chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,6 @@
|
||||
import os
|
||||
|
||||
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"
|
||||
|
||||
from .sparse import *
|
||||
from .tensor import *
|
||||
@@ -0,0 +1,461 @@
|
||||
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)
|
||||
@@ -0,0 +1 @@
|
||||
"""Sparse optimizer is not supported for tensorflow"""
|
||||
@@ -0,0 +1,619 @@
|
||||
"""Tensorflow backend implementation"""
|
||||
from __future__ import absolute_import
|
||||
|
||||
import builtins
|
||||
import numbers
|
||||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
|
||||
from ... import ndarray as nd
|
||||
from ...function.base import TargetCode
|
||||
from ...utils import version
|
||||
|
||||
if version.parse(tf.__version__) < version.parse("2.3.0"):
|
||||
raise RuntimeError(
|
||||
"DGL requires TensorFlow>=2.3.0 for the official DLPack support."
|
||||
)
|
||||
|
||||
|
||||
def zerocopy_to_dlpack(data):
|
||||
return tf.experimental.dlpack.to_dlpack(data)
|
||||
|
||||
|
||||
def zerocopy_from_dlpack(dlpack_tensor):
|
||||
# TODO(Jinjing): Tensorflow requires memory to be 64-bytes aligned. We check the
|
||||
# alignment and make a copy if needed. The functionality is better in TF's main repo.
|
||||
aligned = nd.from_dlpack(dlpack_tensor).to_dlpack(64)
|
||||
return tf.experimental.dlpack.from_dlpack(aligned)
|
||||
|
||||
|
||||
def data_type_dict():
|
||||
return {
|
||||
"bfloat16": tf.bfloat16,
|
||||
"float16": tf.float16,
|
||||
"float32": tf.float32,
|
||||
"float64": tf.float64,
|
||||
"uint8": tf.uint8,
|
||||
"int8": tf.int8,
|
||||
"int16": tf.int16,
|
||||
"int32": tf.int32,
|
||||
"int64": tf.int64,
|
||||
"bool": tf.bool,
|
||||
}
|
||||
|
||||
|
||||
def cpu():
|
||||
return "/cpu:0"
|
||||
|
||||
|
||||
def tensor(data, dtype=None):
|
||||
if isinstance(data, tf.Tensor):
|
||||
if dtype is None or data.dtype == dtype:
|
||||
return data
|
||||
else:
|
||||
return tf.cast(data, dtype=dtype)
|
||||
else:
|
||||
if isinstance(data, numbers.Number):
|
||||
data = [data]
|
||||
return tf.convert_to_tensor(data, dtype=dtype)
|
||||
|
||||
|
||||
def initialize_context():
|
||||
tf.zeros(1)
|
||||
|
||||
|
||||
def as_scalar(data):
|
||||
data = data.numpy()
|
||||
return data if np.isscalar(data) else data.item()
|
||||
|
||||
|
||||
def get_preferred_sparse_format():
|
||||
"""Get the preferred sparse matrix format supported by the backend.
|
||||
|
||||
Different backends have their preferred backend. This info is useful when
|
||||
constructing a sparse matrix.
|
||||
"""
|
||||
return "coo"
|
||||
|
||||
|
||||
def sparse_matrix(data, index, shape, force_format=False):
|
||||
fmt = index[0]
|
||||
if fmt != "coo":
|
||||
raise TypeError(
|
||||
"Tensorflow backend only supports COO format. But got %s." % fmt
|
||||
)
|
||||
# tf.SparseTensor only supports int64 indexing,
|
||||
# therefore manually casting to int64 when input in int32
|
||||
spmat = tf.SparseTensor(
|
||||
indices=tf.cast(tf.transpose(index[1], (1, 0)), tf.int64),
|
||||
values=data,
|
||||
dense_shape=shape,
|
||||
)
|
||||
return spmat, None
|
||||
|
||||
|
||||
def sparse_matrix_indices(spmat):
|
||||
return ("coo", spmat.indices)
|
||||
|
||||
|
||||
def is_tensor(obj):
|
||||
return isinstance(obj, tf.Tensor)
|
||||
|
||||
|
||||
def shape(input):
|
||||
return input.shape
|
||||
|
||||
|
||||
def dtype(input):
|
||||
return input.dtype
|
||||
|
||||
|
||||
def ndim(input):
|
||||
return input.ndim
|
||||
|
||||
|
||||
