explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2024-12-25 17:36:30 +03:00
Code Issues Packages Projects Releases Wiki Activity

Don't normalize gradients

Browse Source
This commit is contained in:
Matthw Honnibal 2020-07-07 17:21:58 +02:00
parent 8177f25b6c
commit a4164f67ca
2 changed files with 6 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
spacy/pipeline/pipes.pyx
View File

@ -334,7 +334,7 @@ class Tagger(Pipe):
losses[self.name] += (gradient**2).sum()
def get_loss(self, examples, scores):
loss_func = SequenceCategoricalCrossentropy(names=self.labels)
loss_func = SequenceCategoricalCrossentropy(names=self.labels, normalize=False)
truths = [eg.get_aligned("tag", as_string=True) for eg in examples]
d_scores, loss = loss_func(scores, truths)
if self.model.ops.xp.isnan(loss):

17
spacy/syntax/nn_parser.pyx
View File

@ -65,7 +65,6 @@ cdef class Parser:
self.set_output(self.moves.n_moves)
self.cfg = dict(cfg)
self.cfg.setdefault("update_with_oracle_cut_size", 100)
self.cfg.setdefault("normalize_gradients_with_batch_size", True)
self._multitasks = []
for multitask in cfg.get("multitasks", []):
self.add_multitask_objective(multitask)
@ -300,17 +299,10 @@ cdef class Parser:
states, golds = zip(*states_golds)
scores, backprop = model.begin_update(states)
d_scores = self.get_batch_loss(states, golds, scores, losses)
if self.cfg["normalize_gradients_with_batch_size"]:
# We have to be very careful how we do this, because of the way we
# cut up the batch. We subdivide long sequences. If we normalize
# naively, we end up normalizing by sequence length, which
# is bad: that would mean that states in long sequences
# consistently get smaller gradients. Imagine if we have two
# sequences, one length 1000, one length 20. If we cut up
# the 1k sequence so that we have a "batch" of 50 subsequences,
# we don't want the gradients to get 50 times smaller!
d_scores /= n_examples
# Note that the gradient isn't normalized by the batch size
# here, because our "samples" are really the states...But we
# can't normalize by the number of states either, as then we'd
# be getting smaller gradients for states in long sequences.
backprop(d_scores)
# Follow the predicted action
self.transition_states(states, scores)
@ -408,6 +400,7 @@ cdef class Parser:
cpu_log_loss(c_d_scores,
costs, is_valid, &scores[i, 0], d_scores.shape[1])
c_d_scores += d_scores.shape[1]
# Note that we don't normalize this. See comment in update() for why.
if losses is not None:
losses.setdefault(self.name, 0.)
losses[self.name] += (d_scores**2).sum()