def __init__(self, **kwargs):
"""
set _node_transformer, _edge_transformer, tdifNodeTextVectorizer
"""
FeatureDefinition.__init__(self)
nbTypes = self._getTypeNumber(kwargs)
print("BETTER FEATURES")
block_transformer = FeatureUnion( [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("xywh", Pipeline([
('selector', NodeTransformerXYWH_v2()),
#v1 ('xywh', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('xywh', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
, ("neighbors", Pipeline([
('selector', NodeTransformerNeighbors()),
#v1 ('neighbors', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('neighbors', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
, ("1hot", Pipeline([
('1hot', Node1HotFeatures()) #does the 1-hot encoding directly
])
)
])
grid_line_transformer = GridLine_NodeTransformer_v2()
self._node_transformer = TransformerListByType([block_transformer, grid_line_transformer])
edge_BB_transformer = FeatureUnion( [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot", Pipeline([
('1hot', Edge1HotFeatures(PageNumberSimpleSequenciality()))
])
)
, ("boolean", Pipeline([
('boolean', EdgeBooleanFeatures_v2())
])
)
, ("numerical", Pipeline([
('selector', EdgeNumericalSelector()),
#v1 ('numerical', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('numerical', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
] )
edge_BL_transformer = Block2GridLine_EdgeTransformer()
edge_LL_transformer = GridLine2GridLine_EdgeTransformer()
self._edge_transformer = TransformerListByType([edge_BB_transformer,
edge_BL_transformer,
edge_BL_transformer, # useless but required
edge_LL_transformer
])
self.tfidfNodeTextVectorizer = None #tdifNodeTextVectorizer
def __init__(self):
FeatureDefinition.__init__(self)
# self.n_tfidf_node, self.t_ngrams_node, self.b_tfidf_node_lc = n_tfidf_node, t_ngrams_node, b_tfidf_node_lc
# self.n_tfidf_edge, self.t_ngrams_edge, self.b_tfidf_edge_lc = n_tfidf_edge, t_ngrams_edge, b_tfidf_edge_lc
# tdifNodeTextVectorizer = TfidfVectorizer(lowercase=self.b_tfidf_node_lc, max_features=self.n_tfidf_node
# , analyzer = 'char', ngram_range=self.t_ngrams_node #(2,6)
# , dtype=np.float64)
node_transformer = FeatureUnion( [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("xywh", Pipeline([
('selector', NodeTransformerXYWH_v2()),
#v1 ('xywh', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('xywh', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
, ("neighbors", Pipeline([
('selector', NodeTransformerNeighbors()),
#v1 ('neighbors', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('neighbors', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
, ("1hot", Pipeline([
('1hot', Node1HotFeatures()) #does the 1-hot encoding directly
])
)
])
lEdgeFeature = [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot", Pipeline([
('1hot', Edge1HotFeatures(PageNumberSimpleSequenciality()))
])
)
, ("boolean", Pipeline([
('boolean', EdgeBooleanFeatures_v2())
])
)
, ("numerical", Pipeline([
('selector', EdgeNumericalSelector()),
#v1 ('numerical', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('numerical', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
]
edge_transformer = FeatureUnion( lEdgeFeature )
#return _node_transformer, _edge_transformer, tdifNodeTextVectorizer
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
self.tfidfNodeTextVectorizer = None #tdifNodeTextVectorizer
def __init__(self, **kwargs):
FeatureDefinition.__init__(self, **kwargs)
nbTypes = self._getTypeNumber(kwargs)
node_transformer = TransformerListByType([ FeatureUnion( [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("xywh", Pipeline([
('selector', NodeTransformerXYWH_v2()),
#v1 ('xywh', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('xywh', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
, ("neighbors", Pipeline([
('selector', NodeTransformerNeighbors()),
#v1 ('neighbors', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('neighbors', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
, ("1hot", Pipeline([
('1hot', Node1HotFeatures()) #does the 1-hot encoding directly
])
)
]) for _i in range(nbTypes) ])
edge_transformer = TransformerListByType([ FeatureUnion( [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot", Pipeline([
('1hot', Edge1HotFeatures(PageNumberSimpleSequenciality()))
