megengine.traced_module.expr 源代码

# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

import builtins
import collections
import copy
import inspect
import re
from typing import Callable, Dict, List

from ..core._imperative_rt import OpDef
from ..core._imperative_rt.core2 import Tensor as RawTensor
from ..core._imperative_rt.core2 import apply, set_module_tracing, unset_module_tracing
from ..core.ops.builtin import FakeQuant
from ..core.ops.special import Const
from ..module import Module
from ..tensor import Parameter, Tensor
from .module_tracer import active_module_tracer, module_tracer
from .node import ModuleNode, Node, NodeMixin, TensorNode
from .pytree import ArgsIndex, TreeDef, _is_const_leaf, _is_leaf, tree_flatten
from .serialization import get_opdef_state, load_opdef_from_state


def rstrip(s: str, __chars: str):
    __chars = re.escape(__chars)
    s = re.sub(r"^(?P<left>.*?)(?:%s)+$" % __chars, "\g<left>", s)
    return s


class Expr:
    r"""``Expr`` represents the operations (i.e. ``CallMethod``, ``CallFunction``, ``Apply``, 
    ``GetAttr``, ``Input``, ``Constant``) on ``Node``.
    """

    inputs = None  # type: List[Node]
    r"""The input Nodes of this Expr."""
    outputs = None  # type: List[Node]
    r"""The output Nodes of this Expr."""
    const_val = None  # type: List[Any]
    r"""The non-tensor object in the input of the operation."""
    arg_def = None  # type: TreeDef
    r"""The :class:`TreeDef` used to reconstruct the input of the operation."""
    out_def = None  # type: TreeDef
    r"""The :class:`TreeDef` used to reconstruct the output of the operation."""
    _top_graph = None  # type: weakref.ReferenceType
    __total_id = 0

    def __init__(self) -> None:
        self._id = Expr.__total_id
        Expr.__total_id += 1
        self._disable_remove = False

    def enable_remove(self):
        self._disable_remove = False

    def disable_remove(self):
        self._disable_remove = True

    def add_inputs(self, vals):
        if not isinstance(vals, collections.abc.Sequence):
            vals = (vals,)
        for val in vals:
            node = NodeMixin.get(val, None)
            if isinstance(node, (TensorNode, ModuleNode)):
                self.inputs.append(node)
                node.users.append(self)
            else:
                assert node is None
                assert _is_leaf(val) and _is_const_leaf(val)
                idx = len(self.inputs) + len(self.const_val)
                self.const_val.append((idx, val))

    def add_outputs(self, outputs):
        self.outputs = []
        if outputs is not None:
            if not isinstance(outputs, collections.Sequence):
                outputs = (outputs,)

            name = None
            orig_name = None
            if isinstance(self, CallMethod):
                name = self.inputs[0]._name
                orig_name = self.inputs[0]._orig_name
                assert isinstance(name, str), "The name of ({}) must be a str".format(
                    self.inputs[0]
                )
                assert isinstance(
                    orig_name, str
                ), "The orig_name of ({}) must be a str".format(self.inputs[0])
                name = rstrip(name, "_out")
                if self.method == "__call__":
                    name += "_out"
                    orig_name += "_out"
                else:
                    strip_method = self.method.strip("_")
                    name = "%s_out" % strip_method
                    orig_name = name
            elif isinstance(self, CallFunction):
                name = self.func.__name__ + "_out"
            elif isinstance(self, Apply):
                name = str(self.opdef).lower() + "_out"

            for i in outputs:
                assert isinstance(i, RawTensor), "The output must be a Tensor"
                o_name = (
                    active_module_tracer().current_scope()._create_unique_name(name)
                )
                self.outputs.append(
                    NodeMixin.get_wrapped_type(i)(
                        expr=self,
                        name=o_name,
                        orig_name=orig_name if orig_name else o_name,
                    )
                )

            for i, node in zip(outputs, self.outputs,):
                NodeMixin.wrap_safe(i, node)

    def unflatten_args(self, inputs):
        if self.arg_def is not None:
            inputs = list(inputs)
            for idx, val in self.const_val:
                inputs.insert(idx, val)
            args, kwargs = self.arg_def.unflatten(inputs)
            return args, kwargs
        else:
            return inputs, {}

    def replace_inputs(self, repl_dict: Dict[Node, Node]):
        r"""Replace the input Nodes of this Expr.
        
