explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-07-19 20:52:23 +03:00
Code Issues Packages Projects Releases Wiki Activity

Cleanup

Browse Source
This commit is contained in:
Paul O'Leary McCann 2021-07-10 22:42:55 +09:00
parent e00bd422d9
commit c25ec292a9
1 changed files with 15 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
spacy/ml/models/coref.py
View File

@ -89,15 +89,17 @@ def build_width_scorer(max_span_width, hidden_size, feature_embed_size=20):
>> Linear(nI=hidden_size, nO=1) >> Linear(nI=hidden_size, nO=1)
) )
span_width_prior.initialize() span_width_prior.initialize()
return Model( model = Model(
"WidthScorer", "WidthScorer",
forward=width_score_forward, forward=width_score_forward,
layers=[span_width_prior]) layers=[span_width_prior])
model.set_ref("width_prior", span_width_prior)
return model
def width_score_forward(model, embeds: SpanEmbeddings, is_train) -> Tuple[Floats1d, Callable]: def width_score_forward(model, embeds: SpanEmbeddings, is_train) -> Tuple[Floats1d, Callable]:
# calculate widths, subtracting 1 so it's 0-index # calculate widths, subtracting 1 so it's 0-index
w_ffnn = model.layers[0] w_ffnn = model.get_ref("width_prior")
idxs = embeds.indices idxs = embeds.indices
widths = idxs[:,1] - idxs[:,0] - 1 widths = idxs[:,1] - idxs[:,0] - 1
wscores, width_b = w_ffnn(widths, is_train) wscores, width_b = w_ffnn(widths, is_train)
@ -227,6 +229,7 @@ def coarse_prune(
cscores = scores[offset:hi] cscores = scores[offset:hi]
# negate it so highest numbers come first # negate it so highest numbers come first
# This is relatively slow but can't be skipped.
tops = (model.ops.xp.argsort(-1 * cscores)).tolist() tops = (model.ops.xp.argsort(-1 * cscores)).tolist()
starts = spanembeds.indices[offset:hi, 0].tolist() starts = spanembeds.indices[offset:hi, 0].tolist()
ends = spanembeds.indices[offset:hi:, 1].tolist() ends = spanembeds.indices[offset:hi:, 1].tolist()
@ -298,7 +301,7 @@ def take_vecs_forward(model, inputs: SpanEmbeddings, is_train) -> Floats2d:
def build_ant_scorer( def build_ant_scorer(
bilinear, dropout, ant_limit=50 bilinear, dropout, ant_limit=50
) -> Model[Tuple[Floats1d, SpanEmbeddings], List[Floats2d]]: ) -> Model[Tuple[Floats1d, SpanEmbeddings], List[Floats2d]]:
return Model( model = Model(
"AntScorer", "AntScorer",
forward=ant_scorer_forward, forward=ant_scorer_forward,
layers=[bilinear, dropout], layers=[bilinear, dropout],
@ -306,6 +309,9 @@ def build_ant_scorer(
"ant_limit": ant_limit, "ant_limit": ant_limit,
}, },
) )
model.set_ref("bilinear", bilinear)
model.set_ref("dropout", dropout)
return model
def ant_scorer_forward( def ant_scorer_forward(
@ -318,13 +324,8 @@ def ant_scorer_forward(
# this contains the coarse bilinear in coref-hoi # this contains the coarse bilinear in coref-hoi
# coarse bilinear is a single layer linear network # coarse bilinear is a single layer linear network
# TODO make these proper refs # TODO make these proper refs
bilinear = model.layers[0] bilinear = model.get_ref("bilinear")
dropout = model.layers[1] dropout = model.get_ref("dropout")
# XXX Note on dimensions: This won't work as a ragged because the floats2ds
# are not all the same dimensions. Each floats2d is a square in the size of
# the number of antecedents in the document. Actually, that will have the
# same size if antecedents are padded... Needs checking.
mscores, sembeds = inputs mscores, sembeds = inputs
vecs = sembeds.vectors # ragged vecs = sembeds.vectors # ragged
@ -362,7 +363,6 @@ def ant_scorer_forward(
# now add the placeholder # now add the placeholder
placeholder = ops.alloc2f(scores.shape[0], 1) placeholder = ops.alloc2f(scores.shape[0], 1)
top_scores = xp.concatenate( (placeholder, top_scores), 1) top_scores = xp.concatenate( (placeholder, top_scores), 1)
#top_scores = ops.softmax(top_scores, axis=1)
out.append((top_scores, top_scores_idx)) out.append((top_scores, top_scores_idx))
@ -389,6 +389,7 @@ def ant_scorer_forward(
offset = 0 offset = 0
for dy, (prod_back, pw_sum_back), ll in zip(dYscores, backprops, veclens): for dy, (prod_back, pw_sum_back), ll in zip(dYscores, backprops, veclens):
hi = offset + ll
dyscore, dyidx = dy dyscore, dyidx = dy
# remove the placeholder # remove the placeholder
dyscore = dyscore[:, 1:] dyscore = dyscore[:, 1:]
@ -398,10 +399,10 @@ def ant_scorer_forward(
for ii, (ridx, rscores) in enumerate(zip(dyidx, dyscore)): for ii, (ridx, rscores) in enumerate(zip(dyidx, dyscore)):
fullscore[ii][ridx] = rscores fullscore[ii][ridx] = rscores
dXembeds.data[offset : offset + ll] = prod_back(fullscore) dXembeds.data[offset : hi] = prod_back(fullscore)
dXscores[offset : offset + ll] = pw_sum_back(fullscore) dXscores[offset : hi] = pw_sum_back(fullscore)
offset += ll offset = hi
# make it fit back into the linear # make it fit back into the linear
dXscores = xp.expand_dims(dXscores, 1) dXscores = xp.expand_dims(dXscores, 1)
return (dXscores, SpanEmbeddings(idxes, dXembeds)) return (dXscores, SpanEmbeddings(idxes, dXembeds))