])
)
, ("boolean", Pipeline([
('boolean', EdgeBooleanFeatures_v2())
])
)
, ("numerical", Pipeline([
('selector', EdgeNumericalSelector()),
#v1 ('numerical', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('numerical', QuantileTransformer(n_quantiles=self.n_QUANTILES, copy=False)) #use in-place scaling
])
)
] ) for _i in range(nbTypes*nbTypes) ])
#return _node_transformer, _edge_transformer, tdifNodeTextVectorizer
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
def __init__(self):
FeatureDefinition.__init__(self)
node_transformer = FeatureUnion(
[ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
(
"xywh",
Pipeline([
('selector', NodeTransformerXYWH_v2()),
('xywh',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
("neighbors", NodeTransformerNeighbors_v2()),
("1hot", Node1HotFeatures()) #does the 1-hot encoding directly
])
lEdgeFeature = [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot", Edge1HotFeatures(PageNumberSimpleSequenciality())),
('boolean', EdgeBooleanFeatures_v2()),
(
"numerical",
Pipeline([
('selector', EdgeNumericalSelector()),
('numerical',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
]))
]
edge_transformer = FeatureUnion(lEdgeFeature)
#return _node_transformer, _edge_transformer, tdifNodeTextVectorizer
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
self.tfidfNodeTextVectorizer = None #tdifNodeTextVectorizer
def __init__(self,
n_tfidf_node=None,
t_ngrams_node=None,
b_tfidf_node_lc=None,
n_tfidf_edge=None,
t_ngrams_edge=None,
b_tfidf_edge_lc=None,
bMirrorPage=True,
bMultiPage=True):
FeatureDefinition.__init__(self)
self.n_tfidf_node, self.t_ngrams_node, self.b_tfidf_node_lc = n_tfidf_node, t_ngrams_node, b_tfidf_node_lc
self.n_tfidf_edge, self.t_ngrams_edge, self.b_tfidf_edge_lc = n_tfidf_edge, t_ngrams_edge, b_tfidf_edge_lc
self.bMirrorPage = bMirrorPage
self.bMultiPage = bMultiPage
tdifNodeTextVectorizer = TfidfVectorizer(
lowercase=self.b_tfidf_node_lc,
max_features=self.n_tfidf_node,
analyzer='char',
ngram_range=self.t_ngrams_node #(2,6)
,
dtype=np.float64)
node_transformer = FeatureUnion(
[ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
(
"text",
Pipeline([
('selector', NodeTransformerTextEnclosed()),
# ('tfidf', TfidfVectorizer(lowercase=self.b_tfidf_node_lc, max_features=self.n_tfidf_node
# , analyzer = 'char', ngram_range=self.tNODE_NGRAMS #(2,6)
# , dtype=np.float64)),
(
'tfidf', tdifNodeTextVectorizer
), #we can use it separately from the pipleline once fitted
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
])),
(
"textlen",
Pipeline([
('selector', NodeTransformerTextLen()),
('textlen',
QuantileTransformer(n_quantiles=self.n_QUANTILES,
copy=False)
) #use in-place scaling
])),
(
"xywh",
Pipeline([
('selector', NodeTransformerXYWH_v2()),
#v1 ('xywh', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('xywh',
QuantileTransformer(n_quantiles=self.n_QUANTILES,
copy=False)
) #use in-place scaling
])),
(
"neighbors",
Pipeline([
('selector', NodeTransformerNeighbors()),
#v1 ('neighbors', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('neighbors',
QuantileTransformer(n_quantiles=self.n_QUANTILES,
copy=False)
) #use in-place scaling
])),
(
"1hot",
Pipeline([('1hot', Node1HotFeatures()
) #does the 1-hot encoding directly
]))
# , ('ocr' , Pipeline([
# ('ocr', NodeOCRFeatures())
# ])
# )
# , ('pnumre' , Pipeline([
# ('pnumre', NodePNumFeatures())
# ])
# )
# , ("doc_tfidf", Pipeline([
# ('zero', Zero2Features())
# #THIS ONE MUST BE LAST, because it include a placeholder column for the doculent-level tfidf
# ])
# )
])
lEdgeFeature = [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot",
Pipeline([('1hot',
Edge1HotFeatures(PageNumberSimpleSequenciality()))])),
("boolean", Pipeline([('boolean', EdgeBooleanFeatures_v2())])),
(
"numerical",
Pipeline([
('selector', EdgeNumericalSelector()),
#v1 ('numerical', StandardScaler(copy=False, with_mean=True, with_std=True)) #use in-place scaling
('numerical',
QuantileTransformer(n_quantiles=self.n_QUANTILES,
copy=False)) #use in-place scaling
])),
(
"sourcetext0",
Pipeline([
('selector',
EdgeTransformerSourceText(0,
bMirrorPage=bMirrorPage,
bMultiPage=bMultiPage)),
(
'tfidf',
TfidfVectorizer(
lowercase=self.b_tfidf_edge_lc,
max_features=self.n_tfidf_edge,
analyzer='char',
