, you add it again to the set of embedders. No new node will be created. If you resume a slot node, it will be added to the set of current embedders. No new slot will be created. Parameters ---------- node: tuple A node reference, obtained by one of the actions `slot` or `node`. Returns ------- None """ curEmbedders = self.curEmbedders oslots = self.oslots self.stats[self.R] += 1 (nType, seq) = node if nType == self.slotType: for eNode in curEmbedders: oslots[eNode].add(seq) else: self._checkSecLevel(node, before=None) curEmbedders.add(node) def link(self, node, slots): """Links the given, existing slots to a node. ``` cv.link(n, [s1, s2]) ``` Sometimes the automatic linking of slots to nodes is not sufficient. This happens when you feel the need to construct a node retro-actively, when the slots that need to be linked to it have already been created. This action is precisely meant for that. Parameters ---------- node: tuple A node reference, obtained by one of the actions `slot` or `node`. slots: iterable of integer Returns ------- boolean """ oslots = self.oslots good = True for seq in slots: oslots[node].add(seq) return good def linked(self, node): """Returns the slots `ss` to which a node is currently linked. ``` ss = cv.linked(n) ``` If you construct non-slot nodes without linking them to slots, they will be removed when TF validates the collective result of the action methods. If you want to prevent that, you can insert an extra slot, but in order to do so, you have to detect that a node is still unlinked. This action is precisely meant for that. Parameters ---------- node: tuple A node reference, obtained by one of the actions `slot` or `node`. Returns ------- tuple of integer The slots are returned as a tuple of integers, sorted. """ oslots = self.oslots return tuple(sorted(oslots.get(node, []))) def feature(self, node, **features): """Adds **node features**. ``` cv.feature(n, name=value, ... , name=value) ``` Parameters ---------- node: tuple A node reference, obtained by one of the actions `slot` or `node`. features: keyword arguments The names and values of features to assign to this node. Returns ------- None !!! caution "None values" If a feature value is `None` it will not be added! """ nodeFeatures = self.nodeFeatures self.stats[self.F] += 1 for (k, v) in features.items(): if v is None: continue # self._checkType(k, v, self.N) nodeFeatures[k][node] = v def edge(self, nodeFrom, nodeTo, **features): """Adds **edge features**. ``` cv.edge(nf, nt, name=value, ... , name=value) ``` Parameters ---------- nodeFrom, nodeTo: tuple Two node references, obtained by one of the actions `slot` or `node`. features: keyword arguments The names and values of features to assign to this edge (i.e. pair of nodes). Returns ------- None !!! note "None values" You may pass values that are `None`, and a corresponding edge will be created. If for all edges the value is `None`, an edge without values will be created. For every `nodeFrom`, such a feature essentially specifies a set of nodes `{ nodeTo }`. """ edgeFeatures = self.edgeFeatures self.stats[self.E] += 1 for (k, v) in features.items(): # self._checkType(k, v, self.E) edgeFeatures[k][nodeFrom][nodeTo] = v def occurs(self, feat): """Whether the feature `featureName` occurs in the resulting data so far. ``` occurs = cv.occurs(featureName) ``` If you have assigned None values to a feature, that will count, i.e. that feature occurs in the data. If you add feature values conditionally, it might happen that some features will not be used at all. For example, if your conversion produces errors, you might add the error information to the result in the form of error features. Later on, when the errors have been weeded out, these features will not occur any more in the result, but then TF will complain that such is feature is declared but not used. At the end of