Module tf.dataset.modify
Modify
Usage
from tf.dataset import modify
modify(
location,
targetLocation,
mergeFeatures=None,
deleteFeatures=None,
addFeatures=None,
mergeTypes=None,
deleteTypes=None,
addTypes=None,
replaceSlotType=None,
featureMeta=None,
silent="deep",
)
Functions
def modify(location,
targetLocation,
targetVersion=None,
mergeFeatures=None,
deleteFeatures=None,
addFeatures=None,
mergeTypes=None,
deleteTypes=None,
addTypes=None,
replaceSlotType=None,
featureMeta=None,
silent='auto')-
Expand source code Browse git
def modify( location, targetLocation, targetVersion=None, mergeFeatures=None, deleteFeatures=None, addFeatures=None, mergeTypes=None, deleteTypes=None, addTypes=None, replaceSlotType=None, featureMeta=None, silent=SILENT_D, ): """Modifies the supply of node types and features in a single data set. Dependent on the presence of the parameters, the following steps will be executed before the result is written out as a new TF dataset: * merge existing features into an other feature, removing the features that went in; * delete any number of existing features; * add any number of features and their data; * merge existing node types into a new one, removing the types that went in, without loss of nodes; So far, no new nodes have been added or removed. But then: * delete any number of node types with their nodes; * add any number of new node types, with nodes and features. The last two actions lead to a shifting of nodes, and all features that map them, will be shifted accordingly. After all that, there is one remaining action that could be performed You can also pass meta data to be merged in. Finally, the resulting features will be written to disk. !!! hint "Only added / merged features" It is possible to output only the added and merged features instead of a complete dataset. Just pass the boolean value `True` to `deleteFeatures` below. Parameters ---------- location: string You can pass just the location of the original dataset in the file system, i.e. the directory that contains the `.tf` files. targetLocation: string The directory into which the result dataset will be written. targetVersion: string, optional None If given, the new version that will be written to the metadata of each feature in the modified dataset. If None, no version modification takes place. mergeFeatures: dict, optional None You can merge several features into one. This is especially useful if there are many features each operating on different node types, and you want to unify them into one feature. The situation may occur that several of the features to be merged supply conflicting values for a node. Then the last feature in the merge list wins. The result feature may exist already. Also then there is a risk of conflict. Again, the merge result wins. Example ------- mergeFeatures=dict( resultFeature1=[feat1, feat2], resultFeature2="feat3, feat4", ), If the resulting feature is new, or needs a new description, you can provide metadata in the `featureMeta` argument. For new features you may want to set the `valueType`, although we try hard to deduce it from the data available. deleteFeatures: boolean | string | iterable, optional None This should be either a boolean value `True` or an iterable or space/comma separated string of features that you want to delete from the result. `True` means: all features will be deleted that are not the result of merging or adding features (see `mergeFeatures` above and `addFeatures` below. addFeatures: dict, optional None You can add as many features as you want, assigning values to all types, including new nodes of new types that have been generated in the steps before. You can also use this parameter to override existing features: if a feature that you are adding already exists, the new data will be merged in, overriding assignments of the existing feature if there is a conflict. The meta data of the old and new feature will also be merged. This parameter must have this shape: Example ------- addFeatures=dict( nodeFeatures=dict( feat1=data1, feat2=data2, ), edgeFeatures=dict( feat3=data3, feat4=data4, ), If the resulting features are new, or need a new description, you can provide metadata in the `featureMeta` argument. For new features you may want to set the `valueType`, although we try hard to deduce it from the data available. mergeTypes: dict, optional None You can merge several node types into one. The merged node type will have the union of nodes of the types that are merged. All relevant features will stay the same, except the `otype` feature of course. You can pass additional information to be added as features to nodes of the new node type. These features can be used to discriminate between the merged types. This parameter must have this shape: Example ------- mergeTypes=dict( outTypeA=( 'inType1', 'inType2', ), outTypeB="inType3, inType4", ) Example ------- mergeTypes=dict( outTypeA=dict( inType1=dict( featureI=valueI, featureK=valueK, ), inType2=dict( featureL=valueL, featureM=valueM, ), ), outTypeB=dict( inType3=dict( featureN=valueN, featureO=valueO, ), inType4=dict( featureP=valueP, featureQ=valueQ, ), ), ) It does not matter if these types and features already occur. The `outTypes` may be existing types of really new types. The new features may be existing or new features. Do not forget to provide meta data for new features in the `featureMeta` argument. This will migrate nodes of type `inType1` or `inType2` to nodes of `outTypeA`. In the extended form, when there are feature specifications associated with the old types, after merging the following assignments will be made: `featureI = valueI` to nodes coming from `inType1` and `featureK = valueK` to nodes coming from `inType2`. No nodes will be removed! !!! caution "slot types" Merging is all about non-slot types. It is an error if a new type or an old type is a slot type. deleteTypes: string | iterable, optional None You can delete node types from the result altogether. You can specify a list of node types as an iterable or as a space separated string. If a node type has to be deleted, all nodes in that type will be removed, and features that assign values to these nodes will have those assignments removed. Example ------- deleteTypes=('line', 'sentence') Example ------- deleteTypes="line sentence" !!! caution "slot types" Deleting is all about non-slot types. It is an error to attempt to delete slot type. addTypes: dict, optional None You may add as many node types as you want. Per node type that you add, you need to specify the current boundaries of that type and how all those nodes map to slots. You can also add features that assign values to those nodes: Example ------- dict( nodeType1=dict( nodeFrom=from1, nodeTo=to1, nodeSlots=slots1, nodeFeatures=nFeatures1, edgeFeatures=eFeatures1, ), nodeType2=dict( nodeFrom=from2, nodeTo=to2, nodeSlots=slots2, nodeFeatures=nFeatures2, edgeFeatures=eFeatures2, ), ), The boundaries may be completely arbitrary, so if you get your nodes from another TF data source, you do not need to align their values. If you also add features about those nodes, the only thing that matters is that the features assign the right values to the nodes within the boundaries. Assignments to nodes outside the boundaries will be ignored. The slots that you link the new nodes to, must exist in the original. You cannot use this function to add slots to your data set. !!! caution "existing node types" It is an error if a new node type already exists in the original, unless that type is meant to be deleted. !!! info "`nodeFeatures`, `edgeFeatures`" You can add any number of features. Per feature you have to provide the mapping that defines the feature. These features may be new, or they may already be present in the original data. If these features have values for nodes that are not within the boundaries of the new node type, those values will not be assigned but silently discarded. Example ------- dict( feat1=dict( n1=val1, n2=val2, ), feat2=dict( n1=val1, n2=val2, ), ), Edge features without values are specified like this: Example ------- dict( feat1=dict( n1={m1, m2}, n2={m3, m4}, ), feat2=dict( n1={m5, m6}, n2={m7, m8}, ), ), Edge features with values are specified like this: Example ------- dict( feat1=dict( n1={m1: v1, m2: v2}, n2={m3: v3, m4: v4}, ), feat2=dict( n1={m5: v5, m6: v6}, n2={m7: v7, m8: v8}, ), ), !!! info "`edgeFeatures` to nodes of other types" However, you may want to define edge features that relate the new nodes to nodes of other types. There is a limited way to do that. * the other type must be the last type that was added before the current type; * the nodes in the specification of the previously added type must be disjoint from the nodes of the currently added type. replaceSlotType: string, optional None If passed, it should be a tuple whose first member is a valid, non-slot node type. The slot type will be replaced by this node type and the original slots will be deleted. The remaining members are features that should be discarded in the process, they are typically features defined for old slot nodes that have little or no meaning for new slot nodes. Other features that are defined for old slots carry over to the corresponding new slots, but only if the new slot does not have already that feature assigned. Only the value of the first old slot that corresponds with a new slot carries over to that new slot. When the original slot type gets replaced, the slot mapping of other nodes needs to be adjusted. The new slots are a coarser division of the corpus than the old slots. It might even be the case that the new slots do not cover the corpus completely. If other nodes are linked to slots that are not covered by the new slots, these links are lost. This may lead to nodes that do not have links to slots anymore. These nodes will be lost, together with the feature values for these nodes and the edges that involve these nodes. Once the slot type is replaced, you may want to adapt the text formats in the OTEXT feature. You can do so by passing appropriate values in the `featureMeta` argument. featureMeta: dict, optional None If the features you have specified in one of the parameters above are new, do not forget to pass metadata for them in this parameter It is especially important to state the value type: Example ------- featureMeta=dict( featureI=dict( valueType='int', description='level of node' ), featureK=dict( valueType='str', description='subtype of node' ), ), You can also tweak the section / structure configuration and the text-formats that are specified in the `otext` feature. Just specify them as keys and values to the `otext` feature. The logic of tweaking meta data is this: what you provide in this parameter will be merged into existing meta data. If you want to remove a key from a feature, give it the value