ngram_range=self.t_ngrams_edge #(2,6)
,
dtype=np.float64)),
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
])),
(
"targettext0",
Pipeline([
('selector',
EdgeTransformerTargetText(0,
bMirrorPage=bMirrorPage,
bMultiPage=bMultiPage)),
(
'tfidf',
TfidfVectorizer(
lowercase=self.b_tfidf_edge_lc,
max_features=self.n_tfidf_edge,
analyzer='char',
ngram_range=self.t_ngrams_edge
#, analyzer = 'word', ngram_range=self.tEDGE_NGRAMS
,
dtype=np.float64)),
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
])),
(
"sourcetext1",
Pipeline([
('selector',
EdgeTransformerSourceText(1,
bMirrorPage=bMirrorPage,
bMultiPage=bMultiPage)),
(
'tfidf',
TfidfVectorizer(
lowercase=self.b_tfidf_edge_lc,
max_features=self.n_tfidf_edge,
analyzer='char',
ngram_range=self.t_ngrams_edge #(2,6)
,
dtype=np.float64)),
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
])),
(
"targettext1",
Pipeline([
('selector',
EdgeTransformerTargetText(1,
bMirrorPage=bMirrorPage,
bMultiPage=bMultiPage)),
(
'tfidf',
TfidfVectorizer(
lowercase=self.b_tfidf_edge_lc,
max_features=self.n_tfidf_edge,
analyzer='char',
ngram_range=self.t_ngrams_edge
#, analyzer = 'word', ngram_range=self.tEDGE_NGRAMS
,
dtype=np.float64)),
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
]))
]
if bMultiPage:
lEdgeFeature.extend([
(
"sourcetext2",
Pipeline([
('selector',
EdgeTransformerSourceText(2,
bMirrorPage=bMirrorPage,
bMultiPage=bMultiPage)),
(
'tfidf',
TfidfVectorizer(
lowercase=self.b_tfidf_edge_lc,
max_features=self.n_tfidf_edge,
analyzer='char',
ngram_range=self.t_ngrams_edge #(2,6)
,
dtype=np.float64)),
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
])),
(
"targettext2",
Pipeline([
('selector',
EdgeTransformerTargetText(2,
bMirrorPage=bMirrorPage,
bMultiPage=bMultiPage)),
(
'tfidf',
TfidfVectorizer(
lowercase=self.b_tfidf_edge_lc,
max_features=self.n_tfidf_edge,
analyzer='char',
ngram_range=self.t_ngrams_edge
#, analyzer = 'word', ngram_range=self.tEDGE_NGRAMS
,
dtype=np.float64)),
('todense', SparseToDense()
) #pystruct needs an array, not a sparse matrix
]))
])
edge_transformer = FeatureUnion(lEdgeFeature)
#return _node_transformer, _edge_transformer, tdifNodeTextVectorizer
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
self.tfidfNodeTextVectorizer = tdifNodeTextVectorizer
def __init__(self,
nbClass=None,
n_feat_node=None,
t_ngrams_node=None,
b_node_lc=None,
n_feat_edge=None,
t_ngrams_edge=None,
b_edge_lc=None,
n_jobs=1):
FeatureDefinition.__init__(self)
assert nbClass, "Error: indicate the number of classes"
self.nbClass = nbClass
self.n_feat_node, self.t_ngrams_node, self.b_node_lc = n_feat_node, t_ngrams_node, b_node_lc
self.n_feat_edge, self.t_ngrams_edge, self.b_edge_lc = n_feat_edge, t_ngrams_edge, b_edge_lc
# tdifNodeTextVectorizer = TfidfVectorizer(lowercase=self.b_node_lc, max_features=self.n_feat_node
# , analyzer = 'char', ngram_range=self.t_ngrams_node #(2,6)
# , dtype=np.float64)
"""
- loading pre-computed data from: CV_5/model_A_fold_1_transf.pkl
no such file : CV_5/model_A_fold_1_transf.pkl
Traceback (most recent call last):
File "/opt/project/read/jl_git/TranskribusDU/src/tasks/DU_GTBooks_5labels.py", line 216, in <module>
oReport = doer._nfold_RunFoldFromDisk(options.iFoldRunNum, options.warm)
File "/opt/project/read/jl_git/TranskribusDU/src/tasks/DU_CRF_Task.py", line 481, in _nfold_RunFoldFromDisk
oReport = self._nfold_RunFold(iFold, ts_trn, lFilename_trn, train_index, test_index, bWarm=bWarm)
File "/opt/project/read/jl_git/TranskribusDU/src/tasks/DU_CRF_Task.py", line 565, in _nfold_RunFold
fe.fitTranformers(lGraph_trn)
File "/opt/project/read/jl_git/TranskribusDU/src/crf/FeatureDefinition_PageXml_logit_v2.py", line 141, in fitTranformers
self._node_transformer.fit(lAllNode)
File "/opt/project/read/VIRTUALENV_PYTHON_FULL_type/lib/python2.7/site-packages/sklearn/pipeline.py", line 712, in fit
for _, trans, _ in self._iter())
File "/opt/project/read/VIRTUALENV_PYTHON_FULL_type/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.py", line 768, in __call__
self.retrieve()
File "/opt/project/read/VIRTUALENV_PYTHON_FULL_type/lib/python2.7/site-packages/sklearn/externals/joblib/parallel.py", line 719, in retrieve
raise exception
RuntimeError: maximum recursion depth exceeded
"""
"""
I guess this is due to the cyclic links to node's neighbours.