your director you can remove unused features conditionally, using this function. Parameters ---------- feat: string The name of a feature Returns ------- boolean """ nodeFeatures = self.nodeFeatures edgeFeatures = self.edgeFeatures if feat in nodeFeatures or feat in edgeFeatures: return True return False def meta(self, feat, **metadata): """Adds, modifies, deletes metadata fields of features. ``` cv.meta(feature, name=value, ... , name=value) ``` Parameters ---------- feat: string The name of a feature metadata: dict If a `value` is `None`, that `name` will be deleted from the metadata fields of the feature. A bare `cv.meta(feature)` will remove the all metadata from the feature. If you modify the field `valueType` of a feature, that feature will be added or removed from the set of `intFeatures`. It will be checked whether you specify either `int` or `str`. Returns ------- None """ errors = self.errors intFeatures = self.intFeatures metaData = self.metaData featMeta = metaData.get(feat, {}) good = True if not metadata: if feat in metaData: del metaData[feat] intFeatures.discard(feat) for (field, text) in metadata.items(): if text is None: if field == "valueType": errors['did not delete metadata field "valueType"'].append(feat) good = False else: if feat in metaData and field in metaData[feat]: del metaData[feat][field] else: metaData.setdefault(feat, {})[field] = text if field == "valueType": if text == "int": intFeatures.add(feat) else: intFeatures.discard(feat) self.good = self._checkFeatMeta(feat, featMeta) and good and self.good def features(self): """Gets the list of all features. ``` featureNames = cv.features() ``` Returns ------- list """ nodeFeatures = self.nodeFeatures edgeFeatures = self.edgeFeatures return sorted(list(nodeFeatures) + list(edgeFeatures)) def active(self, node): """Returns whether a node is currently active. Active nodes are the nodes in the set of current embedders. ``` isActive = cv.active(n) ``` If you construct your nodes in a very dynamic way, it might be hard to keep track for each node whether it has been created, terminated, or resumed, in other words, whether it is active or not. This action is provides a direct and precise way to know whether a node is active. Parameters ---------- node: tuple A node reference, obtained by one of the actions `slot` or `node`. Returns ------- boolean """ return node in self.curEmbedders def activeNodes(self, nTypes=None): """The currently active nodes, i.e. the embedders. ``` nodes = cv.activeNodes() nodes = cv.activeNodes(nTypes=("sentence", "clause")) ``` Parameters ---------- nTypes: iterable optional None If None, all active nodes are returned. Else the iterable lists a few node types, and only active nodes in these types are returned. Returns ------- set """ if nTypes is None: return set(self.curEmbedders) nTypes = set(nTypes) return {n for n in self.curEmbedders if n[0] in nTypes} def activeTypes(self): """The node types of the currently active nodes, i.e. the embedders. ``` nTypes = cv.activeTypes() ``` Parameters ---------- None Returns ------- set """ return {node[0] for node in self.curEmbedders} def get(self, feature, *args): """Retrieves feature values. ``` cv.get(feature, n) and cv.get(feature, nf, nt) ``` `feature` is the name of the feature. For node features, `n` is the node which carries the value. For edge features, `nf, nt` is the pair of from-to nodes which carries the value. Parameters ---------- feature: string The name of a feature node: tuple A node reference, obtained by one of the actions `slot` or `node`. The node in question when retrieving the value of a node feature. nodeFrom, nodeTo: tuple Two node references, obtained by one of the actions `slot` or `node`. The nodes in question when retrieving the value of an edge feature. Returns ------- string or integer """ errors = self.errors nodeFeatures = self.nodeFeatures edgeFeatures = self.edgeFeatures nArgs = len(args) if nArgs == 0 or nArgs > 2: errors["use `cv.get(ft, n)` or `cv.get(ft, nf, nt)`"].append(None) return None return ( nodeFeatures.get(feature, {}).get(args[0], None) if len(args) == 1 else edgeFeatures.get(feature, {}).get(args[0], {}).get(args[1], None) ) def _checkSecLevel(self, node, before=True): levelFromSection = self.levelFromSection sectionFeatures = self.sectionFeatures nodeFeatures = self.nodeFeatures warnings = self.warnings curEmbedders = self.curEmbedders (nType, seq) = node msg = "starts" if before is True else "ends" if before is False else "resumes" if levelFromSection: level = levelFromSection.get(nType, None) if level is None: return headingFeature = sectionFeatures[level - 1] nHeading = nodeFeatures.get(headingFeature, {}).get(node, "??") for em in curEmbedders: eType = em[0] if eType in levelFromSection: eLevel = levelFromSection.get(eType, None) eHeadingFeature = sectionFeatures[eLevel - 1] eHeading = nodeFeatures.get(eHeadingFeature, {}).get(em, "??") if eType == nType: warnings[ f'section {nType} "{nHeading}" of level {level}' f" enclosed in another {nType}: {eHeading}" ].append(None) elif eType in levelFromSection: eLevel = levelFromSection[eType] if eLevel > level: warnings[ f'section {nType} "{nHeading}" of level {level} {msg}' f' inside a {eType} "{eHeading}" of level {eLevel}' ].append(None) def _follow(self, director): # after node = yield ('N', nodeType) all slot nodes that are yielded # will be linked to node, until a ('T', node) is yielded. # If needed, you can resume this node again, after which new slot nodes # continue to be linked to this node. # If you resume a slot node, it all non slot nodes in the current context # will be linked to it. silent = self.silent tmObj = self.TF.tmObj info = tmObj.info if not self.good: return info("Following director... ", force=silent != DEEP) slotType = self.slotType errors = self.errors self.oslots = collections.defaultdict(set) self.nodeFeatures = collections.defaultdict(dict) self.edgeFeatures = collections.defaultdict( lambda: collections.defaultdict(dict) ) self.nodes = collections.defaultdict(set) nodes = self.nodes self.curSeq = collections.Counter() self.curEmbedders = set() curEmbedders = self.curEmbedders self.stats = { actionType: 0 for actionType in (self.S, self.N, self.T, self.R, self.D, self.F, self.E) } director(self) if not self.stats: self.good = False return for (actionType, amount) in sorted(self.stats.items()): info(f'"{actionType}" actions: {amount}', force=silent != DEEP) totalNodes = 0 for (nType, lastSeq) in sorted(self.curSeq.items()): for seq in range(1, lastSeq + 1): nodes[nType].add(seq) slotRep = " = slot type" if nType == slotType else "" info(f'{lastSeq:>8} x "{nType}" node {slotRep}', tm=False) totalNodes += lastSeq info(f"{totalNodes:>8} nodes of all types", tm=False, force=silent != DEEP) self.totalNodes = totalNodes if curEmbedders: embedCount = collections.Counter() for (nType, seq) in curEmbedders: embedCount[nType] += 1 for (nType, amount) in sorted( embedCount.items(), key=lambda x: (-x[1], x[0]), ): errors["Unterminated nodes"].append(f"{nType}: {amount} x") self._showErrors() def _removeUnlinked(self): silent = self.silent tmObj = self.TF.tmObj info = tmObj.info indent = tmObj.indent if not self.good and not self.force: return nodeTypes = self.curSeq nodes = self.nodes slotType = self.slotType oslots = self.oslots discardables = self.discardables nodeFeatures = self.nodeFeatures edgeFeatures = self.edgeFeatures unlinked = {} nDeleted = 0 for (nType, seq) in discardables: unlinked.setdefault(nType, set()).add(seq) nDeleted += 1 if (nType, seq) in oslots: del oslots[(nType, seq)] for nType in nodeTypes: if nType == slotType: continue for seq in range(1, nodeTypes[nType] + 1): if (nType, seq) not in oslots: unlinked.setdefault(nType, set()).add(seq) if unlinked: info( f"Removing unlinked (of which {nDeleted} deleted) nodes ... ", force=silent != DEEP, ) indent(level=2) totalRemoved = 0 for (nType, seqs) in unlinked.items(): seqs = sorted(seqs) theseNodes = nodes[nType] lSeqs = len(seqs) totalRemoved += lSeqs rep = " ..." if lSeqs > 5 else "" pl = "" if lSeqs == 1 else "s" info( f'{lSeqs:>6} unlinked "{nType}" node{pl}: {seqs[0:5]}{rep}', force=silent != DEEP, ) for seq in seqs: node = (nType, seq) theseNodes.discard(seq) for (f, fData) in nodeFeatures.items(): if node in fData: del fData[node] for (f, fData) in edgeFeatures.items(): if node in fData: del fData[node] for (fNode, toValues) in fData: if node in toValues: del toValues[node] pl = "" if totalRemoved == 1 else "s" info(f"{totalRemoved:>6} unlinked node{pl}", force=silent != DEEP) self.totalNodes -= totalRemoved info(f"Leaving {self.totalNodes:>6} nodes", force=silent != DEEP) indent(level=1) def _checkGraph(self): silent = self.silent tmObj = self.TF.tmObj info = tmObj.info if not self.good and not self.force: return info("checking for nodes and edges ... ", force=silent != DEEP) nodes = self.nodes errors = self.errors edgeFeatures = self.edgeFeatures # edges refer to nodes for (k, featureData) in edgeFeatures.items(): for nFrom in featureData: (nType, seq) = nFrom if nType not in nodes or seq not in nodes[nType]: errors["Edge feature: illegal node"].append( f'"{k}": from-node {nFrom} not in node set' ) continue for nTo in featureData[nFrom]: (nType, seq) = nTo if nType not in nodes or seq not in nodes[nType]: errors["Edge feature: illegal node"].append( f'"{k}": to-node {nTo} not in node set' ) self._showErrors() def _checkFeatures(self): silent = self.silent tmObj = self.TF.tmObj info = tmObj.info if not self.good and not self.force: return info("checking (section) features ... ", force=silent != DEEP) intFeatures = self.intFeatures metaData = self.metaData nodes = self.nodes nodeFeatures = self.nodeFeatures edgeFeatures = self.edgeFeatures errors = self.errors for feat in intFeatures: if ( feat not in WARP and feat not in nodeFeatures and feat not in edgeFeatures ): errors["intFeatures"].append( f'"{feat}" is declared as integer valued, ' "but this feature does not occur" ) for nType in self.sectionTypes: if nType not in nodes: errors["sections"].append( f'node type "{nType}" is declared as a section type, ' "but this node type does not occur" ) for feat in self.sectionFeatures: if feat not in nodeFeatures: errors["sections"].append( f'"{feat}" is declared as a section feature, ' "but this node feature does not occur" ) for nType in self.structureTypes: if nType not in nodes: errors["structure"].append( f'node type "{nType}" is declared as a structure type,' f" but this node type does not occur" ) for feat in self.structureFeatures: if feat not in nodeFeatures: errors["structure"].append( f'"{feat}" is declared as a structure feature, ' "but this node feature does not occur" ) nodeFeatures[feat] = {} sectionSet = self.sectionSet structureSet = self.structureSet featFromSectionType = self.featFromSectionType featFromStructureType = self.featFromStructureType for nType in nodes: if nType in structureSet: feat = featFromStructureType[nType] for seq in nodes[nType]: if (nType, seq) not in nodeFeatures[feat]: errors["structure features"].append( f'"structure element "{nType}" {seq} ' f'has no value for "{feat}"' ) if nType in sectionSet: feat = featFromSectionType[nType] for seq in nodes[nType]: if (nType, seq) not in nodeFeatures[feat]: errors["section features"].append( f'"section element "{nType}" {seq} ' f'has no value for "{feat}"' ) for feat in self.textFeatures: if feat not in nodeFeatures: errors["text formats"].append( f'"{feat}" is used in a text format, ' "but this node feature does not occur" ) for feat in WARP: if feat in nodeFeatures or feat in edgeFeatures: errors[feat].append(f'Do not construct the "{feat}" feature yourself') for feat in sorted(nodeFeatures) + sorted(edgeFeatures): if feat not in self.metaData: errors["feature metadata"].append( f'node feature "{feat}" has no metadata' ) for feat in sorted(metaData): if ( feat and feat not in WARP and feat not in nodeFeatures and feat not in edgeFeatures ): errors["feature metadata"].append( f'node feature "{feat}" has metadata but does not occur' ) for (feat, featData) in sorted(nodeFeatures.items()): if None in featData: errors["feature values assigned to None"].append( f'node feature "{feat}" has a node None' ) for (feat, featData) in sorted(edgeFeatures.items()): if None in featData: errors["feature values assigned to None"].append( f'edge feature "{feat}" has a from-node None' ) for toValues in featData.values(): if None in toValues: errors["feature values assigned to None"].append( f'edge feature "{feat}" has a to-node None' ) for (feat, featData) in sorted(edgeFeatures.items()): if feat in WARP: continue hasValues = False for (nodeTo, toValues) in featData.items(): if any(v is not None for v in toValues.values()): hasValues = True break if not hasValues: edgeFeatures[feat] = { nodeTo: set(toValues) for (nodeTo, toValues) in featData.items() } metaData.setdefault(feat, {})["edgeValues"] = hasValues for feat in intFeatures: if feat in WARP: continue if feat in nodeFeatures: featData = nodeFeatures[feat] for (k, v) in featData.items(): if not isInt(v): (nType, node) = k errors["Not a number"].append( f'"node feature "{feat}": {nType} {node} => "{v}"' ) if feat in edgeFeatures and metaData[feat]["edgeValues"]: featData = edgeFeatures[feat] for (fromNode, toValues) in featData.items(): (fType, fNode) = fromNode for (toNode, v) in toValues.items(): (tType, tNode) = toNode if not isInt(v): errors["Not a number"].append( f'"edge feature "{feat}":' f' {fType} {fNode} ="{v}"=> {tType} {tNode}' ) self._showErrors() def _reorderNodes(self): silent = self.silent tmObj = self.TF.tmObj info = tmObj.info if not self.good and not self.force: return info("reordering nodes ...", force=silent != DEEP) nodeTypes = self.curSeq nodes = self.nodes oslots = self.oslots slotType = self.slotType slotKeys = self.slotKeys nTypes = (slotType,) + tuple( sorted(nType for nType in nodes if nType != slotType) ) self.nodeMap = {} self.maxSlot = nodeTypes[slotType] nodeMap = self.nodeMap maxSlot = self.maxSlot newN = 0 # we build a node map that we use later on to remap the features # However, we do not have to remap the otype feature. # And we only have to remap the oslots feature if the order of the slots # is changed. # first we reorder the slots, if needed. As a consequence, we have to # 1. map the old slot nodes to the new slot nodes # 2. adapt the oslots: each node must be linked to the new slots # Because of the node map, _reassign features will adjust the other # node and edge features for the slots sortedSeqs = range(1, maxSlot + 1) if len(slotKeys) > 0: info(f'Sorting {maxSlot} slots (node type "{slotType}")') sortedSeqs = sorted(sortedSeqs, key=lambda x: (slotKeys.get(x, ""), x)) else: info("No slot sorting needed") for seq in sortedSeqs: newN += 1 nodeMap[(slotType, seq)] = newN # now we adapt the oslots feature if len(slotKeys) > 0: self.oslots = { n: {nodeMap[(slotType, s)] for s in slots} for (n, slots) in oslots.items() } oslots = self.oslots for nType in nTypes: canonical = self._canonical(nType) if nType == slotType: continue seqs = nodes[nType] info(f'Sorting {len(seqs)} nodes of type "{nType}"') sortedSeqs = sorted(seqs, key=canonical) for seq in sortedSeqs: newN += 1 nodeMap[(nType, seq)] = newN self.maxNode = newN info(f"Max node = {newN}", force=silent != DEEP) self._showErrors() def _canonical(self, nType): oslots = self.oslots def before(nodeA, nodeB): slotsA = oslots[(nType, nodeA)] slotsB = oslots[(nType, nodeB)] if slotsA == slotsB: return 0 aWithoutB = slotsA - slotsB if not aWithoutB: return 1 bWithoutA = slotsB - slotsA if not bWithoutA: return -1 aMin = min(aWithoutB) bMin = min(bWithoutA) return -1 if aMin < bMin else 1 return functools.cmp_to_key(before) def _reassignFeatures(self): silent = self.silent tmObj = self.TF.tmObj info = tmObj.info indent = tmObj.indent if not self.good and not self.force: return info("reassigning feature values ...", force=silent != DEEP) nodeMap = self.nodeMap oslots = self.oslots nodeFeatures = self.nodeFeatures edgeFeatures = self.edgeFeatures otype = {n: nType for ((nType, seq), n) in nodeMap.items()} oslots = {nodeMap[node]: slots for (node, slots) in oslots.items()} nodeFeaturesProto = self.nodeFeatures edgeFeaturesProto = self.edgeFeatures nodeFeatures = collections.defaultdict(dict) edgeFeatures = collections.defaultdict(lambda: collections.defaultdict(dict)) indent(level=2) for k in sorted(nodeFeaturesProto): featureDataProto = nodeFeaturesProto[k] ln = len(featureDataProto) pl = "" if ln == 1 else "s" info(f'node feature "{k}" with {ln} node{pl}', tm=False) featureData = {} for (node, value) in featureDataProto.items(): featureData[nodeMap[node]] = value nodeFeatures[k] = featureData for k in sorted(edgeFeaturesProto): featureDataProto = edgeFeaturesProto[k] ln = len(featureDataProto) pl = "" if ln == 1 else "s" info(f'edge feature "{k}" with {ln} start node{pl}', tm=False) featureData = {} for (nodeFrom, toValues) in featureDataProto.items(): if type(toValues) is set: toData = {nodeMap[nodeTo] for nodeTo in toValues} else: toData = {} for (nodeTo, value) in toValues.items(): toData[nodeMap[nodeTo]] = value featureData[nodeMap[nodeFrom]] = toData edgeFeatures[k] = featureData nodeFeatures["otype"] = otype edgeFeatures["oslots"] = oslots indent(level=1) self.oslots = None self.otype = None self.nodeFeatures = nodeFeatures self.edgeFeatures = edgeFeatures self._showErrors()Class variables
var D-
The type of the None singleton.
var E-
The type of the None singleton.
var F-
The type of the None singleton.
var N-
The type of the None singleton.
var R-
The type of the None singleton.
var S-
The type of the None singleton.
var T-
The type of the None singleton.
Methods
def active(self, node)-
Returns whether a node is currently active.
Active nodes are the nodes in the set of current embedders.
isActive = cv.active(n)If you construct your nodes in a very dynamic way, it might be hard to keep track for each node whether it has been created, terminated, or resumed, in other words, whether it is active or not.
This action is provides a direct and precise way to know whether a node is active.
Parameters
node:tuple- A node reference, obtained by one of the actions
slotornode.
Returns
boolean
def activeNodes(self, nTypes=None)-
The currently active nodes, i.e. the embedders.
nodes = cv.activeNodes() nodes = cv.activeNodes(nTypes=("sentence", "clause"))Parameters
nTypes:iterable optional None- If None, all active nodes are returned. Else the iterable lists a few node types, and only active nodes in these types are returned.
Returns
set
def activeTypes(self)-
The node types of the currently active nodes, i.e. the embedders.
nTypes = cv.activeTypes()Parameters
None
Returns
set
def delete(self, node)-
deletes a node.
cv.delete(n)The node
nwill be deleted from the set of nodes that will be created. Thisnmust be the result of a previouscv.node()action. slots cannot be deleted.Parameters
node:tuple- A node reference, obtained by the actions
node.
Returns
None
def edge(self, nodeFrom, nodeTo, **features)-
Adds edge features.
cv.edge(nf, nt, name=value, ... , name=value)Parameters
nodeFrom,nodeTo:tuple- Two node references, obtained by one of the actions
slotornode. features:keyword arguments- The names and values of features to assign to this edge (i.e. pair of nodes).