None. silent: string, optional tf.core.timestamp.SILENT_D See `tf.core.timestamp.Timestamp` """ TM = Timestamp() indent = TM.indent info = TM.info error = TM.error warning = TM.warning setSilent = TM.setSilent setSilent(silent) addFeatures = addFeatures or {} onlyDeliverUpdatedFeatures = False if type(deleteFeatures) is bool and deleteFeatures: deleteFeatures = set() onlyDeliverUpdatedFeatures = True deleteFeatures = set(fitemize(deleteFeatures)) mergeFeatures = mergeFeatures or {} addTypes = addTypes or {} deleteTypes = set(fitemize(deleteTypes)) mergeTypes = mergeTypes or {} featureMeta = featureMeta or {} origMaxNode = None origNodeTypes = None origNodeFeatures = None origEdgeFeatures = None origFeatures = None shift = {} shiftNeeded = False slotType = None maxNode = None nodeFeatures = {} edgeFeatures = {} deletedTypes = set() deletedFeatures = set() nodeTypes = {} nodeFeaturesOut = {} edgeFeaturesOut = {} metaDataOut = {} api = None good = True ePrefix = "" eItem = "" targetLocation = ex(targetLocation) def err(msg): nonlocal good error(f"{ePrefix}{eItem}{msg}", tm=False) good = False def warn(msg): warning(f"WARNING: {ePrefix}{eItem}{msg}", tm=False) def inf(msg): info(f"{ePrefix}{eItem}{msg}", tm=False) def meta(feat): return api.TF.features[feat].metaData def valTp(feat): return meta(feat).get(VALTP, None) def otextInfo(): orig = meta(OTEXT) custom = featureMeta.get(OTEXT, {}) combi = {} for key in set(custom) | set(orig): origVal = orig.get(key, "") customVal = custom.get(key, "") combi[key] = customVal or origVal ensureTypes = set() ensureFeatures = set() for kind in (SE_TP, ST_TP): ensureTypes |= set(itemize(combi.get(kind, ""), sep=",")) for kind in (SE_FT, ST_FT): ensureFeatures |= set(itemize(combi.get(kind, ""), sep=",")) ensureFeatures |= set(collectFormats(combi)[-1]) return (ensureTypes, ensureFeatures) def allInt(values): return all(isInt(v) for v in values) and any(True for v in values) def prepare(): nonlocal api nonlocal origNodeTypes nonlocal origFeatures nonlocal origNodeFeatures nonlocal origEdgeFeatures nonlocal origMaxNode nonlocal slotType nonlocal maxNode nonlocal shift nonlocal ePrefix nonlocal eItem indent(level=0, reset=True) info("preparing and checking ...") indent(level=1, reset=True) TF = Fabric(locations=location, silent=silent) origAllFeatures = TF.explore(silent=DEEP, show=True) origNodeFeatures = set(origAllFeatures[NODES]) origEdgeFeatures = set(origAllFeatures[EDGES]) origFeatures = origNodeFeatures | origEdgeFeatures api = TF.load("", silent=silent) if not api: return False F = api.F C = api.C slotType = F.otype.slotType origNodeTypes = {x[0]: (x[2], x[3]) for x in C.levels.data} origMaxSlot = F.otype.maxSlot origMaxNode = F.otype.maxNode maxNode = origMaxNode addedTp = set() addedFt = set() deletedTp = set() deletedFt = set() # check mergeFeatures ePrefix = "Merge features: " for outFeat, inFeats in mergeFeatures.items(): eItem = f"{outFeat}: " inFeats = fitemize(inFeats) if outFeat in WARP: err("Can not merge into standard features") continue if not inFeats: err("Nothing to merge from") continue addedFt.add(outFeat) for inFeat in inFeats: if inFeat in WARP: err(f"Can not merge from standard features: {inFeat}") continue deletedFt.add(inFeat) missingIn = {f for f in inFeats if f not in origFeatures} if missingIn: err(f"Missing features {_rep(missingIn)}") allInIsNode = all(f in origNodeFeatures for f in inFeats) allInIsEdge = all(f in origEdgeFeatures for f in inFeats) outExists = outFeat in origFeatures outIsNode = outExists and outFeat in origNodeFeatures outIsEdge = outExists and outFeat in origEdgeFeatures if outIsNode and not allInIsNode: err("Node Feature can not be merged from an edge feature") if outIsEdge and not allInIsEdge: err("Edge Feature can not be merged from a node feature") if not allInIsNode and not allInIsEdge: err("Feature can not be merged from both node and edge features") allInIsInt = all(valTp(f) == "int" for f in inFeats) correctTp = "int" if allInIsInt else "str" checkValType(outFeat, correctTp=correctTp) # check deleteFeatures ePrefix = "Delete features: " for feat in deleteFeatures: eItem = f"{feat}: " if feat in WARP: err("Can not delete standard features") continue if feat not in origFeatures: err("Not in data set") deletedFt.add(feat) # check addFeatures ePrefix = "Add features: " eItem = "" illegalKeys = set(addFeatures) - ADD_F_KEYS if illegalKeys: err(f"{_rep(illegalKeys)} unrecognized, expected {_rep(ADD_F_KEYS)}") bothFeatures = set(addFeatures.get(NFS, {})) & set(addFeatures.get(EFS, {})) if bothFeatures: err(f"{_rep(bothFeatures)}: Both node and edge features") for kind, otherKind, origSet, origSetOther in ( (NODE, EDGE, origNodeFeatures, origEdgeFeatures), (EDGE, NODE, origEdgeFeatures, origNodeFeatures), ): for feat, data in addFeatures.get(f"{kind}Features", {}).items(): eItem = f"{feat}: " if feat in WARP: err("Cannot add standard features") continue if feat in origSetOther: err(f"{kind} feature already exists as {otherKind} feature") checkValType(feat, vals=data.values()) addedFt.add(feat) # check mergeTypes ePrefix = "Merge types: " mData = {} for outType, inTypes in mergeTypes.items(): eItem = f"{outType}: " if outType == slotType: err("Result cannot be the slot type") withFeatures = type(inTypes) is dict addedTp.add(outType) for inType in inTypes: if inType == slotType: err(f"Slot type {inType} is not mergeable") continue if inType not