But joblib.Parallel uses cPickle, so we cannot specialize the serialization of the Block objects.
JLM April 2017
"""
n_jobs_from_graph = 1 #we cannot pickl the list of graph, so n_jobs = 1 for this part!
# n_jobs_NodeTransformerLogit = max(1, n_jobs/2) #half of the jobs for the NodeTransformerLogit, the rets for the others
n_jobs_NodeTransformerLogit = max(1, n_jobs - 1)
#we keep a ref onto it because its fitting needs not only all the nodes, but also additional info, available on the graph objects
self._node_transf_logit = NodeTransformerLogit(
nbClass,
self.n_feat_node,
self.t_ngrams_node,
self.b_node_lc,
n_jobs=n_jobs_NodeTransformerLogit)
node_transformer = FeatureUnion(
[ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("text", self._node_transf_logit),
(
"textlen",
Pipeline([
('selector', NodeTransformerTextLen()),
('textlen',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"xywh",
Pipeline([
('selector', NodeTransformerXYWH_v2()),
('xywh',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"neighbors",
Pipeline([
('selector', NodeTransformerNeighbors()),
('neighbors',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"1hot",
Pipeline([('1hot', Node1HotFeatures()
) #does the 1-hot encoding directly
]))
# , ('ocr' , Pipeline([
# ('ocr', NodeOCRFeatures())
# ])
# )
# , ('pnumre' , Pipeline([
# ('pnumre', NodePNumFeatures())
# ])
# )
# , ("doc", Pipeline([
# ('zero', Zero2Features())
# #THIS ONE MUST BE LAST, because it include a placeholder column for the doculent-level tfidf
# ])
# )
],
n_jobs=n_jobs_from_graph)
lEdgeFeature = [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot",
Pipeline([('1hot',
Edge1HotFeatures(PageNumberSimpleSequenciality()))])),
("boolean", Pipeline([('boolean', EdgeBooleanFeatures_v2())])),
(
"numerical",
Pipeline([
('selector', EdgeNumericalSelector()),
('numerical',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
])),
("nodetext", EdgeTransformerLogit(nbClass,
self._node_transf_logit))
]
edge_transformer = FeatureUnion(lEdgeFeature, n_jobs=n_jobs_from_graph)
#return _node_transformer, _edge_transformer, tdifNodeTextVectorizer
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
def __init__(self,
nbClass=None,
n_feat_node=None,
t_ngrams_node=None,
b_node_lc=None,
n_feat_edge=None,
t_ngrams_edge=None,
b_edge_lc=None,
n_jobs=1):
FeatureDefinition.__init__(self, nbClass)
assert nbClass, "Error: indicate the numbe of classes"
self.nbClass = nbClass
self.n_feat_node, self.t_ngrams_node, self.b_node_lc = n_feat_node, t_ngrams_node, b_node_lc
self.n_feat_edge, self.t_ngrams_edge, self.b_edge_lc = n_feat_edge, t_ngrams_edge, b_edge_lc
# tdifNodeTextVectorizer = TfidfVectorizer(lowercase=self.b_node_lc, max_features=self.n_feat_node
# , analyzer = 'char', ngram_range=self.t_ngrams_node #(2,6)
# , dtype=np.float64)
"""
I tried to parallelize this code but I'm getting an error on Windows:
File "c:\Local\meunier\git\TranskribusDU\src\crf\FeatureDefinition_PageXml_logit.py", line 144, in fitTranformers
self._node_transformer.fit(lAllNode)
File "C:\Anaconda2\lib\site-packages\sklearn\pipeline.py", line 709, in fit
for _, trans, _ in self._iter())
File "C:\Anaconda2\lib\site-packages\sklearn\externals\joblib\parallel.py", line 768, in __call__
self.retrieve()
File "C:\Anaconda2\lib\site-packages\sklearn\externals\joblib\parallel.py", line 719, in retrieve