def context(input):
|
||||
spec = tf.DeviceSpec.from_string(input.device)
|
||||
return "/{}:{}".format(spec.device_type.lower(), spec.device_index)
|
||||
|
||||
|
||||
def device_type(ctx):
|
||||
return tf.DeviceSpec.from_string(ctx).device_type.lower()
|
||||
|
||||
|
||||
def device_id(ctx):
|
||||
return tf.DeviceSpec.from_string(ctx).device_index
|
||||
|
||||
|
||||
def to_backend_ctx(dglctx):
|
||||
dev_type = dglctx.device_type
|
||||
if dev_type == 1:
|
||||
return "/cpu:0"
|
||||
elif dev_type == 2:
|
||||
return "/gpu:%d" % (dglctx.device_id)
|
||||
else:
|
||||
raise ValueError("Unsupported DGL device context:", dglctx)
|
||||
|
||||
|
||||
def astype(input, ty):
|
||||
with tf.device(input.device):
|
||||
return tf.cast(input, dtype=ty)
|
||||
|
||||
|
||||
def asnumpy(input):
|
||||
if isinstance(input, tf.SparseTensor):
|
||||
# tf.sparse.to_dense assume sorted indices, need to turn off validate_indices in our cases
|
||||
return tf.sparse.to_dense(input, validate_indices=False).numpy()
|
||||
else:
|
||||
return input.numpy()
|
||||
|
||||
|
||||
def copy_to(input, ctx, **kwargs):
|
||||
with tf.device(ctx):
|
||||
new_tensor = tf.identity(input)
|
||||
return new_tensor
|
||||
|
||||
|
||||
def is_pinned(input):
|
||||
return False # not sure how to do this
|
||||
|
||||
|
||||
def sum(input, dim, keepdims=False):
|
||||
if input.dtype == tf.bool:
|
||||
input = tf.cast(input, tf.int32)
|
||||
return tf.reduce_sum(input, axis=dim, keepdims=keepdims)
|
||||
|
||||
|
||||
def floor_div(in1, in2):
|
||||
return astype(in1 / in2, dtype(in1))
|
||||
|
||||
|
||||
def reduce_sum(input):
|
||||
if input.dtype == tf.bool:
|
||||
input = tf.cast(input, tf.int32)
|
||||
return tf.reduce_sum(input)
|
||||
|
||||
|
||||
def cumsum(input, dim):
|
||||
if input.dtype == tf.bool:
|
||||
input = tf.cast(input, tf.int32)
|
||||
return tf.cumsum(input, axis=dim)
|
||||
|
||||
|
||||
def mean(input, dim):
|
||||
return tf.reduce_mean(input, axis=dim)
|
||||
|
||||
|
||||
def reduce_mean(input):
|
||||
return tf.reduce_mean(input)
|
||||
|
||||
|
||||
def max(input, dim):
|
||||
return tf.reduce_max(input, axis=dim)
|
||||
|
||||
|
||||
def reduce_max(input):
|
||||
return tf.reduce_max(input)
|
||||
|
||||
|
||||
def min(input, dim):
|
||||
return tf.reduce_min(input, axis=dim)
|
||||
|
||||
|
||||
def reduce_min(input):
|
||||
return tf.reduce_min(input)
|
||||
|
||||
|
||||
def argsort(input, dim, descending):
|
||||
if descending:
|
||||
return tf.cast(
|
||||
tf.argsort(input, axis=dim, direction="DESCENDING"), dtype=tf.int64
|
||||
)
|
||||
else:
|
||||
return tf.cast(
|
||||
tf.argsort(input, axis=dim, direction="ASCENDING"), dtype=tf.int64
|
||||
)
|
||||
|
||||
|
||||
def topk(input, k, dim, descending=True):
|
||||
if not descending:
|
||||
input = -input
|
||||
shape = np.arange(input.ndim)
|
||||
shape[dim], shape[-1] = shape[-1], shape[dim]
|
||||
out1 = tf.transpose(input, perm=shape)
|
||||
out2 = tf.math.top_k(out1, k=k, sorted=True)
|
||||
out = tf.transpose(out2[0], shape)
|
||||
if not descending:
|
||||
out = -out
|
||||
return out
|
||||
|
||||
|
||||
def argtopk(input, k, dim, descending=True):
|
||||
if not descending:
|
||||
input = -input
|
||||
shape = np.arange(input.ndim)
|
||||
shape[dim], shape[-1] = shape[-1], shape[dim]
|
||||
out1 = tf.transpose(input, perm=shape)
|
||||
out2 = tf.math.top_k(out1, k=k, sorted=True)
|
||||
out = tf.transpose(out2[1], shape)
|
||||
if not descending:
|
||||
out = -out
|
||||
return out
|
