spaCy/spacy/syntax/_parser_model.pyx

468 lines
17 KiB
Cython
Raw Normal View History

# cython: infer_types=True
# cython: cdivision=True
# cython: boundscheck=False
# coding: utf-8
from __future__ import unicode_literals, print_function
from collections import OrderedDict
import numpy
cimport cython.parallel
import numpy.random
cimport numpy as np
from libc.math cimport exp
from libcpp.vector cimport vector
from libc.string cimport memset, memcpy
from libc.stdlib cimport calloc, free, realloc
from cymem.cymem cimport Pool
from thinc.typedefs cimport weight_t, class_t, hash_t
from thinc.extra.search cimport Beam
from thinc.api import chain, clone
from thinc.v2v import Model, Maxout, Affine
from thinc.misc import LayerNorm
from thinc.neural.ops import CupyOps
from thinc.neural.util import get_array_module
from thinc.linalg cimport Vec, VecVec
💫 Use Blis for matrix multiplications (#2966) Our epic matrix multiplication odyssey is drawing to a close... I've now finally got the Blis linear algebra routines in a self-contained Python package, with wheels for Windows, Linux and OSX. The only missing platform at the moment is Windows Python 2.7. The result is at https://github.com/explosion/cython-blis Thinc v7.0.0 will make the change to Blis. I've put a Thinc v7.0.0.dev0 up on PyPi so that we can test these changes with the CI, and even get them out to spacy-nightly, before Thinc v7.0.0 is released. This PR also updates the other dependencies to be in line with the current versions master is using. I've also resolved the msgpack deprecation problems, and gotten spaCy and Thinc up to date with the latest Cython. The point of switching to Blis is to have control of how our matrix multiplications are executed across platforms. When we were using numpy for this, a different library would be used on pip and conda, OSX would use Accelerate, etc. This would open up different bugs and performance problems, especially when multi-threading was introduced. With the change to Blis, we now strictly single-thread the matrix multiplications. This will make it much easier to use multiprocessing to parallelise the runtime, since we won't have nested parallelism problems to deal with. * Use blis * Use -2 arg to Cython * Update dependencies * Fix requirements * Update setup dependencies * Fix requirement typo * Fix msgpack errors * Remove Python27 test from Appveyor, until Blis works there * Auto-format setup.py * Fix murmurhash version
2018-11-26 23:44:04 +00:00
cimport blis.cy
from .._ml import zero_init, PrecomputableAffine, Tok2Vec, flatten
from .._ml import link_vectors_to_models, create_default_optimizer
from ..compat import copy_array
from ..tokens.doc cimport Doc
from ..gold cimport GoldParse
from ..errors import Errors, TempErrors
from .. import util
from .stateclass cimport StateClass
from .transition_system cimport Transition
from . import _beam_utils
from . import nonproj
cdef WeightsC get_c_weights(model) except *:
cdef WeightsC output
cdef precompute_hiddens state2vec = model.state2vec
output.feat_weights = state2vec.get_feat_weights()
output.feat_bias = <const float*>state2vec.bias.data
cdef np.ndarray vec2scores_W = model.vec2scores.W
cdef np.ndarray vec2scores_b = model.vec2scores.b
cdef np.ndarray class_mask = model._class_mask
output.hidden_weights = <const float*>vec2scores_W.data
output.hidden_bias = <const float*>vec2scores_b.data
output.seen_classes = <const float*>class_mask.data
return output
cdef SizesC get_c_sizes(model, int batch_size) except *:
cdef SizesC output
output.states = batch_size
output.classes = model.vec2scores.nO
output.hiddens = model.state2vec.nO
output.pieces = model.state2vec.nP
output.feats = model.state2vec.nF
output.embed_width = model.tokvecs.shape[1]
return output
cdef ActivationsC alloc_activations(SizesC n) nogil:
cdef ActivationsC A
memset(&A, 0, sizeof(A))
resize_activations(&A, n)
return A
cdef void free_activations(const ActivationsC* A) nogil:
free(A.token_ids)
free(A.scores)
free(A.unmaxed)
free(A.hiddens)
free(A.is_valid)