        Args:
            repl_dict: the map {old_Node: new_Node} that specifies how to replace the input Nodes.
        """
        while repl_dict:
            node, repl_node = repl_dict.popitem()
            assert type(node) == type(repl_node)
            assert node in self.inputs, "({}) is not in the ({})".format(node, self)
            assert (
                repl_node.top_graph == node.top_graph
            ), "({}) and ({}) are not in the same graph".format(node, repl_node)
            graph = self.top_graph
            repl_expr_idx = graph._exprs.index(repl_node.expr)
            self_idx = graph._exprs.index(self)
            assert (
                repl_expr_idx < self_idx
            ), "({}) must be generated before ({})".format(repl_node, self)
            idx = self.inputs.index(node)
            self.inputs[idx] = repl_node
            user_idx = node.users.index(self)
            assert user_idx >= 0
            node.users.pop(user_idx)
            repl_node.users.append(self)

    @property
    def kwargs(self):
        r"""Get the the keyword arguments of the operation corresponding to this Expr."""
        _, kwargs = self.unflatten_args(self.inputs)
        return kwargs

    @property
    def args(self):
        r"""Get the the positional arguments of the operation corresponding to this Expr."""
        args, _ = self.unflatten_args(self.inputs)
        return args

    @property
    def top_graph(self):
        r"""Get the parent graph of this Expr."""
        if self._top_graph:
            return self._top_graph()
        return None

    def __getstate__(self):
        state = self.__dict__.copy()
        if "_top_graph" in state:
            state.pop("_top_graph")
        return state

    @classmethod
    def _get_next_id(cls):
        return cls.__total_id

    @classmethod
    def _set_next_id(cls, id: int = 0):
        assert isinstance(id, int)
        cls.__total_id = id


