explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-02-13 18:10:35 +03:00
Code Issues Packages Projects Releases Wiki Activity

Move history size under feature flags

Browse Source
This commit is contained in:
Matthew Honnibal 2017-10-05 19:38:13 -05:00
parent dcdfa071aa
commit e25ffcb11f
1 changed files with 16 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
spacy/syntax/nn_parser.pyx
View File

@ -68,14 +68,10 @@ from ..gold cimport GoldParse
from ..attrs cimport ID, TAG, DEP, ORTH, NORM, PREFIX, SUFFIX, TAG from ..attrs cimport ID, TAG, DEP, ORTH, NORM, PREFIX, SUFFIX, TAG
from . import _beam_utils from . import _beam_utils
USE_HISTORY = True
HIST_SIZE = 8 # Max 8
HIST_DIMS = 8
def get_templates(*args, **kwargs): def get_templates(*args, **kwargs):
return [] return []
USE_FTRL = True
DEBUG = False DEBUG = False
def set_debug(val): def set_debug(val):
global DEBUG global DEBUG
@ -248,6 +244,8 @@ cdef class Parser:
hidden_width = util.env_opt('hidden_width', hidden_width) hidden_width = util.env_opt('hidden_width', hidden_width)
parser_maxout_pieces = util.env_opt('parser_maxout_pieces', 2) parser_maxout_pieces = util.env_opt('parser_maxout_pieces', 2)
embed_size = util.env_opt('embed_size', 7000) embed_size = util.env_opt('embed_size', 7000)
hist_size = util.env_opt('history_feats', cfg.get('history_feats', 0))
hist_width = util.env_opt('history_width', cfg.get('history_width', 0))
tok2vec = Tok2Vec(token_vector_width, embed_size, tok2vec = Tok2Vec(token_vector_width, embed_size,
pretrained_dims=cfg.get('pretrained_dims', 0)) pretrained_dims=cfg.get('pretrained_dims', 0))
tok2vec = chain(tok2vec, flatten) tok2vec = chain(tok2vec, flatten)
@ -263,20 +261,21 @@ cdef class Parser:
with Model.use_device('cpu'): with Model.use_device('cpu'):
if depth == 0: if depth == 0:
if USE_HISTORY: if hist_size:
upper = chain( upper = chain(
HistoryFeatures(nr_class=nr_class, hist_size=HIST_SIZE, HistoryFeatures(nr_class=nr_class, hist_size=hist_size,
nr_dim=HIST_DIMS), nr_dim=hist_width),
zero_init(Affine(nr_class, nr_class+HIST_SIZE*HIST_DIMS, zero_init(Affine(nr_class, nr_class+hist_size*hist_size,
drop_factor=0.0))) drop_factor=0.0)))
upper.is_noop = False upper.is_noop = False
else: else:
upper = chain() upper = chain()
upper.is_noop = True upper.is_noop = True
elif USE_HISTORY: elif hist_size:
upper = chain( upper = chain(
HistoryFeatures(nr_class=nr_class, hist_size=HIST_SIZE, nr_dim=HIST_DIMS), HistoryFeatures(nr_class=nr_class, hist_size=hist_size,
Maxout(hidden_width, hidden_width+HIST_SIZE*HIST_DIMS), nr_dim=hist_width),
Maxout(hidden_width, hidden_width+hist_size*hist_width),
clone(Maxout(hidden_width, hidden_width), depth-2), clone(Maxout(hidden_width, hidden_width), depth-2),
zero_init(Affine(nr_class, hidden_width, drop_factor=0.0)) zero_init(Affine(nr_class, hidden_width, drop_factor=0.0))
) )
@ -296,7 +295,9 @@ cdef class Parser:
'depth': depth, 'depth': depth,
'token_vector_width': token_vector_width, 'token_vector_width': token_vector_width,
'hidden_width': hidden_width, 'hidden_width': hidden_width,
'maxout_pieces': parser_maxout_pieces 'maxout_pieces': parser_maxout_pieces,
'hist_size': hist_size,
'hist_width': hist_width
} }
return (tok2vec, lower, upper), cfg return (tok2vec, lower, upper), cfg
@ -369,7 +370,7 @@ cdef class Parser:
_cleanup(beam) _cleanup(beam)
return output return output
def pipe(self, docs, int batch_size=1000, int n_threads=2, def pipe(self, docs, int batch_size=256, int n_threads=2,
beam_width=None, beam_density=None): beam_width=None, beam_density=None):
""" """
Process a stream of documents. Process a stream of documents.
@ -454,7 +455,7 @@ cdef class Parser:
hists.append([st.get_hist(j+1) for j in range(8)]) hists.append([st.get_hist(j+1) for j in range(8)])
hists = numpy.asarray(hists) hists = numpy.asarray(hists)
vectors = state2vec(token_ids[:next_step.size()]) vectors = state2vec(token_ids[:next_step.size()])
if USE_HISTORY: if self.cfg.get('hist_size'):
scores = vec2scores((vectors, hists)) scores = vec2scores((vectors, hists))
else: else:
scores = vec2scores(vectors) scores = vec2scores(vectors)
@ -577,7 +578,7 @@ cdef class Parser:
mask = vec2scores.ops.get_dropout_mask(vector.shape, drop) mask = vec2scores.ops.get_dropout_mask(vector.shape, drop)
vector *= mask vector *= mask
hists = numpy.asarray([st.history for st in states], dtype='i') hists = numpy.asarray([st.history for st in states], dtype='i')
if USE_HISTORY: if self.cfg.get('hist_size', 0):
scores, bp_scores = vec2scores.begin_update((vector, hists), drop=drop) scores, bp_scores = vec2scores.begin_update((vector, hists), drop=drop)
else: else:
scores, bp_scores = vec2scores.begin_update(vector, drop=drop) scores, bp_scores = vec2scores.begin_update(vector, drop=drop)