explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-07-10 16:22:29 +03:00
Code Issues Packages Projects Releases Wiki Activity

Get non-reference forward func working

Browse Source
This commit is contained in:
Matthew Honnibal 2021-11-01 01:32:29 +01:00
parent 4e894e1076
commit 13b0a24870
1 changed files with 36 additions and 74 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

106
spacy/ml/tb_framework.py
View File

@ -28,7 +28,7 @@ def TransitionModel(
return Model( return Model(
name="parser_model", name="parser_model",
forward=_forward_reference, forward=forward,
init=init, init=init,
layers=[tok2vec_projected], layers=[tok2vec_projected],
refs={"tok2vec": tok2vec_projected}, refs={"tok2vec": tok2vec_projected},
@ -141,12 +141,12 @@ def forward(model, docs_moves: Tuple[List[Doc], TransitionSystem], is_train: boo
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))
feats, backprop_feats = _forward_precomputable_affine(model, tokvecs, is_train) feats, backprop_feats = _forward_precomputable_affine(model, tokvecs, is_train)
all_ids = [] all_ids = []
all_which = [] all_which = []
all_statevecs = [] all_statevecs = []
all_scores = [] all_scores = []
all_tokfeats = []
next_states = [s for s in states if not s.is_final()] next_states = [s for s in states if not s.is_final()]
unseen_mask = _get_unseen_mask(model) unseen_mask = _get_unseen_mask(model)
ids = numpy.zeros((len(states), nF), dtype="i") ids = numpy.zeros((len(states), nF), dtype="i")
@ -155,11 +155,16 @@ def forward(model, docs_moves: Tuple[List[Doc], TransitionSystem], is_train: boo
ids = ids[: len(next_states)] ids = ids[: len(next_states)]
for i, state in enumerate(next_states): for i, state in enumerate(next_states):
state.set_context_tokens(ids, i, nF) state.set_context_tokens(ids, i, nF)
preacts = feats[ids, arange].sum(axis=1) # type: ignore # Sum the state features, add the bias and apply the activation (maxout)
# to create the state vectors.
preacts2f = feats[ids, arange].sum(axis=1) # type: ignore
preacts2f += lower_b
preacts = model.ops.reshape3f(preacts2f, preacts2f.shape[0], nH, nP)
assert preacts.shape[0] == len(next_states), preacts.shape
statevecs, which = ops.maxout(preacts) statevecs, which = ops.maxout(preacts)
# Multiply the state-vector by the scores weights and add the bias, # Multiply the state-vector by the scores weights and add the bias,
# to get the logits. # to get the logits.
scores = ops.gemm(statevecs, upper_W, trans2=True) scores = model.ops.gemm(statevecs, upper_W, trans2=True)
scores += upper_b scores += upper_b
scores[:, unseen_mask == 0] = model.ops.xp.nanmin(scores) scores[:, unseen_mask == 0] = model.ops.xp.nanmin(scores)
# Transition the states, filtering out any that are finished. # Transition the states, filtering out any that are finished.
@ -167,17 +172,15 @@ def forward(model, docs_moves: Tuple[List[Doc], TransitionSystem], is_train: boo
all_scores.append(scores) all_scores.append(scores)
if is_train: if is_train:
# Remember intermediate results for the backprop. # Remember intermediate results for the backprop.
all_tokfeats.append(tokfeats)
all_ids.append(ids.copy()) all_ids.append(ids.copy())
all_statevecs.append(statevecs) all_statevecs.append(statevecs)
all_which.append(which) all_which.append(which)
nS = sum(len(s.history) for s in states)
def backprop_parser(d_states_d_scores): def backprop_parser(d_states_d_scores):
d_tokvecs = model.ops.alloc2f(tokvecs.shape[0], tokvecs.shape[1]) d_tokvecs = model.ops.alloc2f(tokvecs.shape[0], tokvecs.shape[1])
ids = model.ops.xp.vstack(all_ids) ids = model.ops.xp.vstack(all_ids)
which = ops.xp.vstack(all_which) which = ops.xp.vstack(all_which)
statevecs = model.ops.xp.vstack(all_statevecs)
_, d_scores = d_states_d_scores _, d_scores = d_states_d_scores
if model.attrs.get("unseen_classes"): if model.attrs.get("unseen_classes"):
# If we have a negative gradient (i.e. the probability should # If we have a negative gradient (i.e. the probability should
@ -187,32 +190,30 @@ def forward(model, docs_moves: Tuple[List[Doc], TransitionSystem], is_train: boo
if (d_scores[:, clas] < 0).any(): if (d_scores[:, clas] < 0).any():
model.attrs["unseen_classes"].remove(clas) model.attrs["unseen_classes"].remove(clas)
d_scores *= unseen_mask d_scores *= unseen_mask
statevecs = ops.xp.vstack(all_statevecs)
tokfeats = ops.xp.vstack(all_tokfeats)
assert statevecs.shape == (nS, nH), statevecs.shape
assert d_scores.shape == (nS, nO), d_scores.shape
# Calculate the gradients for the parameters of the upper layer. # Calculate the gradients for the parameters of the upper layer.