in origNodeTypes: err(f"Cannot merge non-existing node type {inType}") continue deletedTp.add(inType) mFeatures = inTypes[inType] if withFeatures else {} for feat, val in mFeatures.items(): mData.setdefault(feat, set()).add(val) addedFt.add(feat) for feat, vals in mData.items(): eItem = f"{feat}: " checkValType(feat, vals=vals) # check deleteTypes ePrefix = "Delete types: " for nodeType in deleteTypes: eItem = f"{nodeType}: " if nodeType not in origNodeTypes: err("Not in data set") continue deletedTp.add(nodeType) # check addTypes ePrefix = "Add types: " for nodeType, typeInfo in sorted(addTypes.items()): eItem = f"{nodeType}: " illegalKeys = set(typeInfo) - ADD_T_KEYS if illegalKeys: err(f"{_rep(illegalKeys)} unrecognized, expected {_rep(ADD_T_KEYS)}") continue if nodeType in set(origNodeTypes) - deleteTypes: err("Already occurs") continue addedTp.add(nodeType) nodeSlots = typeInfo.get(NS, {}) if not nodeSlots: err("No slot information given") nF = typeInfo.get(NF, None) if not nF: err("No lower bound given") nT = typeInfo.get(NT, None) if not nT: err("No upper bound given") if nF is not None and nT is not None: unlinked = 0 badlinked = 0 for n in range(nF, nT + 1): slots = nodeSlots.get(n, ()) if not slots: unlinked += 1 else: slotGood = True for slot in slots: if slot < 1 or slot > origMaxSlot: slotGood = False if not slotGood: badlinked += 1 if unlinked: err(f"{unlinked} nodes not linked to slots") if badlinked: err(f"{badlinked} nodes linked to non-slot nodes") for kind in (NODE, EDGE): for feat, data in typeInfo.get(f"{kind}Features", {}).items(): eItem = f"{feat}: " checkValType(feat, vals=data.values()) addedFt.add(feat) (otextTypes, otextFeatures) = otextInfo() if False: # it is no problem to first delete a type and then add a type with # the same name (I think) problemTypes = addedTp & deletedTp if problemTypes: ePrefix = "Add and then delete: " eItem = "types: " err(f"{_rep(problemTypes)}") problemTypes = otextTypes - ((set(origNodeTypes) | addedTp) - deletedTp) if problemTypes: ePrefix = "Missing for text API: " eItem = "types: " warn(f"{_rep(problemTypes)}") problemFeats = addedFt & deletedFt if problemFeats: ePrefix = "Add and then delete: " eItem = "features: " err(f"{_rep(problemFeats)}") problemFeats = otextFeatures - ((origFeatures | addedFt) - deletedFt) if problemFeats: ePrefix = "Missing for text API: " eItem = "features: " warn(f"{_rep(problemFeats)}") if not dirEmpty(targetLocation): ePrefix = "Output directory: " eItem = "not empty: " err("Clean it or remove it or choose another location") if not good: return False api = TF.loadAll(silent=silent) info("done") return True def checkValType(feat, vals=None, correctTp=None): origTp = valTp(feat) if feat in origFeatures else None customTp = featureMeta.get(feat, {}).get(VALTP, None) assignedTp = origTp or customTp if correctTp is None: correctTp = "int" if allInt(vals) else "str" newTp = customTp or correctTp if newTp != assignedTp: featureMeta.setdefault(feat, {})[VALTP] = newTp if customTp and customTp != correctTp and customTp == "int": err("feature values are declared to be int but some values are not int") if assignedTp != newTp: rep1 = f"feature of type {newTp}" rep2 = f" (was {assignedTp})" if assignedTp else "" inf(f"{rep1}{rep2}") def shiftx(vs, offset=None, nF=None, nT=None, nodeMap={}): if offset is None: return ( {shift[m]: v for (m, v) in vs.items()} if type(vs) is dict else {shift[m] for m in vs} ) else: if type(vs) is dict: result = {} for (m, v) in vs.items(): if nF <= m <= nT: result[m + offset] = v elif m in nodeMap: result[nodeMap[m]] = v else: result = set() for m in vs: if nF <= m <= nT: result.add(m + offset) elif m in nodeMap: result.add(nodeMap[m]) return result def shiftFeature(kind, feat, data): return ( {shift[n]: v for (n, v) in data.items() if n in shift} if kind == NODE else {shift[n]: shiftx(v) for (n, v) in data.items() if n in shift} ) def mergeF(): nonlocal deletedFeatures Fs = api.Fs Es = api.Es indent(level=0) if mergeFeatures: info("merge features ...") indent(level=1, reset=True) inF = set() for outFeat, inFeats in mergeFeatures.items(): data = {} inFeats = fitemize(inFeats) if all(f in origNodeFeatures for f in inFeats): featSrc = Fs featDst = nodeFeatures else: featSrc = Es featDst = edgeFeatures for inFeat in inFeats: for n, val in featSrc(inFeat).data.items(): data[n] = val featDst.setdefault(outFeat, {}).update(data) for inFeat in inFeats: inF.add(inFeat) if inFeat in featDst: del featDst[inFeat] deletedFeatures |= inF if mergeFeatures: info(f"done (deleted {len(inF)} and added {len(mergeFeatures)} features)") indent(level=2) info(f"deleted {_rep(inF)}", tm=False) info(f"added {_rep(mergeFeatures)}", tm=False) return True def deleteF(): indent(level=0) if deleteFeatures: info("delete features ...") indent(level=1, reset=True) for feat in deleteFeatures: dest = ( nodeFeatures if feat in origNodeFeatures else edgeFeatures if feat in origEdgeFeatures else None ) if dest and feat in dest: del dest[feat] deletedFeatures.add(feat) if deleteFeatures: info(f"done ({len(deleteFeatures)} features)") indent(level=2) info(_rep(deleteFeatures), tm=False) return True def addF(): indent(level=0) if addFeatures: info("add features ...") indent(level=1, reset=True) added = collections.defaultdict(set) for kind, dest in ( (NODE, nodeFeatures), (EDGE, edgeFeatures), ): for feat, data in