Skip to content

Commit

Permalink
Merge a452764 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 + a452764 commit dbf0420
Show file tree
Hide file tree
Showing 5 changed files with 185 additions and 137 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
239 changes: 143 additions & 96 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 Down Expand Up @@ -48,46 +48,102 @@ 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)
]
# Want to fix this slicing function.
# class MySlice(Function):
# @staticmethod
# def forward(ctx, alpha, beta, s1, s2, e, a, b, c, d):
# indices = torch.tensor([s1, s2, e, a, b, c, d])
# ctx.save_for_backward(indices)
# return alpha[e, :, a:b, c:d]

# @staticmethod
# def backward(ctx, grad_v):
# a, = ctx.saved_tensors
# itertools(a[2], :, a[3]:a[4], a[5]:a[6])
# beta[a[0]][a[1]][a[2], :, a[3]:a[4], a[5]:a[6]] += grad_v
# return None, None, None, None, None, None, None, None, None

# s = MySlice.apply
def s(input, e, a, b, c, d):
return input[e, :, a:b, c:d]

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]

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]
(ACR2,
BCL2,
ACL2,
AIR2,
BIL2,
BCR2) = self._make_chart(6, (batch, N-k, k), arc_scores)

if k > 1:
ACR2[:, :, :k-1] = ACR[:, :N-k, :k-1]
ACR2[:, :, k-1] = ACR_next[:, :-1]
else:
ACR2[:] = alpha[A][C][R, :, :N - k, :k]

if k > 1:
ACL2[:, :, :k-1] = ACL[:, :N-k, :k-1]
ACL2[:, :, k-1] = ACL_next[:, :-1]
else:
ACL2[:] = alpha[A][C][L, :, :N - k, :k]

if k > 1:
BCL2[:, :, 1:] = BCL[:, 1:, :]
BCL2[:, :, 0] = ACL_next[:, 1:]
else:
BCL2[:] = alpha[B][C][L, :, k:N, N - k:]

if k > 1:
BCR2[:, :, 1:] = BCR[:, 1:, :]
BCR2[:, :, 0] = ACR_next[:, 1:]
else:
BCR2[:] = alpha[B][C][R, :, k:N, N-k:]

start = semiring.dot(BCL2, ACR2)
arcs[k] = stack(semiring.times(start, arc_scores[:, f[1], f[0]]),
semiring.times(start, arc_scores[:, f[0], f[1]]))

AIR2[:, :, :k-1] = AIR[:, :N-k, :k-1]
AIR2[:, :, k-1] = arcs[k][R]

BIL2[:, :, 1:] = BIL[:, 1:, :k-1]
BIL2[:, :, 0] = arcs[k][L]

ACL_next = semiring.dot(ACL2, BIL2)
ACR_next = semiring.dot(AIR2, BCR2)

alpha[A][C][R, :, : N - k, k] = ACR_next
ACR = ACR2
BCL = BCL2
ACL = ACL2
AIR = AIR2
BIL = BIL2
BCR = BCR2

v = torch.stack([alpha[A][C][R, i, 0, l] for i, l in enumerate(lengths)])
print(v)
return (v, arcs[1:], alpha)

def _check_potentials(self, arc_scores, lengths=None):
Expand Down Expand Up @@ -116,99 +172,90 @@ 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,
)
)
alpha[A][C][L, :, 1 : N - k, k] = a[1]
alpha[A][C][R, :, : N - k - 1, k] = a[0]

c = semiring.dot(*roll(alpha_in[B][I][R], alpha[B][C][R], N, k, 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)
)

# 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)
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,
)
)

# 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
Loading

0 comments on commit dbf0420

Please sign in to comment.