raise exception
TypeError: can't pickle PyCapsule objects
(virtual_python_pystruct) (C:\Anaconda2) c:\tmp_READ\tuto>python -c "import sklearn; print sklearn.__version__"
0.18.1
=> I force n_jobs to 1
"""
n_jobs = 1
n_jobs_NodeTransformerLogit = max(
1, n_jobs / 2
) #half of the jobs for the NodeTransformerLogit, the rets for the others
#we keep a ref onto it because its fitting needs not only all the nodes, but also additional info, available on the graph objects
self._node_transf_logit = NodeTransformerLogit(
nbClass,
self.n_feat_node,
self.t_ngrams_node,
self.b_node_lc,
n_jobs=n_jobs_NodeTransformerLogit)
node_transformer = FeatureUnion(
[ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("text", self._node_transf_logit),
(
"textlen",
Pipeline([
('selector', NodeTransformerTextLen()),
('textlen',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"xywh",
Pipeline([
('selector', NodeTransformerXYWH()),
('xywh',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"neighbors",
Pipeline([
('selector', NodeTransformerNeighbors()),
('neighbors',
StandardScaler(copy=False,
with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"1hot",
Pipeline([('1hot', Node1HotFeatures()
) #does the 1-hot encoding directly
]))
# , ('ocr' , Pipeline([
# ('ocr', NodeOCRFeatures())
# ])
# )
# , ('pnumre' , Pipeline([
# ('pnumre', NodePNumFeatures())
# ])
# )
# , ("doc", Pipeline([
# ('zero', Zero2Features())
# #THIS ONE MUST BE LAST, because it include a placeholder column for the doculent-level tfidf
# ])
# )
],
n_jobs=max(1, n_jobs - n_jobs_NodeTransformerLogit))
lEdgeFeature = [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot",
Pipeline([('1hot',
Edge1HotFeatures(PageNumberSimpleSequenciality()))])),
("boolean", Pipeline([('boolean', EdgeBooleanFeatures())])),
(
"numerical",
Pipeline([
('selector', EdgeNumericalSelector()),
('numerical',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
])),
("nodetext", EdgeTransformerLogit(nbClass,
self._node_transf_logit))
]
edge_transformer = FeatureUnion(lEdgeFeature, n_jobs=n_jobs)
#return _node_transformer, _edge_transformer, tdifNodeTextVectorizer
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
def __init__(self,
n_tfidf_node=None,
t_ngrams_node=None,
b_tfidf_node_lc=None,
n_tfidf_edge=None,
t_ngrams_edge=None,
b_tfidf_edge_lc=None,
feat_select=None,
text_neighbors=False,
n_tfidf_node_neighbors=500,
XYWH_v2=False,
edge_features=False):
FeatureDefinition.__init__(self)
self.n_tfidf_node, self.t_ngrams_node, self.b_tfidf_node_lc = n_tfidf_node, t_ngrams_node, b_tfidf_node_lc
self.n_tfidf_edge, self.t_ngrams_edge, self.b_tfidf_edge_lc = n_tfidf_edge, t_ngrams_edge, b_tfidf_edge_lc
self.text_neighbors = text_neighbors
self.n_tfidf_node_neighbors = n_tfidf_node_neighbors
self.XYWH_v2 = XYWH_v2
self.edge_features = edge_features
#TODO n_jobs=4
if feat_select == 'chi2':
feat_selector = SelectKBest(chi2, k=self.n_tfidf_node)
elif feat_select == 'mi_rr':
print('Using Mutual Information Round Robin as Feature Selection')
feat_selector = SelectRobinBest(mutual_information,
k=self.n_tfidf_node)
feat_selector_neigh = SelectRobinBest(mutual_information,
k=self.n_tfidf_node)
elif feat_select == 'chi2_rr':
#chi_score = lambda x,y : chi2(x,y)[0] #this can not be pickled ...