addFeatures.get(f"{kind}Features", {}).items(): dest.setdefault(feat, {}).update(data) added[kind].add(feat) if addFeatures: info( f'done (added {len(added["node"])} node + {len(added["edge"])} edge features)' ) indent(level=2) for kind, feats in sorted(added.items()): info(f"{kind} features: {_rep(feats)}") return True def mergeT(): nonlocal deletedTypes indent(level=0) if mergeTypes: info("merge types ...") indent(level=1, reset=True) inT = set() for outType, inTypes in mergeTypes.items(): info(f"Merging {outType}") withFeatures = type(inTypes) is dict for inType in inTypes: addFeatures = inTypes[inType] if withFeatures else {} addFeatures[OTYPE] = outType (nF, nT) = origNodeTypes[inType] for feat, val in addFeatures.items(): for n in range(nF, nT + 1): nodeFeatures.setdefault(feat, {})[n] = val inT.add(inType) deletedTypes |= inT if mergeTypes: info(f"done (merged {len(mergeTypes)} node types)") indent(level=2) info(f"deleted {_rep(inT)}", tm=False) info(f"added {_rep(mergeTypes)}", tm=False) return True def deleteT(): nonlocal maxNode nonlocal shiftNeeded nonlocal ePrefix nonlocal eItem ePrefix = "Delete types: " indent(level=0) if deleteTypes: info("delete types ...") indent(level=1, reset=True) curShift = 0 for nType, (nF, nT) in sorted(origNodeTypes.items(), key=lambda x: x[1][0]): eItem = f"{nType:<20}: " if nType in deleteTypes: curShift -= nT - nF + 1 deletedTypes.add(nType) inf(f"remove: delete nodes {nF:>7}-{nT:>7}") else: nodeTypes[nType] = (nF + curShift, nT + curShift) for n in range(nF, nT + 1): shift[n] = n + curShift inf( f"keep: shift nodes {nF:>7}-{nT:>7} to " f"{nF + curShift:>7}-{nT + curShift:>7}" ) for kind, upd in ( (NODE, nodeFeatures), (EDGE, edgeFeatures), ): for feat, uData in upd.items(): upd[feat] = shiftFeature(kind, feat, uData) maxNode = origMaxNode + curShift eItem = "max node: " inf(f"shifted from {origMaxNode} to {maxNode}") shiftNeeded = curShift != 0 if deleteTypes: info(f"done ({len(deleteTypes)} types)") indent(level=2) info(_rep(deleteTypes), tm=False) return True def addT(): nonlocal maxNode indent(level=0) if addTypes: info("add types ...") indent(level=1, reset=True) nodeMap = {} for nodeType, typeInfo in sorted(addTypes.items()): nF = typeInfo[NF] nT = typeInfo[NT] offset = maxNode - nF + 1 nodeSlots = typeInfo[NS] data = {} newNodeMap = {} for n in range(nF, nT + 1): data[offset + n] = nodeType newNodeMap[n] = offset + n nodeFeatures.setdefault(OTYPE, {}).update(data) data = {} for n in range(nF, nT + 1): data[offset + n] = set(nodeSlots[n]) edgeFeatures.setdefault(OSLOTS, {}).update(data) for feat, addData in typeInfo.get(NFS, {}).items(): data = {} for n in range(nF, nT + 1): value = addData.get(n, None) if value is not None: data[offset + n] = value nodeFeatures.setdefault(feat, {}).update(data) for feat, addData in typeInfo.get(EFS, {}).items(): data = {} for n in range(nF, nT + 1): value = addData.get(n, None) if value: newValue = shiftx( value, offset=offset, nF=nF, nT=nT, nodeMap=nodeMap ) if newValue: data[offset + n] = newValue edgeFeatures.setdefault(feat, {}).update(data) maxNode += nT - nF + 1 nodeMap = newNodeMap if addTypes: info(f"done ({len(addTypes)} types)") indent(level=2) info(_rep(addTypes), tm=False) return True def applyUpdates(): nonlocal good nonlocal ePrefix nonlocal eItem nonlocal replaceSlotType Fs = api.Fs Es = api.Es indent(level=0) info("applying updates ...") indent(level=1, reset=True) mFeat = 0 for kind, featSet, featSrc, featUpd, featOut in ( (NODE, origNodeFeatures, Fs, nodeFeatures, nodeFeaturesOut), (EDGE, origEdgeFeatures, Es, edgeFeatures, edgeFeaturesOut), ): ePrefix = f"Shift {kind} feature: " allFeatSet = set() if onlyDeliverUpdatedFeatures else set(featSet) for feat in sorted((allFeatSet | set(featUpd)) - deletedFeatures): eItem = f"{feat:<20}: " outData = {} outMeta = {} if feat in featSet: # inf("original feature") featObj = featSrc(feat) if feat != OSLOTS and kind == EDGE and featObj.doValues: outMeta["edgeValues"] = True outMeta.update(featObj.meta) if shiftNeeded: outData.update(shiftFeature(kind, feat, featObj)) mFeat += 1 else: outData.update(featObj.items()) if feat in featUpd: # inf("new / updated feature") outData.update(featUpd[feat]) if kind == EDGE: aVal = next(iter(featUpd[feat].values())) hasValues = type(aVal) is dict if outMeta.get("edgeValues", False) != hasValues: outMeta["edgeValues"] = hasValues if feat in featureMeta: for k, v in featureMeta[feat].items(): if v is None: if k in outMeta: del outMeta[k] else: outMeta[k] = v featOut[feat] = outData metaDataOut[feat] = outMeta if replaceSlotType: if type(replaceSlotType) is str: ignoreSlotFeatures = () else: (replaceSlotType, *ignoreSlotFeatures) = replaceSlotType ignoreSlotFeatures = set(ignoreSlotFeatures) ignoreRep = ( f" while ignoring {ignoreSlotFeatures}" if ignoreSlotFeatures else "" ) info(f"Replacing slot type {slotType} by {replaceSlotType}{ignoreRep}") # check replaceSlotType ePrefix = "Replace slot type: " eItem = f"{replaceSlotType}: " allTypes = (set(nodeTypes) - deletedTypes) | set(addTypes) if replaceSlotType not in allTypes: err("Node type does not exist") if replaceSlotType == slotType: err("Node type is already the slot type") if not good: return # map old slots to nodes of the new slot type currentOtype = nodeFeaturesOut[OTYPE] currentOslots = edgeFeaturesOut[OSLOTS] # We have to sort the nodes of the new slot type in the canonical order nextSlots = sorted( ( node for (node, nType) in