cdef void resize_activations(ActivationsC* A, SizesC n) nogil:
if n.states <= A._max_size:
A._curr_size = n.states
return
if A._max_size == 0:
A.token_ids = <int*>calloc(n.states * n.feats, sizeof(A.token_ids[0]))
A.scores = <float*>calloc(n.states * n.classes, sizeof(A.scores[0]))
A.unmaxed = <float*>calloc(n.states * n.hiddens * n.pieces, sizeof(A.unmaxed[0]))
A.hiddens = <float*>calloc(n.states * n.hiddens, sizeof(A.hiddens[0]))
A.is_valid = <int*>calloc(n.states * n.classes, sizeof(A.is_valid[0]))
A._max_size = n.states
else:
A.token_ids = <int*>realloc(A.token_ids,
n.states * n.feats * sizeof(A.token_ids[0]))
A.scores = <float*>realloc(A.scores,
n.states * n.classes * sizeof(A.scores[0]))
A.unmaxed = <float*>realloc(A.unmaxed,
n.states * n.hiddens * n.pieces * sizeof(A.unmaxed[0]))
A.hiddens = <float*>realloc(A.hiddens,
n.states * n.hiddens * sizeof(A.hiddens[0]))
A.is_valid = <int*>realloc(A.is_valid,
n.states * n.classes * sizeof(A.is_valid[0]))
A._max_size = n.states
A._curr_size = n.states
cdef void predict_states(ActivationsC* A, StateC** states,
const WeightsC* W, SizesC n) nogil:
resize_activations(A, n)
memset(A.unmaxed, 0, n.states * n.hiddens * n.pieces * sizeof(float))
memset(A.hiddens, 0, n.states * n.hiddens * sizeof(float))
for i in range(n.states):
states[i].set_context_tokens(&A.token_ids[i*n.feats], n.feats)
sum_state_features(A.unmaxed,
W.feat_weights, A.token_ids, n.states, n.feats, n.hiddens * n.pieces)
for i in range(n.states):
VecVec.add_i(&A.unmaxed[i*n.hiddens*n.pieces],
W.feat_bias, 1., n.hiddens * n.pieces)
for j in range(n.hiddens):
index = i * n.hiddens * n.pieces + j * n.pieces
which = Vec.arg_max(&A.unmaxed[index], n.pieces)
A.hiddens[i*n.hiddens + j] = A.unmaxed[index + which]
memset(A.scores, 0, n.states * n.classes * sizeof(float))
💫 Use Blis for matrix multiplications (#2966) Our epic matrix multiplication odyssey is drawing to a close... I've now finally got the Blis linear algebra routines in a self-contained Python package, with wheels for Windows, Linux and OSX. The only missing platform at the moment is Windows Python 2.7. The result is at https://github.com/explosion/cython-blis Thinc v7.0.0 will make the change to Blis. I've put a Thinc v7.0.0.dev0 up on PyPi so that we can test these changes with the CI, and even get them out to spacy-nightly, before Thinc v7.0.0 is released. This PR also updates the other dependencies to be in line with the current versions master is using. I've also resolved the msgpack deprecation problems, and gotten spaCy and Thinc up to date with the latest Cython. The point of switching to Blis is to have control of how our matrix multiplications are executed across platforms. When we were using numpy for this, a different library would be used on pip and conda, OSX would use Accelerate, etc. This would open up different bugs and performance problems, especially when multi-threading was introduced. With the change to Blis, we now strictly single-thread the matrix multiplications. This will make it much easier to use multiprocessing to parallelise the runtime, since we won't have nested parallelism problems to deal with. * Use blis * Use -2 arg to Cython * Update dependencies * Fix requirements * Update setup dependencies * Fix requirement typo * Fix msgpack errors * Remove Python27 test from Appveyor, until Blis works there * Auto-format setup.py * Fix murmurhash version
2018-11-26 23:44:04 +00:00
cdef double one = 1.0