# expr: None (i.e. fake expression which is used to mark input)
[文档]class Input(Expr): r"""A fake Expr which is used to mark the input of graph.""" name = None def __init__(self, name=None, type=None, orig_name=None): super().__init__() self.inputs = [] node_cls = type if type else Node if orig_name is None: orig_name = name self.outputs = [ node_cls(self, name=name, orig_name=orig_name), ] self.name = name
[文档] @classmethod def make(cls, *args, **kwargs): expr = cls(*args, **kwargs) oup_node = expr.outputs[0] name = ( active_module_tracer().current_scope()._create_unique_name(oup_node._name) ) oup_node._name = name active_module_tracer().current_scope()._add_input(oup_node) return expr.outputs[0]
def __repr__(self): return "%{}:\t{} = Input()".format(self._id, self.outputs[0])
# expr: outputs = getattr(inputs[0], self.name)
[文档]class GetAttr(Expr): r"""``Getattr`` represents the fetch of an attribute from the ``Module`` hierarchy.""" name = None r"""name: the qualified name of the attribute to be retrieved.""" def __init__(self, module, name, type=None, orig_name=None): super().__init__() assert isinstance(module, ModuleNode) self.inputs = [ module, ] module.users.append(self) self.name = name node_cls = type if type else Node self.outputs = [ node_cls(self, name=name, orig_name=orig_name), ]
[文档] @classmethod def make(cls, *args, **kwargs): expr = cls(*args, **kwargs) module = expr.inputs[0] oup_name = expr.name while module._name != "self": oup_name = module._name + "_" + oup_name module = module.expr.inputs[0] oup_name = active_module_tracer().current_scope()._create_unique_name(oup_name) expr.outputs[0]._name = oup_name active_module_tracer().current_scope()._insert(expr) return expr.outputs[0]
[文档] def interpret(self, *inputs): return (getattr(inputs[0], self.name),)
def __repr__(self): out_type = "Tensor" if isinstance(self.outputs[0], ModuleNode): out_type = self.outputs[0].module_type.__name__ return '%{}:\t{} = getattr({}, "{}") -> ({})'.format( self._id, self.outputs[0], self.inputs[0], self.name, out_type )
# expr: outputs = inputs[0].__call__(*inputs[1:])
[文档]class CallMethod(Expr): r"""``CallMethod`` represents a call to the ``__call__`` method of ``Module`` or a method of ``Tensor``. Args: node: the Node to be called. method: the method name. Default: "__call__" """ def __init__(self, node, method="__call__"): super().__init__() if isinstance(node, type): assert issubclass(node, Tensor) cls = Parameter if issubclass(node, Parameter) else Tensor self.inputs = [] self.const_val = [(0, cls)] else: assert isinstance(node, (TensorNode, ModuleNode)) node.users.append(self) self.inputs = [ node, ] self.const_val = [] self.method = method
[文档] @classmethod def make(cls, *args, **kwargs): expr = cls(*args, **kwargs) active_module_tracer().current_scope()._insert(expr) return expr
@property def graph(self): if isinstance(self.inputs[0], ModuleNode): m_node = self.inputs[0] if ( hasattr(m_node.owner, "argdef_graph_map") and m_node.owner.argdef_graph_map ): assert self.arg_def in m_node.owner.argdef_graph_map return m_node.owner.argdef_graph_map[self.arg_def] return None
[文档] def interpret(self, *inputs): args, kwargs = self.unflatten_args(inputs) obj = args[0] meth = getattr(obj, self.method) if inspect.ismethod(meth): args = args[1:] outputs = getattr(obj, self.method)(*args, **kwargs) if self.method == "__setitem__": outputs = obj if outputs is None: return outputs outputs, _ = tree_flatten(outputs, is_leaf=lambda x: isinstance(x, RawTensor)) return outputs
def __repr__(self): args = ", ".join(str(i) for i in self.args[1:]) kwargs = ", ".join("{}={}".format(k, v) for k, v in self.kwargs.items()) outputs = self.outputs if self.out_def: outputs = self.out_def.unflatten(outputs) method = ".%s" % self.method if method == ".__call__": method = "" return "%{}:\t{}{}{}({})".format( self._id, str(outputs) + " = " if outputs else "", self.args[0], method, ", ".join([args, kwargs]), )
# expr: outputs = apply(self.opdef, *inputs)
[文档]class Apply(Expr): r"""``Apply`` represents a call to :func:`apply`. Args: opdef: the applied :class:`OpDef`. """ opdef = None def __init__(self, opdef): super().__init__() assert isinstance(opdef, OpDef) self.opdef = opdef self.inputs = []
[文档] @classmethod def make(cls, *args, **kwargs): expr = cls(*args, **kwargs) active_module_tracer().current_scope()._insert(expr) return expr
