Skip to content

Commit

Permalink
Merge 47fc6b2 into 6f06de9
Browse files Browse the repository at this point in the history
  • Loading branch information
@srush
srush committed Sep 9, 2019
2 parents 6f06de9 + 47fc6b2 commit 31c9002
Show file tree
Hide file tree
Showing 5 changed files with 143 additions and 142 deletions.
4 changes: 2 additions & 2 deletions torch_struct/cky.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@ def backward(ctx, grad_v):


class CKY(_Struct):
def sum(self, scores, lengths=None, force_grad=False, _autograd=False):
def sum(self, scores, lengths=None, force_grad=False, _autograd=True):
"""
Compute the inside pass of a CFG using CKY.
Expand Down Expand Up @@ -162,7 +162,7 @@ def _dp_backward(self, scores, lengths, alpha_in, v, force_grad=False):

return (term_marginals, edge_marginals, root_marginals)

def marginals(self, scores, lengths=None, _autograd=False):
def marginals(self, scores, lengths=None, _autograd=True):
"""
Compute the marginals of a CFG using CKY.
Expand Down
202 changes: 101 additions & 101 deletions torch_struct/deptree.py
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
import torch
import itertools
from .helpers import _Struct, roll
from .helpers import _Struct, roll2


def _convert(logits):
Expand All @@ -10,9 +10,10 @@ def _convert(logits):
).type_as(logits.data)
new_logits.fill_(-1e9)
new_logits[:, 1:, 1:] = logits
for i in range(0, logits.size(1)):
new_logits[:, 0, i + 1] = logits[:, i, i]
new_logits[:, i + 1, i + 1] = -1e9

N = logits.size(1)
new_logits[:, 0, 1:] = logits[:, torch.arange(N), torch.arange(N)]
new_logits[:, torch.arange(1, N), torch.arange(1, N)] = -1e9
return new_logits


Expand Down Expand Up @@ -48,46 +49,54 @@ def _dp(self, arc_scores, lengths=None, force_grad=False):

DIRS = 2

alpha = [
self._make_chart(2, (DIRS, batch, N, N), arc_scores, force_grad)
for _ in range(2)
]

def stack(a, b):
return torch.stack([a, b])

def sstack(a):
return torch.stack([a, a])

alpha = [
self._make_chart(2, (DIRS, batch, N, N), arc_scores, force_grad)
for _ in range(2)
]
arcs = self._make_chart(N, (DIRS, batch, N), arc_scores, force_grad)
arcs = [self._make_chart(1, (DIRS, batch, N-k), arc_scores, force_grad)[0]
for k in range(N)]

# Inside step. assumes first token is root symbol
alpha[A][C][:, :, :, 0].data.fill_(semiring.one())
alpha[B][C][:, :, :, -1].data.fill_(semiring.one())

k = 0

AIR = alpha[A][I][R, :, :N-k, 1:k]
BIL = alpha[B][I][L, :, k:N, N-k:N-1]
k = 1
AC2 = alpha[A][C][:, :, :N - k, :k]
BC2 = alpha[B][C][:, :, k:, N - k:]
ends = [None]
for k in range(1, N):
f = torch.arange(N - k), torch.arange(k, N)
arcs[k] = semiring.dot(
sstack(alpha[A][C][R, :, : N - k, :k]),
sstack(alpha[B][C][L, :, k:, N - k :]),
stack(arc_scores[:, f[1], f[0]], arc_scores[:, f[0], f[1]]).unsqueeze(
-1
),
)
alpha[A][I][:, :, : N - k, k] = arcs[k]
alpha[B][I][:, :, k:N, N - k - 1] = alpha[A][I][:, :, : N - k, k]
alpha[A][C][:, :, : N - k, k] = semiring.dot(
stack(
alpha[A][C][L, :, : N - k, :k],
alpha[A][I][R, :, : N - k, 1 : k + 1],
),
stack(
alpha[B][I][L, :, k:, N - k - 1 : N - 1],
alpha[B][C][R, :, k:, N - k :],
),
)
alpha[B][C][:, :, k:N, N - k - 1] = alpha[A][C][:, :, : N - k, k]
v = torch.stack([alpha[A][C][R, i, 0, l] for i, l in enumerate(lengths)])
print(v)
if k > 1:
AC2 = torch.cat([AC[:, :, :-1], AC_next[:, :, :-1].unsqueeze(-1)],
dim=3)
if k > 1:
BC2 = torch.cat([AC_next[:, :, 1:].unsqueeze(-1), BC[:, :, 1:]], dim=3)

start = semiring.dot(BC2[L], AC2[R])
arcs[k] = stack(semiring.times(start, arc_scores[:, f[1], f[0]]),
semiring.times(start, arc_scores[:, f[0], f[1]]))