||||
|
||||
|
||||
def exp(input):
|
||||
return tf.exp(input)
|
||||
|
||||
|
||||
def inverse(input):
|
||||
return tf.linalg.inv(input)
|
||||
|
||||
|
||||
def sqrt(input):
|
||||
return tf.sqrt(input)
|
||||
|
||||
|
||||
def softmax(input, dim=-1):
|
||||
return tf.math.softmax(input, axis=dim)
|
||||
|
||||
|
||||
def cat(seq, dim):
|
||||
return tf.concat(seq, axis=dim)
|
||||
|
||||
|
||||
def stack(seq, dim):
|
||||
return tf.stack(seq, axis=dim)
|
||||
|
||||
|
||||
def split(input, sizes_or_sections, dim):
|
||||
return [
|
||||
copy_to(_, input.device)
|
||||
for _ in tf.split(input, sizes_or_sections, axis=dim)
|
||||
]
|
||||
|
||||
|
||||
def repeat(input, repeats, dim):
|
||||
return tf.repeat(input, repeats, dim)
|
||||
|
||||
|
||||
def gather_row(data, row_index):
|
||||
return tf.gather(data, row_index)
|
||||
|
||||
|
||||
def slice_axis(data, axis, begin, end):
|
||||
# assert axis == 0
|
||||
# tf doesn't behave well with negative
|
||||
s = [slice(None) for i in range(data.ndim)]
|
||||
if end == 0:
|
||||
end = data.shape[axis]
|
||||
s[axis] = slice(begin, end, None)
|
||||
return data[tuple(s)]
|
||||
|
||||
|
||||
def take(data, indices, dim):
|
||||
return tf.gather_nd(data, indices, dim)
|
||||
|
||||
|
||||
def narrow_row(x, start, stop):
|
||||
return x[start:stop]
|
||||
|
||||
|
||||
def scatter_row(data, row_index, value):
|
||||
row_index = tf.expand_dims(row_index, 1)
|
||||
# XXX(minjie): Normally, the copy_to here is unnecessary. However, TF has this
|
||||
# notorious legacy issue that int32 type data is always on CPU, which will
|
||||
# crash the program since DGL requires feature data to be on the same device
|
||||
# as graph structure.
|
||||
return copy_to(
|
||||
tf.tensor_scatter_nd_update(data, row_index, value), data.device
|
||||
)
|
||||
|
||||
|
||||
def index_add_inplace(data, row_idx, value):
|
||||
raise NotImplementedError("Tensorflow doesn't support inplace index_add")
|
||||
|
||||
|
||||
def scatter_row_inplace(data, row_index, value):
|
||||
raise NotImplementedError("Tensorflow doesn't support inplace update")
|
||||
|
||||
|
||||
def squeeze(input, dim):
|
||||
return tf.squeeze(input, axis=dim)
|
||||
|
||||
|
||||
def unsqueeze(input, dim):
|
||||
return tf.expand_dims(input, axis=dim)
|
||||
|
||||
|
||||
def reshape(input, shape):
|
||||
return tf.reshape(input, shape)
|
||||
|
||||
|
||||
def swapaxes(input, axis1, axis2):
|
||||
ndim = input.ndim
|
||||
t = list(range(ndim))
|
||||
t[axis1], t[axis2] = axis2 % ndim, axis1 % ndim
|
||||
return tf.transpose(input, perm=t)
|
||||
|
||||
|
||||
def empty(shape, dtype, ctx):
|
||||
# tf doesn't have tf.empty(), use zeros() as a workaround
|
||||
return zeros(shape, dtype, ctx)
|
||||
|
||||
|
||||
def zeros(shape, dtype, ctx):
|
||||
with tf.device(ctx):
|
||||
t = tf.zeros(shape, dtype=dtype)
|
||||
return t
|
||||
|
||||
|
||||
def zeros_like(input):
|
||||
return tf.zeros_like(input)
|
||||
|
||||
|
||||
def ones(shape, dtype, ctx):
|
||||
with tf.device(ctx):
|
||||
t = tf.ones(shape, dtype=dtype)
|
||||
return t
|
||||
|
||||
|
||||
def uniform(shape, dtype, ctx, low, high):
|
||||
with tf.device(ctx):
|
||||
t = tf.random.uniform(shape, dtype=dtype, minval=low, maxval=high)
|
||||
return t
|
||||
|
||||
|
||||
def randint(shape, dtype, ctx, low, high):
|
||||
with tf.device(ctx):
|
||||
t = tf.random.uniform(shape, dtype=dtype, minval=low, maxval=high)
|
||||
return t
|
||||
|
||||
|
||||
def pad_packed_tensor(input, lengths, value, l_min=None):
