explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2024-11-13 13:17:06 +03:00
Code Issues Packages Projects Releases Wiki Activity

Add padding vector in parser, to make gradient more correct

Browse Source
This commit is contained in:
Matthew Honnibal 2017-11-04 00:23:23 +01:00
parent c2bbf076a4
commit 98c29b7912
1 changed files with 8 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
spacy/syntax/nn_parser.pyx
View File

@ -532,7 +532,9 @@ cdef class Parser:
return None return None
backprops = [] backprops = []
d_tokvecs = state2vec.ops.allocate(tokvecs.shape) # Add a padding vector to the d_tokvecs gradient, so that missing
# values don't affect the real gradient.
d_tokvecs = state2vec.ops.allocate((tokvecs.shape[0]+1, tokvecs.shape[1]))
cdef float loss = 0. cdef float loss = 0.
n_steps = 0 n_steps = 0
while todo: while todo:
@ -615,7 +617,9 @@ cdef class Parser:
bp_vectors)) bp_vectors))
else: else:
backprop_lower.append((ids, d_vector, bp_vectors)) backprop_lower.append((ids, d_vector, bp_vectors))
d_tokvecs = self.model[0].ops.allocate(tokvecs.shape) # Add a padding vector to the d_tokvecs gradient, so that missing
# values don't affect the real gradient.
d_tokvecs = state2vec.ops.allocate((tokvecs.shape[0]+1, tokvecs.shape[1]))
self._make_updates(d_tokvecs, bp_tokvecs, backprop_lower, sgd, self._make_updates(d_tokvecs, bp_tokvecs, backprop_lower, sgd,
cuda_stream) cuda_stream)
@ -668,7 +672,8 @@ cdef class Parser:
(ids.size, d_state_features.shape[2])) (ids.size, d_state_features.shape[2]))
self.model[0].ops.scatter_add(d_tokvecs, ids, self.model[0].ops.scatter_add(d_tokvecs, ids,
d_state_features) d_state_features)
bp_tokvecs(d_tokvecs, sgd=sgd) # Padded -- see update()
bp_tokvecs(d_tokvecs[:-1], sgd=sgd)
@property @property
def move_names(self): def move_names(self):