Files
dmlc--dgl/python/dgl/backend/tensorflow/sparse.py
T
2026-07-13 13:35:51 +08:00

462 lines
14 KiB
Python

import numpy as np
import tensorflow as tf
from ..._sparse_ops import (
_bwd_segment_cmp,
_csrmask,
_csrmm,
_csrsum,
_gsddmm,
_gspmm,
_scatter_add,
_segment_reduce,
)
from ...base import ALL, is_all
from ...heterograph_index import create_unitgraph_from_csr
from .tensor import asnumpy, context, copy_to, tensor, zerocopy_from_numpy
__all__ = [
"gspmm",
"gsddmm",
"edge_softmax",
"segment_reduce",
"scatter_add",
"csrmm",
"csrsum",
"csrmask",
]
def _scatter_nd(index, src, n_rows):
assert index.shape == src.shape
shp = index.shape
ctx = context(src)
ndim = index.ndim
offsets = []
stride = 1
for i in reversed(range(1, ndim)):
di = shp[i]
offset_i = tf.range(di, dtype=index.dtype)
offsets.append(
tf.reshape(
(stride * offset_i), (1,) * i + (di,) + (1,) * (ndim - 1 - i)
)
)
stride *= di
if ndim > 1:
new_idx = index * stride + copy_to(sum(offsets), ctx)
else:
new_idx = index
src = tf.reshape(src, (-1,))
new_idx = tf.reshape(new_idx, (-1, 1))
rst = tf.reshape(
tf.scatter_nd(new_idx, src, (stride * n_rows,)), (n_rows, *shp[1:])
)
return rst
def _gather_nd(index, src):
shp = index.shape
ctx = context(src)
ndim = index.ndim
offsets = []
stride = 1
for i in reversed(range(1, ndim)):
di = shp[i]
offset_i = tf.range(di, dtype=index.dtype)
offsets.append(
tf.reshape(
(stride * offset_i), (1,) * i + (di,) + (1,) * (ndim - 1 - i)
)
)
stride *= di
if ndim > 1:
new_idx = index * stride + copy_to(sum(offsets), ctx)
else:
new_idx = index
src = tf.reshape(src, (-1,))
new_idx = tf.reshape(new_idx, (-1))
rst = tf.reshape(tf.gather(src, new_idx), shp)
return rst
def _reduce_grad(grad, shape):
"""Reduce gradient on the broadcast dimension
If there is broadcast in forward pass, gradients need to be reduced on
broadcast dimension. This function checks the input tensor shape and
gradient shape and perform the reduction.
Parameters
----------
grad: Tensor
Gradient tensor
shape: tuple
Shape of input tensor
Returns
-------
Tensor
"""
grad_shape = grad.shape[1:]
in_shape = shape[1:]
if in_shape == grad_shape:
# no need to reduce
return grad
num_to_squeeze = len(grad_shape) - len(in_shape)
# pad inshape
in_shape = (1,) * num_to_squeeze + in_shape
reduce_idx = np.asarray(
np.nonzero(np.asarray(grad_shape) - np.asarray(in_shape))
)
reduce_idx += 1 # skip batch dim
reduce_idx_tensor = tf.constant(
tuple(reduce_idx.flatten().tolist()), dtype=tf.int32
)
grad = tf.reduce_sum(grad, axis=reduce_idx_tensor, keepdims=True)
return tf.reshape(grad, shape)
def _need_reduce_last_dim(ufeat, efeat):
"""Indicates whether to reduce the last dimension on edges
in the backward pass of spmm,
if so, use dot instead of mul."""