|
||||
old_shape = input.shape
|
||||
if isinstance(lengths, tf.Tensor):
|
||||
max_len = as_scalar(tf.reduce_max(lengths))
|
||||
else:
|
||||
max_len = builtins.max(lengths)
|
||||
|
||||
if l_min is not None:
|
||||
max_len = builtins.max(max_len, l_min)
|
||||
|
||||
batch_size = len(lengths)
|
||||
ndim = input.ndim
|
||||
tensor_list = []
|
||||
cum_row = 0
|
||||
pad_nparray = np.zeros((ndim, 2), dtype=np.int32)
|
||||
for l in lengths:
|
||||
t = input[cum_row : cum_row + l]
|
||||
pad_nparray[0, 1] = max_len - l
|
||||
t = tf.pad(
|
||||
t, tf.constant(pad_nparray), mode="CONSTANT", constant_values=value
|
||||
)
|
||||
tensor_list.append(t)
|
||||
cum_row += l
|
||||
return tf.stack(tensor_list, axis=0)
|
||||
|
||||
|
||||
def pack_padded_tensor(input, lengths):
|
||||
out_list = []
|
||||
for i, l in enumerate(lengths):
|
||||
t = input[i]
|
||||
out = t[:l]
|
||||
out_list.append(out)
|
||||
return tf.concat(out_list, axis=0)
|
||||
|
||||
|
||||
def boolean_mask(input, mask):
|
||||
return tf.boolean_mask(input, mask)
|
||||
|
||||
|
||||
def equal(x, y):
|
||||
return x == y
|
||||
|
||||
|
||||
def allclose(x, y, rtol=1e-4, atol=1e-4):
|
||||
return np.allclose(
|
||||
tf.convert_to_tensor(x).numpy(),
|
||||
tf.convert_to_tensor(y).numpy(),
|
||||
rtol=rtol,
|
||||
atol=atol,
|
||||
)
|
||||
|
||||
|
||||
def logical_not(input):
|
||||
return ~input
|
||||
|
||||
|
||||
def logical_and(input1, input2):
|
||||
return tf.math.logical_and(input1, input2)
|
||||
|
||||
|
||||
def clone(input):
|
||||
# TF tensor is always immutable so returning the input is safe.
|
||||
return input
|
||||
|
||||
|
||||
def clamp(data, min_val, max_val):
|
||||
return tf.clip_by_value(data, min_val, max_val)
|
||||
|
||||
|
||||
def replace_inf_with_zero(x):
|
||||
return tf.where(tf.abs(x) == np.inf, 0, x)
|
||||
|
||||
|
||||
def count_nonzero(input):
|
||||
return int(tf.math.count_nonzero(input))
|
||||
|
||||
|
||||
def unique(input, return_inverse=False, return_counts=False):
|
||||
if return_inverse and return_counts:
|
||||
return tf.unique_with_counts(input)
|
||||
elif return_counts:
|
||||
result = tf.unique_with_counts(input)
|
||||
return result.y, result.count
|
||||
elif return_inverse:
|
||||
return tf.unique(input)
|
||||
else:
|
||||
return tf.unique(input).y
|
||||
|
||||
|
||||
def full_1d(length, fill_value, dtype, ctx):
|
||||
with tf.device(ctx):
|
||||
t = tf.fill([length], value=fill_value)
|
||||
t = tf.cast(t, dtype=dtype)
|
||||
return t
|
||||
|
||||
|
||||
def nonzero_1d(input):
|
||||
nonzero_bool = tf.cast(input, tf.bool)
|
||||
return tf.reshape(tf.where(nonzero_bool), (-1,))
|
||||
|
||||
|
||||
def sort_1d(input):
|
||||
return tf.sort(input), tf.cast(tf.argsort(input), dtype=tf.int64)
|
||||
|
||||
|
||||
def arange(start, stop, dtype=tf.int64, ctx=None):
|
||||
if not ctx:
|
||||
ctx = "/cpu:0"
|
||||
with tf.device(ctx):
|
||||
t = tf.range(start, stop, dtype=dtype)
|
||||
return t
|
||||
|
||||
|
||||
def rand_shuffle(arr):
|
||||
return tf.random.shuffle(arr)
|
||||
|
||||
|
||||
def zerocopy_to_numpy(input):
|
||||
return np.asarray(memoryview(input))
|
||||
|
||||
|
||||
def zerocopy_from_numpy(np_array):
|
||||
# NOTE: not zerocopy
|
||||
# This assumes tensor should be on cpu
|
||||
with tf.device("/cpu:0"):
|
||||
t = tf.convert_to_tensor(np_array)
|
||||
return t
|
||||
|
||||
|
||||
def zerocopy_to_dgl_ndarray(data):
|
||||
if device_type(data.device) == "gpu" and data.dtype in (tf.int32, tf.int64):