# Compute hidden-to-output
💫 Use Blis for matrix multiplications (#2966) Our epic matrix multiplication odyssey is drawing to a close... I've now finally got the Blis linear algebra routines in a self-contained Python package, with wheels for Windows, Linux and OSX. The only missing platform at the moment is Windows Python 2.7. The result is at https://github.com/explosion/cython-blis Thinc v7.0.0 will make the change to Blis. I've put a Thinc v7.0.0.dev0 up on PyPi so that we can test these changes with the CI, and even get them out to spacy-nightly, before Thinc v7.0.0 is released. This PR also updates the other dependencies to be in line with the current versions master is using. I've also resolved the msgpack deprecation problems, and gotten spaCy and Thinc up to date with the latest Cython. The point of switching to Blis is to have control of how our matrix multiplications are executed across platforms. When we were using numpy for this, a different library would be used on pip and conda, OSX would use Accelerate, etc. This would open up different bugs and performance problems, especially when multi-threading was introduced. With the change to Blis, we now strictly single-thread the matrix multiplications. This will make it much easier to use multiprocessing to parallelise the runtime, since we won't have nested parallelism problems to deal with. * Use blis * Use -2 arg to Cython * Update dependencies * Fix requirements * Update setup dependencies * Fix requirement typo * Fix msgpack errors * Remove Python27 test from Appveyor, until Blis works there * Auto-format setup.py * Fix murmurhash version
2018-11-26 23:44:04 +00:00
blis.cy.gemm(blis.cy.NO_TRANSPOSE, blis.cy.TRANSPOSE,
n.states, n.classes, n.hiddens, one,
<float*>A.hiddens, n.hiddens, 1,
<float*>W.hidden_weights, n.hiddens, 1,
one,
<float*>A.scores, n.classes, 1)
# Add bias
for i in range(n.states):
VecVec.add_i(&A.scores[i*n.classes],
W.hidden_bias, 1., n.classes)
# Set unseen classes to minimum value
i = 0
min_ = A.scores[0]
for i in range(1, n.states * n.classes):
if A.scores[i] < min_:
min_ = A.scores[i]
for i in range(n.states):
for j in range(n.classes):
if not W.seen_classes[j]:
A.scores[i*n.classes+j] = min_
cdef void sum_state_features(float* output,
const float* cached, const int* token_ids, int B, int F, int O) nogil:
cdef int idx, b, f, i
cdef const float* feature
padding = cached
cached += F * O
cdef int id_stride = F*O
cdef float one = 1.
for b in range(B):
for f in range(F):
if token_ids[f] < 0:
feature = &padding[f*O]
else:
idx = token_ids[f] * id_stride + f*O
feature = &cached[idx]
💫 Use Blis for matrix multiplications (#2966) Our epic matrix multiplication odyssey is drawing to a close... I've now finally got the Blis linear algebra routines in a self-contained Python package, with wheels for Windows, Linux and OSX. The only missing platform at the moment is Windows Python 2.7. The result is at https://github.com/explosion/cython-blis Thinc v7.0.0 will make the change to Blis. I've put a Thinc v7.0.0.dev0 up on PyPi so that we can test these changes with the CI, and even get them out to spacy-nightly, before Thinc v7.0.0 is released. This PR also updates the other dependencies to be in line with the current versions master is using. I've also resolved the msgpack deprecation problems, and gotten spaCy and Thinc up to date with the latest Cython. The point of switching to Blis is to have control of how our matrix multiplications are executed across platforms. When we were using numpy for this, a different library would be used on pip and conda, OSX would use Accelerate, etc. This would open up different bugs and performance problems, especially when multi-threading was introduced. With the change to Blis, we now strictly single-thread the matrix multiplications. This will make it much easier to use multiprocessing to parallelise the runtime, since we won't have nested parallelism problems to deal with. * Use blis * Use -2 arg to Cython * Update dependencies * Fix requirements * Update setup dependencies * Fix requirement typo * Fix msgpack errors * Remove Python27 test from Appveyor, until Blis works there * Auto-format setup.py * Fix murmurhash version
2018-11-26 23:44:04 +00:00
blis.cy.axpyv(blis.cy.NO_CONJUGATE, O, one,
<float*>feature, 1,
&output[b*O], 1)
token_ids += F
cdef void cpu_log_loss(float* d_scores,
const float* costs, const int* is_valid, const float* scores,
int O) nogil:
"""Do multi-label log loss"""
cdef double max_, gmax, Z, gZ
best = arg_max_if_gold(scores, costs, is_valid, O)
guess = Vec.arg_max(scores, O)
if best == -1 or guess == -1:
# These shouldn't happen, but if they do, we want to make sure we don't
# cause an OOB access.
return
Z = 1e-10
gZ = 1e-10
max_ = scores[guess]
gmax = scores[best]
for i in range(O):
Z += exp(scores[i] - max_)
if costs[i] <= costs[best]:
gZ += exp(scores[i] - gmax)
for i in range(O):
if costs[i] <= costs[best]:
d_scores[i] = (exp(scores[i]-max_) / Z) - (exp(scores[i]-gmax)/gZ)
else:
d_scores[i] = exp(scores[i]-max_) / Z
cdef int arg_max_if_gold(const weight_t* scores, const weight_t* costs,
const int* is_valid, int n) nogil:
# Find minimum cost
cdef float cost = 1
for i in range(n):
if is_valid[i] and costs[i] < cost:
cost = costs[i]
# Now find best-scoring with that cost
cdef int best = -1
for i in range(n):
if costs[i] <= cost and is_valid[i]:
if best == -1 or scores[i] > scores[best]:
best = i
return best
cdef int arg_max_if_valid(const weight_t* scores, const int* is_valid, int n) nogil:
cdef int best = -1
for i in range(n):
if is_valid[i] >= 1:
if best == -1 or scores[i] > scores[best]:
best = i
return best
class ParserModel(Model):
def __init__(self, tok2vec, lower_model, upper_model, unseen_classes=None):
Model.__init__(self)
self._layers = [tok2vec, lower_model, upper_model]
self.unseen_classes = set()
if unseen_classes:
for class_ in unseen_classes:
self.unseen_classes.add(class_)
def begin_update(self, docs, drop=0.):
step_model = ParserStepModel(docs, self._layers, drop=drop,
unseen_classes=self.unseen_classes)
def finish_parser_update(golds, sgd=None):
step_model.make_updates(sgd)
return None
return step_model, finish_parser_update
def resize_output(self, new_output):
if new_output == self.upper.nO:
return
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 15:25:33 +00:00
smaller = self.upper
2019-02-21 16:12:23 +00:00
with Model.use_device('cpu'):
larger = Affine(new_output, smaller.nI)
larger.W.fill(0.0)
larger.b.fill(0.0)