currentOtype.items() if nType == replaceSlotType ), key=_canonical(currentOslots), ) # Now we walk through the nodes of the new slot type and see # to what slots they are linked. # There might be old slots that are not linked to any of these nodes. # These old slots must be removed. currentSlotMap = {} removeNodes = set() for newSlot in nextSlots: for oldSlot in currentOslots[newSlot]: currentSlotMap[oldSlot] = newSlot for oldSlot, nType in currentOtype.items(): if nType == slotType: if oldSlot not in currentSlotMap: removeNodes.add(oldSlot) info(f"{len(removeNodes)} old {slotType}s do not map to {replaceSlotType}s") # All features (except those in ignoreSlotFeatures) on old slots # have to be extended to the new slots # If a new slot has conflicting feature values for the old slots # it is linked to, the first defined feature value will be taken. # Likewise, all edge features that involve old slots # have to be extended to the new slots. for feat, featData in nodeFeaturesOut.items(): if feat == OTYPE: continue if feat in ignoreSlotFeatures: continue updates = {} for oldSlot, value in featData.items(): if oldSlot in removeNodes: continue nType = currentOtype[oldSlot] if nType != slotType: continue newSlot = currentSlotMap[oldSlot] currentVal = featData.get(newSlot, None) if currentVal is None: alreadyUpdated = updates.get(newSlot, None) if alreadyUpdated is None: updates[newSlot] = value if updates: for node, val in updates.items(): featData[node] = val for feat, featData in edgeFeaturesOut.items(): if feat == OSLOTS: continue updates = {} for fromNode, toNodes in featData.items(): if fromNode in removeNodes: continue nTypeFrom = currentOtype[fromNode] if nTypeFrom == slotType: newFromNode = currentSlotMap[fromNode] else: newFromNode = fromNode if type(toNodes) is dict: for toNode, value in toNodes.items(): if toNode in removeNodes: continue nTypeTo = currentOtype[toNode] if nTypeTo == slotType: newToNode = currentSlotMap[toNode] else: newToNode = toNode if nTypeFrom != slotType and nTypeTo != slotType: continue currentVals = featData.get(newFromNode, {}) if newToNode not in currentVals: alreadyUpdateds = updates.get(newFromNode, {}) if newToNode not in alreadyUpdateds: doUpdate = True else: alreadyVal = alreadyUpdateds[newToNode] if alreadyVal is None: doUpdate = value is not None else: doUpdate = False if doUpdate: updates.setdefault(newFromNode, {})[ newToNode ] = value else: for toNode in toNodes: if toNode in removeNodes: continue nTypeTo = currentOtype[toNode] if nTypeTo == slotType: newToNode = currentSlotMap[toNode] else: newToNode = toNode if nTypeFrom != slotType and nTypeTo != slotType: continue currentVals = featData.get(newFromNode, set()) if newToNode not in currentVals: alreadyUpdateds = updates.get(newFromNode, set()) if newToNode not in alreadyUpdateds: updates.setdefault(newFromNode, set()).add( newToNode ) if updates: for fromNode, toNodes in updates.items(): if type(toNodes) is dict: for toNode, val in toNodes.items(): featData.setdefault(fromNode, {})[toNode] = val else: for toNode in toNodes: featData.setdefault(fromNode, set()).add(toNode) # link all nodes, except slot nodes and nodes of the new slot type to # new slots. # N.B. the new slots are still ordinary nodes, so this is just # the next version of oslots, not the final version # Gather the nodes that end up unlinked to new slots. nextOslots = {} for node, slots in currentOslots.items(): newSlots = {currentSlotMap[s] for s in slots if s in currentSlotMap} if len(newSlots): nextOslots[node] = newSlots else: removeNodes.add(node) # build the final otype and oslots features # we will shuffle nodes and delete nodes, so we maintain a map newOtype = {} newOslots = {} newFromCurrent = {} newNode = 0 # first the new slots for node in nextSlots: newNode += 1 newFromCurrent[node] = newNode newOtype[newNode] = replaceSlotType # now the rest for node, nType in currentOtype.items(): if nType == slotType or nType == replaceSlotType or node in removeNodes: continue newNode += 1 newFromCurrent[node] = newNode newOtype[newNode] = nType newOslots[newNode] = {newFromCurrent[s] for s in nextOslots[node]} # now apply the node mapping to the remaining node and edge features # we may have to delete features removeNodeFeatures = set() removeEdgeFeatures = set() for feat, featData in nodeFeaturesOut.items(): if feat == OTYPE: nodeFeaturesOut[OTYPE] = newOtype continue newFeatData = {} for node, value in featData.items(): if node in removeNodes: continue if node not in newFromCurrent: continue newFeatData[newFromCurrent[node]] = value if len(newFeatData): nodeFeaturesOut[feat] = newFeatData else: removeNodeFeatures.add(feat) for feat, featData in edgeFeaturesOut.items(): if feat == OSLOTS: edgeFeaturesOut[OSLOTS] = newOslots continue newFeatData = {} for fromNode, toNodes in featData.items(): if fromNode in removeNodes: continue newFromNode = newFromCurrent[fromNode] if type(toNodes) is dict: newToNodes = {} for toNode, value in toNodes.items(): if toNode in removeNodes: continue newTNode = newFromCurrent[toNode] newToNodes[newTNode] = value else: newToNodes = set() for toNode in toNodes: if toNode in removeNodes: continue newTNode = newFromCurrent[toNode] newToNodes.add(newTNode) if not len(newToNodes): continue newFeatData[newFromNode] = newToNodes if len(newFeatData): edgeFeaturesOut[feat] = newFeatData else: removeEdgeFeatures.add(feat) # remove empty features for feat in