Returns
None
None values
You may pass values that are
None, and a corresponding edge will be created. If for all edges the value isNone, an edge without values will be created. For everynodeFrom, such a feature essentially specifies a set of nodes{ nodeTo }. def feature(self, node, **features)-
Adds node features.
cv.feature(n, name=value, ... , name=value)Parameters
node:tuple- A node reference, obtained by one of the actions
slotornode. features:keyword arguments- The names and values of features to assign to this node.
Returns
None
None values
If a feature value is
Noneit will not be added! def features(self)-
Gets the list of all features.
featureNames = cv.features()Returns
list
def get(self, feature, *args)-
Retrieves feature values.
cv.get(feature, n) and cv.get(feature, nf, nt)featureis the name of the feature.For node features,
nis the node which carries the value.For edge features,
nf, ntis the pair of from-to nodes which carries the value.Parameters
feature:string- The name of a feature
node:tuple- A node reference, obtained by one of the actions
slotornode. The node in question when retrieving the value of a node feature. nodeFrom,nodeTo:tuple- Two node references, obtained by one of the actions
slotornode. The nodes in question when retrieving the value of an edge feature.
Returns
stringorinteger
def link(self, node, slots)-
Links the given, existing slots to a node.
cv.link(n, [s1, s2])Sometimes the automatic linking of slots to nodes is not sufficient.
This happens when you feel the need to construct a node retro-actively, when the slots that need to be linked to it have already been created.
This action is precisely meant for that.
Parameters
node:tuple- A node reference, obtained by one of the actions
slotornode. slots:iterableofinteger
Returns
boolean
def linked(self, node)-
Returns the slots
ssto which a node is currently linked.ss = cv.linked(n)If you construct non-slot nodes without linking them to slots, they will be removed when TF validates the collective result of the action methods.
If you want to prevent that, you can insert an extra slot, but in order to do so, you have to detect that a node is still unlinked.
This action is precisely meant for that.
Parameters
node:tuple- A node reference, obtained by one of the actions
slotornode.
Returns
tupleofinteger
The slots are returned as a tuple of integers, sorted.
def meta(self, feat, **metadata)-
Adds, modifies, deletes metadata fields of features.
cv.meta(feature, name=value, ... , name=value)Parameters
feat:string- The name of a feature
metadata:dict- If a
valueisNone, thatnamewill be deleted from the metadata fields of the feature. A barecv.meta(feature)will remove the all metadata from the feature. If you modify the fieldvalueTypeof a feature, that feature will be added or removed from the set ofintFeatures. It will be checked whether you specify eitherintorstr.
Returns
None
def node(self, nType, slots=None)-
Makes a non-slot node and return the handle to it in
n.n = cv.node(nodeType)You have to pass its node type, i.e. a string. Think of
sentence,paragraph,phrase,word,sign, whatever.There are two modes for this function:
- Auto: (
slots=None): Non slot nodes will be automatically added to the set of embedders. - Explicit: (
slots=iterable): The slots in iterable will be assigned to this node and nothing else. The node will not be added to the set of embedders. Put otherwise: the node will be terminated after construction. However: you could resume it later to add other slots.
Remember that you can add features to the node by later
cv.feature(n, key=value, ...)calls.
Parameters
nType:string- A node type, not the slot type
slots:iterableofinteger, optionalNone- The slots to assign to this node.
If left out, the node is left as an embedding node and
subsequent slots will be added to it automatically.
All slots in the iterable must have been generated before
by means of the
cv.slot()action.
Returns
node referenceorNone- If an error occurred,
Noneis returned. The node reference consists of a node type and a sequence number, but normally you do not have to dig these out. Just pass the tuple as a whole to actions that require a node argument.
- Auto: (
def occurs(self, feat)-
Whether the feature
featureNameoccurs in the resulting data so far.occurs = cv.occurs(featureName)If you have assigned None values to a feature, that will count, i.e. that feature occurs in the data.
If you add feature values conditionally, it might happen that some features will not be used at all. For example, if your conversion produces errors, you might add the error information to the result in the form of error features.
Later on, when the errors have been weeded out, these features will not occur any more in the result, but then TF will complain that such is feature is declared but not used. At the end of your director you can remove unused features conditionally, using this function.