# The weight gemm is (nS, nO) @ (nS, nH).T
model.inc_grad("upper_b", d_scores.sum(axis=0)) model.inc_grad("upper_b", d_scores.sum(axis=0))
model.inc_grad("upper_W", model.ops.gemm(d_scores, statevecs, trans1=True)) model.inc_grad("upper_W", model.ops.gemm(d_scores, statevecs, trans1=True))
# Now calculate d_statevecs, by backproping through the upper linear layer. # Now calculate d_statevecs, by backproping through the upper linear layer.
# This gemm is (nS, nO) @ (nO, nH)
d_statevecs = model.ops.gemm(d_scores, upper_W) d_statevecs = model.ops.gemm(d_scores, upper_W)
# Backprop through the maxout activation # Backprop through the maxout activation
d_preacts = model.ops.backprop_maxout(d_statevecs, which, model.get_dim("nP")) d_preacts = model.ops.backprop_maxout(d_statevecs, which, nP)
model.inc_grad("lower_b", d_preacts.sum(axis=0)) d_preacts2f = model.ops.reshape2f(d_preacts, d_preacts.shape[0], nH * nP)
model.inc_grad("lower_W", model.ops.gemm(d_preacts, tokfeats, trans1=True)) model.inc_grad("lower_b", d_preacts2f.sum(axis=0))
# We don't need to backprop the summation, because we pass back the IDs instead # We don't need to backprop the summation, because we pass back the IDs instead
d_state_features = backprop_feats((d_preacts, all_ids)) d_state_features = backprop_feats((d_preacts2f, ids))
ids1d = model.ops.xp.vstack(all_ids).flatten() d_tokvecs = model.ops.alloc2f(tokvecs.shape[0], tokvecs.shape[1])
d_state_features = d_state_features.reshape((ids1d.size, -1)) model.ops.scatter_add(d_tokvecs, ids, d_state_features)
d_tokvecs = model.ops.alloc((tokvecs.shape[0] + 1, tokvecs.shape[1])) model.inc_grad("lower_pad", d_tokvecs[-1])
model.ops.scatter_add(d_tokvecs, ids1d, d_state_features) return (backprop_tok2vec(d_tokvecs[:-1]), None)
return (backprop_tok2vec(d_tokvecs), None)
return (states, all_scores), backprop_parser return (states, all_scores), backprop_parser
def _forward_reference(
def _forward_reference(model, docs_moves: Tuple[List[Doc], TransitionSystem], is_train: bool): model, docs_moves: Tuple[List[Doc], TransitionSystem], is_train: bool
):
"""Slow reference implementation, without the precomputation""" """Slow reference implementation, without the precomputation"""
nF = model.get_dim("nF") nF = model.get_dim("nF")
tok2vec = model.get_ref("tok2vec") tok2vec = model.get_ref("tok2vec")
@ -311,7 +312,6 @@ def _forward_reference(model, docs_moves: Tuple[List[Doc], TransitionSystem], is
return (states, all_scores), backprop_parser return (states, all_scores), backprop_parser
def _get_unseen_mask(model: Model) -> Floats1d: def _get_unseen_mask(model: Model) -> Floats1d:
mask = model.ops.alloc1f(model.get_dim("nO")) mask = model.ops.alloc1f(model.get_dim("nO"))
mask.fill(1) mask.fill(1)
@ -321,17 +321,18 @@ def _get_unseen_mask(model: Model) -> Floats1d:
def _forward_precomputable_affine(model, X: Floats2d, is_train: bool): def _forward_precomputable_affine(model, X: Floats2d, is_train: bool):
W: Floats2d = model.get_param("lower_W")
W: Floats4d = model.get_param("lower_W")
pad: Floats4d = model.get_param("lower_pad")
nF = model.get_dim("nF") nF = model.get_dim("nF")
nH = model.get_dim("nH") nH = model.get_dim("nH")
