explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-07-05 20:33:10 +03:00
Code Issues Packages Projects Releases Wiki Activity

Improve indexing on reference implementation

Browse Source
This commit is contained in:
Matthew Honnibal 2021-11-01 00:23:15 +01:00
parent 0a7d9278cd
commit 4e894e1076
1 changed files with 5 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
spacy/ml/tb_framework.py
View File

@ -230,6 +230,7 @@ def _forward_reference(model, docs_moves: Tuple[List[Doc], TransitionSystem], is
docs, moves = docs_moves docs, moves = docs_moves
states = moves.init_batch(docs) states = moves.init_batch(docs)
tokvecs, backprop_tok2vec = tok2vec(docs, is_train) tokvecs, backprop_tok2vec = tok2vec(docs, is_train)
tokvecs = model.ops.xp.vstack((tokvecs, lower_pad))
all_ids = [] all_ids = []
all_which = [] all_which = []
all_statevecs = [] all_statevecs = []
@ -244,13 +245,7 @@ def _forward_reference(model, docs_moves: Tuple[List[Doc], TransitionSystem], is
state.set_context_tokens(ids, i, nF) state.set_context_tokens(ids, i, nF)
# Sum the state features, add the bias and apply the activation (maxout) # Sum the state features, add the bias and apply the activation (maxout)
# to create the state vectors. # to create the state vectors.
tokfeats3f = model.ops.alloc3f(ids.shape[0], nF, nI) tokfeats3f = tokvecs[ids]
for i in range(ids.shape[0]):
for j in range(nF):
if ids[i, j] == -1:
tokfeats3f[i, j] = lower_pad
else:
tokfeats3f[i, j] = tokvecs[ids[i, j]]
tokfeats = model.ops.reshape2f(tokfeats3f, tokfeats3f.shape[0], -1) tokfeats = model.ops.reshape2f(tokfeats3f, tokfeats3f.shape[0], -1)
preacts2f = model.ops.gemm(tokfeats, lower_W, trans2=True) preacts2f = model.ops.gemm(tokfeats, lower_W, trans2=True)
preacts2f += lower_b preacts2f += lower_b
@ -309,16 +304,9 @@ def _forward_reference(model, docs_moves: Tuple[List[Doc], TransitionSystem], is
d_tokfeats = model.ops.gemm(d_preacts2f, lower_W) d_tokfeats = model.ops.gemm(d_preacts2f, lower_W)
# Get the gradients of the tokvecs and the padding # Get the gradients of the tokvecs and the padding
d_tokfeats3f = model.ops.reshape3f(d_tokfeats, nS, nF, nI) d_tokfeats3f = model.ops.reshape3f(d_tokfeats, nS, nF, nI)
d_lower_pad = model.ops.alloc1f(nI) model.ops.scatter_add(d_tokvecs, ids, d_tokfeats3f)
assert ids.shape[0] == nS model.inc_grad("lower_pad", d_tokvecs[-1])
for i in range(ids.shape[0]): return (backprop_tok2vec(d_tokvecs[:-1]), None)
for j in range(ids.shape[1]):
if ids[i, j] == -1:
d_lower_pad += d_tokfeats3f[i, j]
else:
d_tokvecs[ids[i, j]] += d_tokfeats3f[i, j]
model.inc_grad("lower_pad", d_lower_pad)
return (backprop_tok2vec(d_tokvecs), None)
return (states, all_scores), backprop_parser return (states, all_scores), backprop_parser