Parameters
feat:string- The name of a feature
Returns
boolean
def resume(self, node)-
resumes a node.
cv.resume(n)If you resume a non-slot node, you add it again to the set of embedders. No new node will be created.
If you resume a slot node, it will be added to the set of current embedders. No new slot will be created.
Parameters
node:tuple- A node reference, obtained by one of the actions
slotornode.
Returns
None
def slot(self, key=None)-
Makes a slot node and return the handle to it in
n.n = cv.slot()No further information is needed. Remember that you can add features to the node by later
cv.feature(n, key=value, ...)calls.
Parameters
key:string, optionalNone-
If passed, it acts as a sort key on the slot. At the end of the walk, all slots will be sorted by their key and then by their original order. Care will be taken that slots retain their features and linkages.
Keys are strings
Note that the key must be a string. If you want to sort on numbers, make sure to pad all numbers with leading zeros.
Returns
node reference: tuple- The node reference consists of a node type and a sequence number, but normally you do not have to dig these out. Just pass the tuple as a whole to actions that require a node argument.
def stop(self, msg)-
Stops the director. No further input will be read.
cv.stop(msg)The director will exit with a non-good status and issue the message
msg. If you have calledwalk()withforce=True, indicating that the director must proceed after errors, then this stop command will cause termination nevertheless.Parameters
msg:string- A message to display upon stopping.
Returns
None
def terminate(self, node)-
terminates a node.
cv.terminate(n)The node
nwill be removed from the set of current embedders. Thisnmust be the result of a previouscv.slot()orcv.node()action.Parameters
node:tuple- A node reference, obtained by one of the actions
slotornode.
Returns
None
def walk(self, director, slotType, otext={}, generic={}, intFeatures=set(), featureMeta={}, warn=True, generateTf=True, force=False)-
Asks a director function to walk through source data and receives its actions.
The
directorfunction should unravel the source. You have to program thedirector, which takes one argument:cv. From thecvyou can use a few standard actions that instruct TF to build a graph.This function will check whether the metadata makes sense and is minimally complete.
During node creation the section structure will be watched, and you will be warned if irregularities occur.
After the creation of the feature data, some extra checks will be performed to see whether the metadata matches the data and vice versa.
If the slots need to be sorted by their keys, it will happen at this point, and the generated features will be adapted to the sorted slots.
The new feature data will be written to the output directory of the underlying TF object. In fact, the rules are exactly the same as for
FabricCore.save().Parameters
slotType:string- The node type that acts as the type of the slots in the data set.
oText:dict-
The configuration information to be stored in the
otextfeature (seetf.core.text):- section types
- section features
- structure types
- structure features
- text formats
generic:dict-
Metadata that will be written into the header of all generated TF features.
You can make changes to this later on, dynamically in your director.
intFeatures:iterable-
The set of features that have integer values only.
You can make changes to this later on, dynamically in your director.
featureMeta:dictofdict-
For each node or edge feature descriptive metadata can be supplied.
You can make changes to this later on, dynamically in your director.
warn:boolean, optionalTrue-
This regulates the response to warnings:
True(default): stop after warnings (as if they are errors);Falsecontinue after warnings but do show them;Nonesuppress all warnings. force:boolean, optionalFalse- This forces the process to continue after errors.
Your TF set might not be valid.
Yet this can be useful during testing, when you know
that not everything is OK, but you want to check some results.
Especially when dealing with large datasets, you might want to test
with little pieces. But then you get a kind of non-fatal errors that
stand in the way of testing. For those cases:
force=True. generateTf:boolean, optionalTrue- You can pass
Falsehere to suppress the actual writing of TF data. In that way you can dry-run the director to check for errors and warnings director:function-
An ordinary function that takes one argument, the
cvobject, and should not deliver anything.Writing this function is the main job to do when you want to convert a data source to TF.
See
tf.convert.walkerfor more details.
Returns
boolean- Whether the operation was successful