nP = model.get_dim("nP") nP = model.get_dim("nP")
nI = model.get_dim("nI") nI = model.get_dim("nI")
# The weights start out (nH * nP, nF * nI). Transpose and reshape to (nF * nH *nP, nI)
W3f = model.ops.reshape3f(W, nH * nP, nF, nI)
W3f = W3f.transpose((1, 0, 2))
W2f = model.ops.reshape2f(W3f, nF * nH * nP, nI)
assert X.shape == (X.shape[0], nI), X.shape assert X.shape == (X.shape[0], nI), X.shape
Yf_ = model.ops.gemm(X, model.ops.reshape2f(W, nF * nH * nP, nI), trans2=True) Yf_ = model.ops.gemm(X, W2f, trans2=True)
Yf = model.ops.reshape4f(Yf_, Yf_.shape[0], nF, nH, nP) Yf = model.ops.reshape3f(Yf_, Yf_.shape[0], nF, nH * nP)
Yf = model.ops.xp.vstack((Yf, pad))
def backward(dY_ids: Tuple[Floats3d, Ints2d]): def backward(dY_ids: Tuple[Floats3d, Ints2d]):
# This backprop is particularly tricky, because we get back a different # This backprop is particularly tricky, because we get back a different
@ -348,54 +349,15 @@ def _forward_precomputable_affine(model, X: Floats2d, is_train: bool):
# However, we avoid building that array for efficiency -- and just pass # However, we avoid building that array for efficiency -- and just pass
# in the indices. # in the indices.
dY, ids = dY_ids dY, ids = dY_ids
assert dY.ndim == 3 dXf = model.ops.gemm(dY, W)
assert dY.shape[1] == nH, dY.shape Xf = X[ids].reshape((ids.shape[0], -1))
assert dY.shape[2] == nP, dY.shape dW = model.ops.gemm(dY, Xf, trans1=True)
# nB = dY.shape[0] model.inc_grad("lower_W", dW)
# model.inc_grad(
# "lower_pad", _backprop_precomputable_affine_padding(model, dY, ids)
# )
# model.inc_grad("lower_b", dY.sum(axis=0)) # type: ignore
dY = model.ops.reshape2f(dY, dY.shape[0], nH * nP)
Wopfi = W.transpose((1, 2, 0, 3))
Wopfi = Wopfi.reshape((nH * nP, nF * nI))
dXf = model.ops.gemm(dY.reshape((dY.shape[0], nH * nP)), Wopfi)
ids1d = model.ops.xp.vstack(ids).flatten()
Xf = model.ops.reshape2f(X[ids1d], -1, nF * nI)
dWopfi = model.ops.gemm(dY, Xf, trans1=True)
dWopfi = dWopfi.reshape((nH, nP, nF, nI))
# (o, p, f, i) --> (f, o, p, i)
dWopfi = dWopfi.transpose((2, 0, 1, 3))
model.inc_grad("lower_W", dWopfi)
return model.ops.reshape3f(dXf, dXf.shape[0], nF, nI) return model.ops.reshape3f(dXf, dXf.shape[0], nF, nI)
return Yf, backward return Yf, backward
def _backprop_precomputable_affine_padding(model, dY, ids):
ids = model.ops.xp.vstack(ids)
nB = dY.shape[0]
nF = model.get_dim("nF")
nP = model.get_dim("nP")
nH = model.get_dim("nH")
# Backprop the "padding", used as a filler for missing values.
# Values that are missing are set to -1, and each state vector could
# have multiple missing values. The padding has different values for
# different missing features. The gradient of the padding vector is:
#
# for b in range(nB):
# for f in range(nF):
# if ids[b, f] < 0:
# d_pad[f] += dY[b]
#
# Which can be rewritten as:
#
# (ids < 0).T @ dY
mask = model.ops.asarray(ids < 0, dtype="f")
d_pad = model.ops.gemm(mask, dY.reshape(nB, nH * nP), trans1=True)
return d_pad.reshape((1, nF, nH, nP))
def _infer_nO(Y: Optional[Tuple[List[State], List[Floats2d]]]) -> Optional[int]: def _infer_nO(Y: Optional[Tuple[List[State], List[Floats2d]]]) -> Optional[int]:
if Y is None: if Y is None:
return None return None