|
||||
# NOTE: TF doesn't keep signed tensors on GPU due to legacy issues with
|
||||
# shape inference. Convert it to unsigned and cast it back afterwards.
|
||||
if data.dtype == tf.int32:
|
||||
data = tf.cast(data, tf.uint32)
|
||||
elif data.dtype == tf.int64:
|
||||
data = tf.cast(data, tf.uint64)
|
||||
return nd.cast_to_signed(nd.from_dlpack(zerocopy_to_dlpack(data)))
|
||||
else:
|
||||
return nd.from_dlpack(zerocopy_to_dlpack(data))
|
||||
|
||||
|
||||
def zerocopy_to_dgl_ndarray_for_write(input):
|
||||
return zerocopy_to_dgl_ndarray(input)
|
||||
|
||||
|
||||
def zerocopy_from_dgl_ndarray(input):
|
||||
return zerocopy_from_dlpack(input.to_dlpack())
|
||||
|
||||
|
||||
def sync():
|
||||
context = context().context()
|
||||
context.async_wait()
|
||||
|
||||
|
||||
class GradContext:
|
||||
def __init__(self):
|
||||
self.tensor_for_grad = []
|
||||
self.grad_list = []
|
||||
self.tape = None
|
||||
|
||||
def set_tape(self, tape):
|
||||
self.tape = tape
|
||||
|
||||
def add_tensor(self, x):
|
||||
idx_pop = []
|
||||
for idx, ele in enumerate(self.tensor_for_grad):
|
||||
if ele._id == x._id:
|
||||
idx_pop.append(idx)
|
||||
if len(idx_pop) > 0:
|
||||
self.tensor_for_grad.pop(idx_pop[0])
|
||||
if self.tape is not None:
|
||||
self.tape.watch(x)
|
||||
self.tensor_for_grad.append(x)
|
||||
|
||||
def backward(self, x, head_gradient=None):
|
||||
if head_gradient is not None:
|
||||
x = x * head_gradient
|
||||
self.grad_list = self.tape.gradient(x, self.tensor_for_grad)
|
||||
|
||||
def is_no_grad(self, x):
|
||||
idx_pop = []
|
||||
for idx, ele in enumerate(self.tensor_for_grad):
|
||||
if ele._id == x._id:
|
||||
idx_pop.append(idx)
|
||||
if len(idx_pop) == 0:
|
||||
return True
|
||||
else:
|
||||
return self.grad_list[idx_pop[0]] is None
|
||||
|
||||
def grad(self, x):
|
||||
idx_pop = []
|
||||
for idx, ele in enumerate(self.tensor_for_grad):
|
||||
if ele._id == x._id:
|
||||
idx_pop.append(idx)
|
||||
assert len(idx_pop) == 1
|
||||
t = self.grad_list[idx_pop[0]]
|
||||
return tf.convert_to_tensor(t)
|
||||
|
||||
|
||||
cgrad = GradContext()
|
||||
|
||||
|
||||
def get_cgrad():
|
||||
return cgrad
|
||||
|
||||
|
||||
class record_grad:
|
||||
def __init__(self):
|
||||
self.tape = tf.GradientTape()
|
||||
|
||||
def __enter__(self):
|
||||
cgrad.set_tape(self.tape)
|
||||
self.tape.__enter__()
|
||||
for x in cgrad.tensor_for_grad:
|
||||
self.tape.watch(x)
|
||||
|
||||
def __exit__(self, exc_type, exc_value, exc_traceback):
|
||||
# pass
|
||||
self.tape.__exit__(exc_type, exc_value, exc_traceback)
|
||||
cgrad.tape = None
|
||||
|
||||
|
||||
def attach_grad(x):
|
||||
cgrad.add_tensor(x)
|
||||
return x
|
||||
|
||||
|
||||
def backward(x, head_gradient=None):
|
||||
cgrad.backward(x, head_gradient)
|
||||
|
||||
|
||||
def grad(x):
|
||||
return cgrad.grad(x)
|
||||
|
||||
|
||||
def is_no_grad(x):
|
||||
return cgrad.is_no_grad(x)
|
||||
|
||||
|
||||
def is_recording():
|
||||
raise NotImplementedError("Tensorflow doesn't support is_recording")
|
||||
|
||||
|
||||
no_grad = None
|
||||
|
||||
initialize_context()
|
||||
Reference in New Issue
Block a user