removeNodeFeatures: del nodeFeaturesOut[feat] if feat in metaDataOut: del metaDataOut[feat] for feat in removeEdgeFeatures: del edgeFeaturesOut[feat] if feat in metaDataOut: del metaDataOut[feat] otextMeta = {} otextMeta.update(meta(OTEXT)) mK = 0 if OTEXT in featureMeta: for k, v in featureMeta[OTEXT].items(): if v is None: if k in otextMeta: del otextMeta[k] mK += 1 else: if k not in otextMeta or otextMeta[k] != v: otextMeta[k] = v mK += 1 metaDataOut[OTEXT] = otextMeta if mFeat or mK: fRep = f" (shifted {mFeat} features)" if mFeat else "" kRep = f" (adapted {mK} keys in otext)" if mK else "" info(f"done{fRep}{kRep}") return True def _canonical(oslots): def before(nodeA, nodeB): slotsA = oslots[nodeA] slotsB = oslots[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 writeTf(): indent(level=0) if targetVersion is None: versionMsg = "with the original version number" else: versionMsg = f"as version {targetVersion}" for feat in set(nodeFeaturesOut) | set(edgeFeaturesOut): metaDataOut.setdefault(feat, {})["version"] = targetVersion info(f"write TF data {versionMsg} ... ") indent(level=1, reset=True) TF = Fabric(locations=targetLocation, silent=silent) TF.save( metaData=metaDataOut, nodeFeatures=nodeFeaturesOut, edgeFeatures=edgeFeaturesOut, silent=silent, ) return True def finalize(): indent(level=0) info("all done") return True def process(): for step in ( prepare, mergeF, deleteF, addF, mergeT, deleteT, addT, applyUpdates, writeTf, finalize, ): if not step(): return False return True result = process() return resultModifies the supply of node types and features in a single data set.
Dependent on the presence of the parameters, the following steps will be executed before the result is written out as a new TF dataset:
- merge existing features into an other feature, removing the features that went in;
- delete any number of existing features;
- add any number of features and their data;
- merge existing node types into a new one, removing the types that went in, without loss of nodes;
So far, no new nodes have been added or removed. But then:
- delete any number of node types with their nodes;
- add any number of new node types, with nodes and features.
The last two actions lead to a shifting of nodes, and all features that map them, will be shifted accordingly.
After all that, there is one remaining action that could be performed
You can also pass meta data to be merged in.
Finally, the resulting features will be written to disk.
Only added / merged features
It is possible to output only the added and merged features instead of a complete dataset. Just pass the boolean value
TruetodeleteFeaturesbelow.Parameters
location:string- You can pass just the location of the original dataset in the file system,
i.e. the directory that contains the
.tffiles. targetLocation:string- The directory into which the result dataset will be written.
targetVersion:string, optionalNone- If given, the new version that will be written to the metadata of each feature in the modified dataset. If None, no version modification takes place.
mergeFeatures:dict, optionalNone-
You can merge several features into one. This is especially useful if there are many features each operating on different node types, and you want to unify them into one feature. The situation may occur that several of the features to be merged supply conflicting values for a node. Then the last feature in the merge list wins.
The result feature may exist already. Also then there is a risk of conflict. Again, the merge result wins.
Example
mergeFeatures=dict( resultFeature1=[feat1, feat2], resultFeature2="feat3, feat4", ),If the resulting feature is new, or needs a new description, you can provide metadata in the
featureMetaargument. For new features you may want to set thevalueType, although we try hard to deduce it from the data available. deleteFeatures:boolean | string | iterable, optionalNone-
This should be either a boolean value
Trueor an iterable or space/comma separated string of features that you want to delete from the result.Truemeans: all features will be deleted that are not the result of merging or adding features (seemergeFeaturesabove andaddFeaturesbelow. addFeatures:dict, optionalNone-
You can add as many features as you want, assigning values to all types, including new nodes of new types that have been generated in the steps before.
You can also use this parameter to override existing features: if a feature that you are adding already exists, the new data will be merged in, overriding assignments of the existing feature if there is a conflict. The meta data of the old and new feature will also be merged.
This parameter must have this shape:
Example
addFeatures=dict( nodeFeatures=dict( feat1=data1, feat2=data2, ), edgeFeatures=dict( feat3=data3, feat4=data4, ),If the resulting features are new, or need a new description, you can provide metadata in the
featureMetaargument. For new features you may want to set thevalueType, although we try hard to deduce it from the data available. mergeTypes:dict, optionalNone-
You can merge several node types into one. The merged node type will have the union of nodes of the types that are merged. All relevant features will stay the same, except the
otypefeature of course.You can pass additional information to be added as features to nodes of the new node type. These features can be used to discriminate between the merged types.