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 15:25:33 +00:00
# It seems very unhappy if I pass these as smaller.W?
# Seems to segfault. Maybe it's a descriptor protocol thing?
smaller_W = smaller.W
larger_W = larger.W
smaller_b = smaller.b
larger_b = larger.b
# Weights are stored in (nr_out, nr_in) format, so we're basically
# just adding rows here.
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 15:25:33 +00:00
larger_W[:smaller.nO] = smaller_W
larger_b[:smaller.nO] = smaller_b
self._layers[-1] = larger
for i in range(smaller.nO, new_output):
self.unseen_classes.add(i)
2018-09-14 18:50:59 +00:00
def begin_training(self, X, y=None):
self.lower.begin_training(X, y=y)
@property
def tok2vec(self):
return self._layers[0]
@property
def lower(self):
return self._layers[1]
@property
def upper(self):
return self._layers[2]
class ParserStepModel(Model):
def __init__(self, docs, layers, unseen_classes=None, drop=0.):
self.tokvecs, self.bp_tokvecs = layers[0].begin_update(docs, drop=drop)
self.state2vec = precompute_hiddens(len(docs), self.tokvecs, layers[1],
drop=drop)
self.vec2scores = layers[-1]
self.cuda_stream = util.get_cuda_stream()
self.backprops = []
self._class_mask = numpy.zeros((self.vec2scores.nO,), dtype='f')
self._class_mask.fill(1)
if unseen_classes is not None:
for class_ in unseen_classes:
self._class_mask[class_] = 0.
@property
def nO(self):
return self.state2vec.nO
def class_is_unseen(self, class_):
return self._class_mask[class_]
def mark_class_unseen(self, class_):
self._class_mask[class_] = 0
def mark_class_seen(self, class_):
self._class_mask[class_] = 1
def begin_update(self, states, drop=0.):
token_ids = self.get_token_ids(states)
vector, get_d_tokvecs = self.state2vec.begin_update(token_ids, drop=0.0)
2018-05-19 17:24:34 +00:00
mask = self.vec2scores.ops.get_dropout_mask(vector.shape, drop)
if mask is not None:
vector *= mask
scores, get_d_vector = self.vec2scores.begin_update(vector, drop=drop)
# If the class is unseen, make sure its score is minimum
scores[:, self._class_mask == 0] = numpy.nanmin(scores)
def backprop_parser_step(d_scores, sgd=None):
# Zero vectors for unseen classes
d_scores *= self._class_mask
d_vector = get_d_vector(d_scores, sgd=sgd)
if mask is not None:
d_vector *= mask
2018-05-19 17:24:34 +00:00
if isinstance(self.state2vec.ops, CupyOps) \
and not isinstance(token_ids, self.state2vec.ops.xp.ndarray):
# Move token_ids and d_vector to GPU, asynchronously
self.backprops.append((
util.get_async(self.cuda_stream, token_ids),
util.get_async(self.cuda_stream, d_vector),
get_d_tokvecs
))
else:
self.backprops.append((token_ids, d_vector, get_d_tokvecs))
return None
return scores, backprop_parser_step
def get_token_ids(self, batch):
states = _beam_utils.collect_states(batch)
cdef StateClass state
states = [state for state in states if not state.is_final()]
cdef np.ndarray ids = numpy.zeros((len(states), self.state2vec.nF),
dtype='i', order='C')
ids.fill(-1)
c_ids = <int*>ids.data
for state in states:
state.c.set_context_tokens(c_ids, ids.shape[1])
c_ids += ids.shape[1]
return ids
def make_updates(self, sgd):
# Tells CUDA to block, so our async copies complete.
if self.cuda_stream is not None:
self.cuda_stream.synchronize()
# Add a padding vector to the d_tokvecs gradient, so that missing
# values don't affect the real gradient.
d_tokvecs = self.ops.allocate((self.tokvecs.shape[0]+1, self.tokvecs.shape[1]))
for ids, d_vector, bp_vector in self.backprops:
d_state_features = bp_vector((d_vector, ids), sgd=sgd)
ids = ids.flatten()
d_state_features = d_state_features.reshape(
(ids.size, d_state_features.shape[2]))
self.ops.scatter_add(d_tokvecs, ids,
d_state_features)
# Padded -- see update()
self.bp_tokvecs(d_tokvecs[:-1], sgd=sgd)
return d_tokvecs
cdef class precompute_hiddens:
"""Allow a model to be "primed" by pre-computing input features in bulk.
This is used for the parser, where we want to take a batch of documents,
and compute vectors for each (token, position) pair. These vectors can then
be reused, especially for beam-search.