AIR2 = torch.cat([AIR[:, :-1], arcs[k][R].unsqueeze(-1)], dim=2)
BIL2 = torch.cat([arcs[k][L].unsqueeze(-1), BIL[:, 1:]], dim=2)
AC_next = stack(semiring.dot(AC2[L], BIL2), semiring.dot(AIR2, BC2[R]))

ends.append(AC_next[R, :, 0])
AC = AC2
BC = BC2
AIR = AIR2
BIL = BIL2
v = torch.stack([ends[l][i] for i, l in enumerate(lengths)])
# v = torch.stack([alpha[A][C][R, i, 0, l] for i, l in enumerate(lengths)])
return (v, arcs[1:], alpha)

def _check_potentials(self, arc_scores, lengths=None):
Expand Down Expand Up @@ -116,99 +125,90 @@ def _dp_backward(self, arc_scores, lengths, alpha_in, v=None, force_grad=False):
for _ in range(2)
]

def stack(a, b):
return torch.stack([a, b], dim=-1)
def stack(a, b, dim=-1):
return torch.stack([a, b], dim=dim)

def sstack(a):
return torch.stack([a, a], dim=-1)

for k in range(N - 1, -1, -1):
# Initialize
for b, l in enumerate(lengths):
alpha[A][C][R, b, 0, l] = semiring.one()
alpha[B][C][R, b, l, N - l - 1] = semiring.one()

# R completes
# I -> C* C
# I -> C* C
# C -> I C*
a = semiring.dot(
*roll(
stack(alpha[A][I][R], alpha[A][I][L]),
sstack(alpha_in[A][C][L]),
N,
k,
1,
if N - k - 1 > 0:
# R completes
# I -> C* C
# I -> C* C
# C -> I C*
a = semiring.sum(
semiring.times(
*roll2(
stack(
stack(alpha[A][I][R], alpha[A][I][L]),
sstack(alpha_in[B][C][R]),
dim=0,
),
stack(
sstack(alpha_in[A][C][L]),
stack(alpha[B][I][L], alpha[B][I][R]),
dim=0,
),
N,
k,
1,
)
).view(2, batch, N - k - 1, -1),
dim=-1,
)
)

c = semiring.dot(*roll(alpha_in[B][I][R], alpha[B][C][R], N, k, 0))
alpha[A][C][L, :, 1 : N - k, k] = a[1]
alpha[A][C][R, :, : N - k - 1, k] = a[0]

alpha[A][C][R, :, : N - k - 1, k] = semiring.plus(
semiring.sum(a), alpha[A][C][R, :, : N - k - 1, k]
)

alpha[A][C][R][:, : N - k, k] = semiring.plus(
alpha[A][C][R][:, : N - k, k], c
)

# L completes
# I -> C* C
# I -> C* C
# C -> I C*
a = semiring.dot(
*roll(
sstack(alpha_in[B][C][R]),
stack(alpha[B][I][L], alpha[B][I][R]),
N,
k,
1,
for b, l in enumerate(lengths):
if l == k:
alpha[A][C][R, b, 0, l] = semiring.one()
alpha[B][C][R, b, l, N - l - 1] = semiring.one()

c = semiring.sum(
semiring.times(
*roll2(
stack(alpha[A][C][L], alpha_in[B][I][R], dim=0),
stack(alpha_in[A][I][L], alpha[B][C][R], dim=0),
N,
k,
0,
)
)
)

c = semiring.dot(*roll(alpha[A][C][L], alpha_in[A][I][L], N, k, 0))

alpha[A][C][L, :, 1 : N - k, k] = semiring.plus(
semiring.sum(a), alpha[A][C][L, :, 1 : N - k, k]
)
alpha[A][C][L][:, : N - k, k] = semiring.plus(
c, alpha[A][C][L][:, : N - k, k]
alpha[A][C][:, :, : N - k, k] = semiring.plus(
alpha[A][C][:, :, : N - k, k], c
)

# Compute reverses.
alpha[B][C][:, :, k:N, N - k - 1] = alpha[A][C][:, :, : N - k, k]

if k > 0:
f = torch.arange(N - k), torch.arange(k, N)

# Incomplete
alpha[A][I][R][:, : N - k, k] = semiring.dot(
arc_scores[:, f[0], f[1]].unsqueeze(-1),
*roll(alpha[A][C][R], alpha_in[A][C][R], N, k)
alpha[A][I][:, :, : N - k, k] = semiring.dot(
stack(
arc_scores[:, f[1], f[0]], arc_scores[:, f[0], f[1]], dim=0
).unsqueeze(-1),
*roll2(
stack(alpha_in[B][C][L], alpha[A][C][R], dim=0),
stack(alpha[B][C][L], alpha_in[A][C][R], dim=0),
N,
k,
)
)