This parameter must have this shape:
Example
mergeTypes=dict( outTypeA=( 'inType1', 'inType2', ), outTypeB="inType3, inType4", )Example
mergeTypes=dict( outTypeA=dict( inType1=dict( featureI=valueI, featureK=valueK, ), inType2=dict( featureL=valueL, featureM=valueM, ), ), outTypeB=dict( inType3=dict( featureN=valueN, featureO=valueO, ), inType4=dict( featureP=valueP, featureQ=valueQ, ), ), )It does not matter if these types and features already occur. The
outTypesmay be existing types of really new types. The new features may be existing or new features.Do not forget to provide meta data for new features in the
featureMetaargument.This will migrate nodes of type
inType1orinType2to nodes ofoutTypeA.In the extended form, when there are feature specifications associated with the old types, after merging the following assignments will be made:
featureI = valueIto nodes coming frominType1and
featureK = valueKto nodes coming frominType2.No nodes will be removed!
slot types
Merging is all about non-slot types. It is an error if a new type or an old type is a slot type.
deleteTypes:string | iterable, optionalNone-
You can delete node types from the result altogether. You can specify a list of node types as an iterable or as a space separated string.
If a node type has to be deleted, all nodes in that type will be removed, and features that assign values to these nodes will have those assignments removed.
Example
deleteTypes=('line', 'sentence')Example
deleteTypes="line sentence"slot types
Deleting is all about non-slot types. It is an error to attempt to delete slot type.
addTypes:dict, optionalNone-
You may add as many node types as you want.
Per node type that you add, you need to specify the current boundaries of that type and how all those nodes map to slots. You can also add features that assign values to those nodes:
Example
dict( nodeType1=dict( nodeFrom=from1, nodeTo=to1, nodeSlots=slots1, nodeFeatures=nFeatures1, edgeFeatures=eFeatures1, ), nodeType2=dict( nodeFrom=from2, nodeTo=to2, nodeSlots=slots2, nodeFeatures=nFeatures2, edgeFeatures=eFeatures2, ), ),The boundaries may be completely arbitrary, so if you get your nodes from another TF data source, you do not need to align their values.
If you also add features about those nodes, the only thing that matters is that the features assign the right values to the nodes within the boundaries. Assignments to nodes outside the boundaries will be ignored.
The slots that you link the new nodes to, must exist in the original. You cannot use this function to add slots to your data set.
existing node types
It is an error if a new node type already exists in the original, unless that type is meant to be deleted.
nodeFeatures,edgeFeaturesYou can add any number of features. Per feature you have to provide the mapping that defines the feature.
These features may be new, or they may already be present in the original data.
If these features have values for nodes that are not within the boundaries of the new node type, those values will not be assigned but silently discarded.
Example
dict( feat1=dict( n1=val1, n2=val2, ), feat2=dict( n1=val1, n2=val2, ), ),Edge features without values are specified like this:
Example
dict( feat1=dict( n1={m1, m2}, n2={m3, m4}, ), feat2=dict( n1={m5, m6}, n2={m7, m8}, ), ),Edge features with values are specified like this:
Example
dict( feat1=dict( n1={m1: v1, m2: v2}, n2={m3: v3, m4: v4}, ), feat2=dict( n1={m5: v5, m6: v6}, n2={m7: v7, m8: v8}, ), ),edgeFeaturesto nodes of other typesHowever, you may want to define edge features that relate the new nodes to nodes of other types. There is a limited way to do that.
- the other type must be the last type that was added before the current type;
- the nodes in the specification of the previously added type must be disjoint from the nodes of the currently added type.
replaceSlotType:string, optionalNone-
If passed, it should be a tuple whose first member is a valid, non-slot node type. The slot type will be replaced by this node type and the original slots will be deleted. The remaining members are features that should be discarded in the process, they are typically features defined for old slot nodes that have little or no meaning for new slot nodes. Other features that are defined for old slots carry over to the corresponding new slots, but only if the new slot does not have already that feature assigned. Only the value of the first old slot that corresponds with a new slot carries over to that new slot.
When the original slot type gets replaced, the slot mapping of other nodes needs to be adjusted. The new slots are a coarser division of the corpus than the old slots. It might even be the case that the new slots do not cover the corpus completely.
If other nodes are linked to slots that are not covered by the new slots, these links are lost. This may lead to nodes that do not have links to slots anymore. These nodes will be lost, together with the feature values for these nodes and the edges that involve these nodes.
Once the slot type is replaced, you may want to adapt the text formats in the OTEXT feature. You can do so by passing appropriate values in the
featureMetaargument. featureMeta:dict, optionalNone-
If the features you have specified in one of the parameters above are new, do not forget to pass metadata for them in this parameter It is especially important to state the value type:
Example
featureMeta=dict( featureI=dict( valueType='int', description='level of node' ), featureK=dict( valueType='str', description='subtype of node' ), ),You can also tweak the section / structure configuration and the text-formats that are specified in the
otextfeature. Just specify them as keys and values to theotextfeature.The logic of tweaking meta data is this: what you provide in this parameter will be merged into existing meta data.
If you want to remove a key from a feature, give it the value None.
silent:string, optionalSILENT_D- See
Timestamp