You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

128 lines
4.0 KiB

import torch
from torch import Tensor
aten = torch.ops.aten
import inspect
import warnings
from typing import Dict, List, Optional, Set
from torch.types import Number
decomposition_table: Dict[str, torch.jit.ScriptFunction] = {}
function_name_set: Set[str] = set()
def check_decomposition_has_type_annotations(f):
inspect_empty = inspect._empty # type: ignore[attr-defined]
sig = inspect.signature(f)
for param in sig.parameters.values():
assert (
param.annotation != inspect_empty
), f"No signature on param {param.name} for function {f.name}"
assert (
sig.return_annotation != inspect_empty
), f"No return annotation for function {f.name}"
def signatures_match(decomposition_sig, torch_op_sig):
decomp_params = decomposition_sig.parameters
op_params = torch_op_sig.parameters
if len(decomp_params) != len(op_params):
return False
for decomp_param, op_param in zip(decomp_params.values(), op_params.values()):
# can't check full equality yet because not all fields are correcly deduced
# in the torch_op_sig - like default value
# can't check 'kind' bc
# kwarg-only values with defaults not yet supported in TS
inspect_empty = inspect._empty # type: ignore[attr-defined]
for field in ["name", "annotation"]:
if field == "name" and decomp_param.name == "self":
warnings.warn("PyTorch uses 'input' instead of 'self' on public api")
if getattr(decomp_param, field) != getattr(op_param, field):
return False
decomp_default = decomp_param.default
op_default = op_param.default
# default value not always correctly inferred as being present on torch schema,
# but if specified on both they should be equal
if decomp_default != inspect_empty and op_default != inspect_empty:
if decomp_default != op_default:
return False
return decomposition_sig.return_annotation == torch_op_sig.return_annotation
def register_decomposition(aten_op, registry=None):
def decomposition_decorator(f):
nonlocal registry
if registry is None:
registry = decomposition_table
assert isinstance(aten_op, torch._ops.OpOverload)
# Need unique name for jit function serialization
assert (
f.__name__ not in function_name_set
), f"Duplicated function name {f.__name__}"
function_name_set.add(f.__name__)
scripted_func = torch.jit.script(f)
torch._C._jit_pass_inline(scripted_func.graph)
for _ in range(2):
torch._C._jit_pass_peephole(scripted_func.graph)
torch._C._jit_pass_constant_propagation(scripted_func.graph)
registry[str(aten_op._schema)] = scripted_func
return f
return decomposition_decorator
# TODO: replace torch.sigmoid -> aten.sigmoid
@register_decomposition(aten.var.correction)
def var_decomposition(
input: Tensor,
dim: Optional[List[int]] = None,
correction: Optional[Number] = None,
keepdim: bool = False,
) -> Tensor:
if dim is None:
dim_i: List[int] = []
dim = dim_i
if isinstance(dim, (tuple, list)) and len(dim) == 0:
n = input.numel()
else:
n = 1
for dim_i in dim: # type: ignore[assignment]
n *= input.shape[dim_i] # type: ignore[call-overload]
mean = aten.mean(input, dim, True)
sub = input - mean
sq = sub * sub
sum = aten.sum(sq, dim, keepdim)
if correction is None:
denom = float(n - 1)
else:
if isinstance(correction, int):
denom = float(n - correction)
elif isinstance(correction, float):
denom = float(n) - correction
else:
raise RuntimeError("correction must be int or float")
return sum / max(0, denom)
@register_decomposition(aten.var.default)
def var(input: Tensor, unbiased: bool = True) -> Tensor:
return var_decomposition(input, correction=(1 if unbiased else 0))