# C -> C I
alpha[A][I][L][:, : N - k, k] = semiring.dot(
arc_scores[:, f[1], f[0]].unsqueeze(-1),
*roll(alpha_in[B][C][L], alpha[B][C][L], N, k)
)

# Compute reverses
alpha[B][I][:, :, k:N, N - k - 1] = alpha[A][I][:, :, : N - k, k]

v = alpha[A][C][R, :, 0, 0]
left = semiring.times(alpha[A][I][L, :, :, :], alpha_in[A][I][L, :, :, :])
right = semiring.times(alpha[A][I][R, :, :, :], alpha_in[A][I][R, :, :, :])

ret = torch.zeros(batch, N, N).type_as(left)
for k in range(N):
for d in range(N - k):
ret[:, k + d, k] = semiring.div_exp(
left[:, k, d] - arc_scores[:, k + d, k], v.view(batch)
)
ret[:, k, k + d] = semiring.div_exp(
right[:, k, d] - arc_scores[:, k, k + d], v.view(batch)
)
for k in torch.arange(N):
f = torch.arange(N - k), torch.arange(k, N)
ret[:, f[1], k] = left[:, k, f[0]]
ret[:, k, f[1]] = right[:, k, f[0]]

ret = semiring.div_exp(ret - arc_scores, v.view(batch, 1, 1))
return _unconvert(ret)

def _arrange_marginals(self, grads):
Expand Down
49 changes: 22 additions & 27 deletions torch_struct/helpers.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,29 +7,10 @@ def roll(a, b, N, k, gap=0):
return (a[:, : N - (k + gap), (k + gap) :], b[:, k + gap :, : N - (k + gap)])


class DPManual(Function):
@staticmethod
def forward(ctx, obj, input, lengths):
with torch.no_grad():
v, _, alpha = obj._dp(input, lengths, False)
ctx.obj = obj
ctx.lengths = lengths
ctx.alpha = alpha

if isinstance(input, tuple):
ctx.save_for_backward(*input)
else:
ctx.save_for_backward(input)
return v
def roll2(a, b, N, k, gap=0):
return (a[:, :, : N - (k + gap), (k + gap) :], b[:, :, k + gap :, : N - (k + gap)])


@staticmethod
def backward(ctx, grad_v):
input = ctx.saved_tensors
if len(input) == 1:
input = input[0]
with torch.no_grad():
marginals = ctx.obj._dp_backward(input, ctx.lengths, ctx.alpha)
return None, marginals, None


class _Struct:
Expand All @@ -40,7 +21,7 @@ def score(self, potentials, parts):
batch = potentials.shape[0]
return torch.mul(potentials, parts).view(batch, -1).sum(-1)

def _make_chart(self, N, size, potentials, force_grad):
def _make_chart(self, N, size, potentials, force_grad=False):
return [
(
torch.zeros(*size)
Expand All @@ -61,16 +42,30 @@ def sum(self, edge, lengths=None, _autograd=True):
Returns:
v: b tensor of total sum
"""


if (
_autograd
or self.semiring is not LogSemiring
or not hasattr(self, "_dp_backward")
):
return self._dp(edge, lengths)[0]
else:
return DPManual.apply(self, edge, lengths)
v, _, alpha = self._dp(edge, lengths, False)

class DPManual(Function):
@staticmethod
def forward(ctx, input):
return v

@staticmethod
def backward(ctx, grad_v):
marginals = self._dp_backward(edge, lengths, alpha)
return marginals.mul(
grad_v.view((grad_v.shape[0],) + tuple([1] * marginals.dim())))

return DPManual.apply(edge)

def marginals(self, edge, lengths=None, _autograd=True):
"""
Expand All @@ -83,15 +78,15 @@ def marginals(self, edge, lengths=None, _autograd=True):
marginals: b x (N-1) x C x C table
"""
v, edge, alpha = self._dp(edge, lengths=lengths, force_grad=True)
v, edges, alpha = self._dp(edge, lengths=lengths, force_grad=True)
if (
_autograd
or self.semiring is not LogSemiring
or not hasattr(self, "_dp_backward")
):
marg = torch.autograd.grad(
v.sum(dim=0),
edge,
edges,
create_graph=True,
only_inputs=True,
allow_unused=False,
Expand Down
2 changes: 1 addition & 1 deletion torch_struct/semirings.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ def one():

@staticmethod
def div_exp(a, b):
return a.exp().div(b.exp())
return (a - b).exp()


class LogSemiring(_BaseLog):
Expand Down
Loading

0 comments on commit 31c9002

Please sign in to comment.