Let's say we're using 12 features for each state, e.g. word at start of
buffer, three words on stack, their children, etc. In the normal arc-eager
system, a document of length N is processed in 2*N states. This means we'll
create 2*N*12 feature vectors --- but if we pre-compute, we only need
N*12 vector computations. The saving for beam-search is much better:
if we have a beam of k, we'll normally make 2*N*12*K computations --
so we can save the factor k. This also gives a nice CPU/GPU division:
we can do all our hard maths up front, packed into large multiplications,
and do the hard-to-program parsing on the CPU.
"""
cdef readonly int nF, nO, nP
cdef bint _is_synchronized
cdef public object ops
cdef np.ndarray _features
cdef np.ndarray _cached
cdef np.ndarray bias
cdef object _cuda_stream
cdef object _bp_hiddens
def __init__(self, batch_size, tokvecs, lower_model, cuda_stream=None,
drop=0.):
gpu_cached, bp_features = lower_model.begin_update(tokvecs, drop=drop)
cdef np.ndarray cached
if not isinstance(gpu_cached, numpy.ndarray):
# Note the passing of cuda_stream here: it lets
# cupy make the copy asynchronously.
# We then have to block before first use.
cached = gpu_cached.get(stream=cuda_stream)
else:
cached = gpu_cached
if not isinstance(lower_model.b, numpy.ndarray):
self.bias = lower_model.b.get()
else:
self.bias = lower_model.b
self.nF = cached.shape[1]
self.nP = getattr(lower_model, 'nP', 1)
self.nO = cached.shape[2]
self.ops = lower_model.ops
self._is_synchronized = False
self._cuda_stream = cuda_stream
self._cached = cached
self._bp_hiddens = bp_features
cdef const float* get_feat_weights(self) except NULL:
if not self._is_synchronized and self._cuda_stream is not None:
self._cuda_stream.synchronize()
self._is_synchronized = True
return <float*>self._cached.data
def __call__(self, X):
return self.begin_update(X)[0]
def begin_update(self, token_ids, drop=0.):
cdef np.ndarray state_vector = numpy.zeros(
(token_ids.shape[0], self.nO, self.nP), dtype='f')
# This is tricky, but (assuming GPU available);
# - Input to forward on CPU
# - Output from forward on CPU
# - Input to backward on GPU!
# - Output from backward on GPU
bp_hiddens = self._bp_hiddens
feat_weights = self.get_feat_weights()
cdef int[:, ::1] ids = token_ids
sum_state_features(<float*>state_vector.data,
feat_weights, &ids[0,0],
token_ids.shape[0], self.nF, self.nO*self.nP)
state_vector += self.bias
state_vector, bp_nonlinearity = self._nonlinearity(state_vector)
def backward(d_state_vector_ids, sgd=None):
d_state_vector, token_ids = d_state_vector_ids
d_state_vector = bp_nonlinearity(d_state_vector, sgd)
# This will usually be on GPU
if not isinstance(d_state_vector, self.ops.xp.ndarray):
d_state_vector = self.ops.xp.array(d_state_vector)
d_tokens = bp_hiddens((d_state_vector, token_ids), sgd)
return d_tokens
return state_vector, backward
def _nonlinearity(self, state_vector):
if self.nP == 1:
state_vector = state_vector.reshape(state_vector.shape[:-1])
mask = state_vector >= 0.
state_vector *= mask
else:
state_vector, mask = self.ops.maxout(state_vector)
def backprop_nonlinearity(d_best, sgd=None):
# Fix nans (which can occur from unseen classes.)
d_best[self.ops.xp.isnan(d_best)] = 0.
if self.nP == 1:
d_best *= mask
d_best = d_best.reshape((d_best.shape + (1,)))
return d_best
else:
return self.ops.backprop_maxout(d_best, mask, self.nP)
return state_vector, backprop_nonlinearity