feat_selector = SelectRobinBest(chi2_scores, k=self.n_tfidf_node)
feat_selector_neigh = SelectRobinBest(chi2_scores,
k=self.n_tfidf_node)
elif feat_select == 'tf' or feat_select is None:
feat_selector = None
else:
raise ValueError('Invalid Feature Selection method', feat_select)
if feat_selector:
tdifNodeTextVectorizer = TfidfVectorizer(
lowercase=self.b_tfidf_node_lc,
max_features=10000,
analyzer='char',
ngram_range=self.t_ngrams_node) #(2,6)
text_pipeline = Pipeline([
('selector', NodeTransformerTextEnclosed()),
('tf', tdifNodeTextVectorizer),
('word_selector', feat_selector),
#('todense', SparseToDense()) #Here we don't need to convert to Dense anymore
])
else:
tdifNodeTextVectorizer = TfidfVectorizer(
lowercase=self.b_tfidf_node_lc,
max_features=self.n_tfidf_node,
analyzer='char',
ngram_range=self.t_ngrams_node #(2,6)
,
dtype=np.float64)
text_pipeline = Pipeline([
('selector', NodeTransformerTextEnclosed()),
('tf', tdifNodeTextVectorizer),
#('todense', SparseToDense()) #Here we don't need to convert to Dense anymore
])
node_transformer_ops = [
("text", text_pipeline),
(
"textlen",
Pipeline([
('selector', NodeTransformerTextLen()),
('textlen',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"neighbors",
Pipeline([
('selector', NodeTransformerNeighbors()),
('neighbors',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
])),
(
"1hot",
Pipeline([('1hot', Node1HotFeatures()
) #does the 1-hot encoding directly
]))
]
if self.XYWH_v2 is True:
feat_xy = (
"xywh",
Pipeline([
('selector', NodeTransformerXYWH_v2()),
('xywh',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
]))
else:
feat_xy = (
"xywh",
Pipeline([
('selector', NodeTransformerXYWH()),
('xywh',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
]))
node_transformer_ops.append(feat_xy)
if text_neighbors:
#BY DEFAULT we use chi2
if self.n_tfidf_node_neighbors > 0:
feat_selector_neigh = SelectKBest(
chi2, k=self.n_tfidf_node_neighbors)
neighborsTextVectorizer = TfidfVectorizer(
lowercase=self.b_tfidf_node_lc,
analyzer='char',
ngram_range=self.t_ngrams_node) #(2,6)
neighbors_text_pipeline = Pipeline([
('selector', NodeTransformerNeighborsAllText()),
('tf_neighbors', neighborsTextVectorizer),
('feat_selector', feat_selector_neigh),
])
else:
neighborsTextVectorizer = TfidfVectorizer(
lowercase=self.b_tfidf_node_lc,
analyzer='char',
ngram_range=self.t_ngrams_node) #(2,6)
neighbors_text_pipeline = Pipeline([
('selector', NodeTransformerNeighborsAllText()),
('tf_neighbors', neighborsTextVectorizer)
])
node_transformer_ops.append(
('text_neighbors', neighbors_text_pipeline))
print(node_transformer_ops)
node_aggregated_edge_features = [('1hot_edge',
NodeEdgeTransformer(Edge1HotFeatures(
PageNumberSimpleSequenciality()),
agg_func='sum'))]
node_aggregated_edge_features.append(
('boolean_edge',
NodeEdgeTransformer(EdgeBooleanFeatures_v2(), agg_func='sum')))
#Aggregated Numerical Features do not make a lot of sense here ....
if edge_features:
node_transformer_ops.extend(node_aggregated_edge_features)
print(node_transformer_ops)
node_transformer = FeatureUnion(node_transformer_ops)
#Minimal EdgeFeature Here
lEdgeFeature = [ #CAREFUL IF YOU CHANGE THIS - see cleanTransformers method!!!!
("1hot",
Pipeline([('1hot',
Edge1HotFeatures(PageNumberSimpleSequenciality()))])),
("boolean", Pipeline([('boolean', EdgeBooleanFeatures())])),
(
"numerical",
Pipeline([
('selector', EdgeNumericalSelector()),
('numerical',
StandardScaler(copy=False, with_mean=True,
with_std=True)) #use in-place scaling
]))
]
edge_transformer = FeatureUnion(lEdgeFeature)
self._node_transformer = node_transformer
self._edge_transformer = edge_transformer
self.tfidfNodeTextVectorizer = tdifNodeTextVectorizer