explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2024-11-11 04:08:09 +03:00
Code Issues Packages Projects Releases Wiki Activity

Fix error in beam gradient calculation

Browse Source
This commit is contained in:
Matthew Honnibal 2018-05-09 02:44:09 +02:00
parent a6ae1ee6f7
commit a61fd60681
1 changed files with 15 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
spacy/syntax/_beam_utils.pyx
View File

@ -133,7 +133,7 @@ cdef class ParserBeam(object):
if follow_gold: if follow_gold:
min_cost = 0 min_cost = 0
for j in range(beam.nr_class): for j in range(beam.nr_class):
if beam.costs[i][j] < min_cost: if beam.is_valid[i][j] and beam.costs[i][j] < min_cost:
min_cost = beam.costs[i][j] min_cost = beam.costs[i][j]
for j in range(beam.nr_class): for j in range(beam.nr_class):
if beam.costs[i][j] > min_cost: if beam.costs[i][j] > min_cost:
@ -160,7 +160,8 @@ nr_update = 0
def update_beam(TransitionSystem moves, int nr_feature, int max_steps, def update_beam(TransitionSystem moves, int nr_feature, int max_steps,
states, golds, states, golds,
state2vec, vec2scores, state2vec, vec2scores,
int width, losses=None, drop=0.): int width, losses=None, drop=0.,
early_update=True):
global nr_update global nr_update
cdef MaxViolation violn cdef MaxViolation violn
nr_update += 1 nr_update += 1
@ -201,13 +202,16 @@ def update_beam(TransitionSystem moves, int nr_feature, int max_steps,
for indices in p_indices] for indices in p_indices]
g_scores = [numpy.ascontiguousarray(scores[indices], dtype='f') g_scores = [numpy.ascontiguousarray(scores[indices], dtype='f')
for indices in g_indices] for indices in g_indices]
# Now advance the states in the beams. The gold beam is contrained to # Now advance the states in the beams. The gold beam is constrained to
# to follow only gold analyses. # to follow only gold analyses.
pbeam.advance(p_scores) pbeam.advance(p_scores)
gbeam.advance(g_scores, follow_gold=True) gbeam.advance(g_scores, follow_gold=True)
# Track the "maximum violation", to use in the update. # Track the "maximum violation", to use in the update.
for i, violn in enumerate(violns): for i, violn in enumerate(violns):
violn.check_crf(pbeam[i], gbeam[i]) violn.check_crf(pbeam[i], gbeam[i])
if pbeam[i].loss > 0 and pbeam[i].min_score > (gbeam[i].score * 1.5):
pbeam.dones[i] = True
gbeam.dones[i] = True
histories = [] histories = []
losses = [] losses = []
for violn in violns: for violn in violns:
@ -271,14 +275,15 @@ def get_gradient(nr_class, beam_maps, histories, losses):
Each batch has multiple beams Each batch has multiple beams
So history is list of lists of lists of ints So history is list of lists of lists of ints
""" """
nr_step = len(beam_maps)
grads = [] grads = []
nr_step = 0 nr_steps = []
for eg_id, hists in enumerate(histories): for eg_id, hists in enumerate(histories):
nr_step = 0
for loss, hist in zip(losses[eg_id], hists): for loss, hist in zip(losses[eg_id], hists):
if loss != 0.0 and not numpy.isnan(loss): if loss != 0.0 and not numpy.isnan(loss):
nr_step = max(nr_step, len(hist)) nr_step = max(nr_step, len(hist))
for i in range(nr_step): nr_steps.append(nr_step)
for i in range(max(nr_steps)):
grads.append(numpy.zeros((max(beam_maps[i].values())+1, nr_class), grads.append(numpy.zeros((max(beam_maps[i].values())+1, nr_class),
dtype='f')) dtype='f'))
if len(histories) != len(losses): if len(histories) != len(losses):
@ -289,8 +294,11 @@ def get_gradient(nr_class, beam_maps, histories, losses):
continue continue
key = tuple([eg_id]) key = tuple([eg_id])
# Adjust loss for length # Adjust loss for length
# We need to do this because each state in a short path is scored
# multiple times, as we add in the average cost when we run out
# of actions.
avg_loss = loss / len(hist) avg_loss = loss / len(hist)
loss += avg_loss * (nr_step - len(hist)) loss += avg_loss * (nr_steps[eg_id] - len(hist))
for j, clas in enumerate(hist): for j, clas in enumerate(hist):
i = beam_maps[j][key] i = beam_maps[j][key]
# In step j, at state i action clas # In step j, at state i action clas