ushp = ufeat.shape
eshp = efeat.shape
return ushp[1:-1] == eshp[1:-1] and eshp[-1] == 1 and ushp[-1] > 1
def _muldiv(op, x):
return 1.0 / x if op == "div" else x
def _addsub(op, x):
return -x if op == "sub" else x
def _expand(x, shape):
return tf.broadcast_to(x, (x.shape[0], *shape))
def gspmm_real(gidx, op, reduce_op, X, Y):
out, (argX, argY) = _gspmm(gidx, op, reduce_op, X, Y)
def grad(dZ):
dZ = tensor(dZ)
if op != "copy_rhs":
g_rev = gidx.reverse()
if reduce_op == "sum":
if op in ["mul", "div"]:
dX = _gspmm(g_rev, "mul", "sum", dZ, _muldiv(op, Y))[0]
elif op in ["add", "sub"]:
dX = _gspmm(g_rev, "copy_lhs", "sum", dZ, Y)[0]
elif op == "copy_lhs":
dX = _gspmm(g_rev, "copy_lhs", "sum", dZ, None)[0]
else:
if op in ["mul", "div"]:
dX = _scatter_nd(
argX,
_muldiv(op, _gather_nd(argY, _expand(Y, dZ.shape[1:])))
* dZ,
X.shape[0],
)
elif op in ["add", "sub", "copy_lhs"]:
dX = _scatter_nd(argX, dZ, X.shape[0])
dX = _reduce_grad(dX, X.shape)
else:
dX = tf.zeros_like(X)
if op != "copy_lhs":
if reduce_op == "sum":
if op == "mul" and _need_reduce_last_dim(X, Y):
dY = _gsddmm(gidx, "dot", X, dZ)
elif op in ["mul", "div"]:
dY = _gsddmm(gidx, "mul", X, dZ)
if op == "div":
dY = -dY / (Y**2)
elif op in ["add", "sub", "copy_rhs"]:
dY = _gsddmm(gidx, "copy_rhs", X, _addsub(op, dZ))
else:
out_shp = (Y.shape[0],) + dZ.shape[1:]
if op in ["mul", "div"]:
dY = _scatter_nd(
argY,
_gather_nd(argX, _expand(X, dZ.shape[1:])) * dZ,
Y.shape[0],
)
if op == "div":
dY = -dY / (Y**2)
elif op in ["add", "sub", "copy_rhs"]:
dY = _scatter_nd(argY, _addsub(op, dZ), Y.shape[0])
dY = _reduce_grad(dY, Y.shape)
else:
dY = tf.zeros_like(Y)
return dX, dY
return out, grad
def gspmm(gidx, op, reduce_op, X, Y):
@tf.custom_gradient
def _lambda(X, Y):
return gspmm_real(gidx, op, reduce_op, X, Y)
if X is None:
X = tf.zeros(())
if Y is None:
Y = tf.zeros(())
return _lambda(X, Y)
def gsddmm_real(gidx, op, X, Y, lhs_target, rhs_target):
out = _gsddmm(gidx, op, X, Y, lhs_target, rhs_target)
def grad(dZ):
if op != "copy_rhs":
if lhs_target in ["u", "v"]:
_gidx = gidx if lhs_target == "v" else gidx.reverse()
if op in ["add", "sub", "copy_lhs"]:
dX = _gspmm(_gidx, "copy_rhs", "sum", None, dZ)[0]
else: # mul, div, dot
if rhs_target == lhs_target:
dX = _gspmm(_gidx, "copy_rhs", "sum", None, dZ)[
0
] * _muldiv(op, Y)
elif rhs_target == "e":
dX = _gspmm(
_gidx, "copy_rhs", "sum", None, dZ * _muldiv(op, Y)
)[0]
else: # rhs_target = !lhs_target
dX = _gspmm(_gidx, "mul", "sum", _muldiv(op, Y), dZ)[0]
else: # lhs_target == 'e'
if op in ["add", "sub", "copy_lhs"]:
dX = dZ
else: # mul, div, dot
dX = _gsddmm(
gidx, "mul", dZ, _muldiv(op, Y), "e", rhs_target
)
dX = _reduce_grad(dX, X.shape)
else:
dX = tf.zeros_like(X)
if op != "copy_lhs":
if rhs_target in ["u", "v"]:
_gidx = gidx if rhs_target == "v" else gidx.reverse()
if op in ["add", "sub", "copy_rhs"]:
dY = _gspmm(
_gidx, "copy_rhs", "sum", None, _addsub(op, dZ)
)[0]
else: # mul, div, dot
if lhs_target == rhs_target:
dY = _gspmm(_gidx, "copy_rhs", "sum", None, dZ)[0] * X
elif lhs_target == "e":
dY = _gspmm(_gidx, "copy_rhs", "sum", None, dZ * X)[0]
else: # rhs_target = !lhs_target
dY = _gspmm(_gidx, "mul", "sum", X, dZ)[0]
if op == "div":
dY = -dY / (Y**2)
else:
if op in ["add", "sub", "copy_rhs"]:
dY = _addsub(op, dZ)
else: # mul, div, dot
dY = _gsddmm(gidx, "mul", dZ, X, "e", lhs_target)
if op == "div":
dY = -dY / (Y**2)
dY = _reduce_grad(dY, Y.shape)
else:
dY = tf.zeros_like(Y)
return dX, dY
return out, grad
def gsddmm(gidx, op, X, Y, lhs_target="u", rhs_target="v"):
@tf.custom_gradient
def _lambda(X, Y):
return gsddmm_real(gidx, op, X, Y, lhs_target, rhs_target)
if X is None:
X = tf.zeros(())
if Y is None:
Y = tf.zeros(())
return _lambda(X, Y)
def edge_softmax_real(gidx, score, eids=ALL, norm_by="dst"):
if not is_all(eids):
gidx = gidx.edge_subgraph([eids], True).graph
if norm_by == "src":
gidx = gidx.reverse()