[文档] def interpret(self, *inputs): return apply(self.opdef, *inputs)
def __repr__(self): return "%{}:\t{} = {}({})".format( self._id, ", ".join(str(i) for i in self.outputs), self.opdef, ", ".join(str(i) for i in self.inputs), ) def __getstate__(self): state = super().__getstate__() state["opdef"] = get_opdef_state(state["opdef"]) return state def __setstate__(self, state): state["opdef"] = load_opdef_from_state(state["opdef"]) for k, v in state.items(): setattr(self, k, v)
[文档] @classmethod def apply_module_trace_hook(cls, opdef, *inputs): for i in inputs: node = NodeMixin.get(i, None) if node is None: # capture as constant NodeMixin.wrap_safe(i, Constant.make(i)) if isinstance(opdef, FakeQuant): inp_nodes = [NodeMixin.get(inputs[0])] for i in inputs[1:]: node = Constant.make(i) inp_nodes.append(node) apply_node = cls.make(opdef) for n in inp_nodes: n.users.append(apply_node) apply_node.inputs = inp_nodes else: apply_node = cls.make(opdef) apply_node.add_inputs(inputs) assert not apply_node.const_val unset_module_tracing() outputs = apply(opdef, *inputs) set_module_tracing() apply_node.add_outputs(outputs) for n, v in zip(apply_node.outputs, outputs): NodeMixin.wrap_safe(v, n) return list(outputs)
[文档]class CallFunction(Expr): r"""``CallFunction`` represents a call to a built-in function. Args: func: a built-in function. """ def __init__(self, func): super().__init__() assert isinstance(func, Callable) self.func = func self.const_val = [] self.inputs = []
[文档] @classmethod def make(cls, *args, **kwargs): expr = cls(*args, **kwargs) active_module_tracer().current_scope()._insert(expr) return expr
[文档] def interpret(self, *inputs): args, kwargs = self.unflatten_args(inputs) outputs = self.func(*args, **kwargs) if outputs is None: return outputs outputs, _ = tree_flatten(outputs, is_leaf=lambda x: isinstance(x, RawTensor)) return outputs
def __repr__(self): args = ", ".join(str(i) for i in self.args) kwargs = ", ".join("{}={}".format(k, v) for k, v in self.kwargs.items()) outputs = self.outputs if self.out_def: outputs = self.out_def.unflatten(outputs) return "%{}:\t{}{}({})".format( self._id, str(outputs) + " = " if outputs else "", self.func.__module__.rsplit(".")[-1] + "." + self.func.__name__, ", ".join([args, kwargs]), )
# expr outputs = self.value
[文档]class Constant(Expr): r"""``Constant`` represents a ``Tensor`` or "Module" which is not the attribute of a Module. Args: c: a const Tensor or Module. name: the name of output Node. """ value = None r"""The const Tensor or Module""" # TODO: constant cache to reduce the size of dumped model _constant_cache = {} def __init__(self, c, name=None): super().__init__() assert isinstance(c, (RawTensor, Module)) if isinstance(c, Module): assert module_tracer.is_builtin(c) or c.is_qat self.value = c self.name = name self.inputs = [] node_cls = NodeMixin.get_wrapped_type(c) self.outputs = [ node_cls(self, name=name, orig_name=name), ] self.outputs[0]._name = name if name else "const_" + str(self._id)
[文档] @classmethod def make(cls, *args, **kwargs): expr = cls(*args, **kwargs) name = "const_module" if isinstance(expr.value, Module) else "const_tensor" full_name = name if ( isinstance(expr.value, RawTensor) and id(expr.value) in active_module_tracer().id2name ): full_name = active_module_tracer().id2name[id(expr.value)] scope_name = active_module_tracer().current_scope()._module_name if full_name and scope_name: full_name = ("self." + full_name)[len(scope_name) + 1 :] else: full_name = name else: full_name = name name = active_module_tracer().current_scope()._create_unique_name(full_name) expr.outputs[0]._name = name expr.outputs[0]._orig_name = full_name active_module_tracer().current_scope()._insert(expr) return expr.outputs[0]
[文档] def interpret(self, *inputs): if isinstance(self.value, RawTensor): return Const(self.value.numpy())() return (self.value,)
def __repr__(self): name = self.name if name is None: name = type(self.value) node_type = "Module" if isinstance(self.outputs[0], TensorNode): node_type = "Tensor" return "%{}:\t{} = Constant({}) -> ({})".format( self._id, self.outputs[0], name, node_type ) def __getstate__(self): state = self.__dict__.copy() if "_top_graph" in state: state.pop("_top_graph") if isinstance(self.value, RawTensor): state["value"] = Tensor(self.value) return state