score_max = _gspmm(gidx, "copy_rhs", "max", None, score)[0]
score = tf.math.exp(_gsddmm(gidx, "sub", score, score_max, "e", "v"))
score_sum = _gspmm(gidx, "copy_rhs", "sum", None, score)[0]
out = _gsddmm(gidx, "div", score, score_sum, "e", "v")
def edge_softmax_backward(grad_out):
sds = out * grad_out
accum = gspmm(gidx, "copy_rhs", "sum", None, sds)
grad_score = sds - gsddmm(gidx, "mul", out, accum, "e", "v")
return grad_score
return out, edge_softmax_backward
def edge_softmax(gidx, logits, eids=ALL, norm_by="dst"):
@tf.custom_gradient
def _lambda(logits):
return edge_softmax_real(gidx, logits, eids, norm_by)
return _lambda(logits)
def segment_reduce_real(op, x, offsets):
y, arg = _segment_reduce(op, x, offsets)
def segment_reduce_backward(dy):
m = x.shape[0]
if op == "sum":
offsets_np = asnumpy(offsets[1:])
indices_np = np.zeros((m + 1,), dtype=offsets_np.dtype)
np.add.at(indices_np, offsets_np, np.ones_like(offsets_np))
indices_np = np.cumsum(indices_np, -1)[:-1]
indices = zerocopy_from_numpy(indices_np)
dx = tf.gather(dy, indices)
else:
dx = _bwd_segment_cmp(dy, arg, m)
return dx
return y, segment_reduce_backward
def segment_reduce(op, x, offsets):
@tf.custom_gradient
def _lambda(x):
return segment_reduce_real(op, x, offsets)
return _lambda(x)
def scatter_add_real(x, idx, m):
y = _scatter_add(x, idx, m)
def scatter_add_backward(dy):
return tf.gather(dy, idx)
return y, scatter_add_backward
def scatter_add(x, idx, m):
@tf.custom_gradient
def _lambda(x):
return scatter_add_real(x, idx, m)
return _lambda(x)
def csrmm_real(gidxA, A_weights, gidxB, B_weights, num_vtypes):
gidxC, C_weights = _csrmm(gidxA, A_weights, gidxB, B_weights, num_vtypes)
nrows, ncols, C_indptr, C_indices, C_eids = gidxC.adjacency_matrix_tensors(
0, False, "csr"
)
def grad(dnrows, dncols, dC_indptr, dC_indices, dC_eids, dC_weights):
# Only the last argument is meaningful.
dgidxA, dA_weights = _csrmm(
gidxC,
dC_weights,
gidxB.reverse(),
B_weights,
gidxA.number_of_ntypes(),
)
dgidxB, dB_weights = _csrmm(
gidxA.reverse(),
A_weights,
gidxC,
dC_weights,
gidxB.number_of_ntypes(),
)
dA_weights = _csrmask(dgidxA, dA_weights, gidxA)
dB_weights = _csrmask(dgidxB, dB_weights, gidxB)
return dA_weights, dB_weights
return (
tf.constant(nrows),
tf.constant(ncols),
C_indptr,
C_indices,
C_eids,
C_weights,
), grad
def csrmm(gidxA, A_weights, gidxB, B_weights, num_vtypes):
@tf.custom_gradient
def _lambda(A_weights, B_weights):
return csrmm_real(gidxA, A_weights, gidxB, B_weights, num_vtypes)
nrows, ncols, C_indptr, C_indices, C_eids, C_weights = _lambda(
A_weights, B_weights
)
gidxC = create_unitgraph_from_csr(
num_vtypes,
nrows.numpy(),
ncols.numpy(),
C_indptr,
C_indices,
C_eids,
["coo", "csr", "csc"],
)
return gidxC, C_weights
def csrsum_real(gidxs, weights):
gidxC, C_weights = _csrsum(gidxs, weights)
nrows, ncols, C_indptr, C_indices, C_eids = gidxC.adjacency_matrix_tensors(
0, False, "csr"
)
def grad(dnrows, dncols, dC_indptr, dC_indices, dC_eids, dC_weights):
# Only the last argument is meaningful.
return tuple(_csrmask(gidxC, dC_weights, gidx) for gidx in gidxs)
return (
tf.constant(nrows),
tf.constant(ncols),
C_indptr,
C_indices,
C_eids,
C_weights,
), grad
def csrsum(gidxs, weights):
@tf.custom_gradient
def _lambda(*weights):
return csrsum_real(gidxs, weights)
nrows, ncols, C_indptr, C_indices, C_eids, C_weights = _lambda(*weights)
num_vtypes = gidxs[0].number_of_ntypes()
gidxC = create_unitgraph_from_csr(
num_vtypes,
nrows.numpy(),
ncols.numpy(),
C_indptr,
C_indices,
C_eids,
["coo", "csr", "csc"],
)
return gidxC, C_weights
def csrmask_real(gidxA, A_weights, gidxB):
B_weights = _csrmask(gidxA, A_weights, gidxB)
def grad(dB_weights):
return _csrmask(gidxB, dB_weights, gidxA)
return B_weights, grad
def csrmask(gidxA, A_weights, gidxB):
@tf.custom_gradient
def _lambda(A_weights):
return csrmask_real(gidxA, A_weights, gidxB)
return _lambda(A_weights)