feat(traced_module): add some functions of graph modification

GitOrigin-RevId: ac0603057a

imperative/python/megengine/experimental/traced_module/expr.py

@@ -9,6 +9,7 @@
 
 import builtins
 import collections
+import inspect
 from typing import Callable, List
 
 from ...core._imperative_rt import OpDef
@@ -16,10 +17,10 @@ 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.special import Const
 from ...module import Module
-from ...tensor import Tensor
+from ...tensor import Parameter, Tensor
 from .module_tracer import active_module_tracer, module_tracer
 from .node import ModuleNode, Node, NodeMixin, TensorNode
-from .pytree import TreeDef
+from .pytree import TreeDef, tree_flatten
 
 
 class Expr:
@@ -38,25 +39,28 @@ class Expr:
         for val in vals:
             node = NodeMixin.get(val, None)
             if isinstance(node, (TensorNode, ModuleNode)):
-                if node not in self.inputs:
-                    self.inputs.append(node)
+                self.inputs.append(node)
+                node.users.append(self)
             else:
                 assert node is None
-                assert type(val) in builtins.__dict__.values()
                 idx = len(self.inputs) + len(self.const_val)
                 self.const_val.append((idx, val))
 
-    def add_outputs(self, outputs):
+    def add_outputs(self, outputs, check_inplace=True):
         self.outputs = []
-        if not isinstance(outputs, collections.Sequence):
-            outputs = (outputs,)
+        if outputs is not None:
+            if not isinstance(outputs, collections.Sequence):
+                outputs = (outputs,)
 
-        for i in outputs:
-            assert isinstance(i, RawTensor)
-            self.outputs.append(NodeMixin.get_wrapped_type(i)(self))
+            for i in outputs:
+                assert isinstance(i, RawTensor)
+                node = NodeMixin.get(i, None) if check_inplace else None
+                self.outputs.append(
+                    node if node else NodeMixin.get_wrapped_type(i)(self)
+                )
 
-        for i, node in zip(outputs, self.outputs,):
-            NodeMixin.wrap_safe(i, node)
+            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:
@@ -110,6 +114,7 @@ class GetAttr(Expr):
         self.inputs = [
             module,
         ]
+        module.users.append(self)
         self.name = name
         node_cls = type if type else Node
         self.outputs = [
@@ -134,12 +139,20 @@ class GetAttr(Expr):
 
 # expr: outputs = inputs[0].__call__(*inputs[1:])
 class CallMethod(Expr):
-    def __init__(self, module, method="__call__"):
-        assert isinstance(module, (TensorNode, ModuleNode))
-        self.inputs = [
-            module,
-        ]
-        self.const_val = []
+    def __init__(self, node, method="__call__"):
+        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
 
     @classmethod
@@ -160,10 +173,13 @@ class CallMethod(Expr):
     def interpret(self, *inputs):
         args, kwargs = self.unflatten_args(inputs)
         obj = args[0]
-        args = args[1:]
+        meth = getattr(obj, self.method)
+        if inspect.ismethod(meth):
+            args = args[1:]
         outputs = getattr(obj, self.method)(*args, **kwargs)
-        if isinstance(outputs, RawTensor):
-            outputs = (outputs,)
+        if outputs is None:
+            return outputs
+        outputs, _ = tree_flatten(outputs, is_leaf=lambda x: isinstance(x, RawTensor))
         return outputs
 
     def __repr__(self):
@@ -171,7 +187,7 @@ class CallMethod(Expr):
         kwargs = ", ".join("{}={}".format(k, v) for k, v in self.kwargs.items())
         return "{} = {}.{}({})".format(
             ", ".join(str(i) for i in self.outputs),
-            self.inputs[0],
+            self.args[0],
             self.method,
             ", ".join([args, kwargs]),
         )
@@ -209,9 +225,8 @@ class Apply(Expr):
             if node is None:  # capture as constant
                 NodeMixin.wrap_safe(i, Constant.make(i))
         apply_node = cls.make(opdef)
-        for i in inputs:
-            assert isinstance(i, RawTensor)
-            apply_node.inputs.append(NodeMixin.get(i))
+        apply_node.add_inputs(inputs)
+        assert not apply_node.const_val
 
         unset_module_tracing()
         outputs = apply(opdef, *inputs)
@@ -283,7 +298,7 @@ class Constant(Expr):
         return (self.value,)
 
     def __repr__(self):
-        return "{} = Constant({})".format(self.outputs[0], self.value)
+        return "{} = Constant({})".format(self.outputs[0], type(self.value))
 
     def __getstate__(self):
         state = self.__dict__.copy()

imperative/python/megengine/experimental/traced_module/module_tracer.py

@@ -79,6 +79,8 @@ BUILTIN_ARRAY_METHOD = [
     "min",
     "max",
     "mean",
+    "__getitem__",
+    "__setitem__",
 ]
 
 
@@ -176,7 +178,8 @@ class Patcher:
             self.patch_module(module)
         for meth in BUILTIN_ARRAY_METHOD:
             self.patch_method(ArrayMethodMixin, meth, self.wrap_fn)
-
+        self.patch_method(Tensor, "detach", self.wrap_fn)
+        self.patch_method(Tensor, "__new__", self.wrap_fn)
         for i, j in self._builtin_functions:
             if id(i) not in self.visited_frames_ids:
                 self.patch_function(i, j, self.wrap_fn)
@@ -203,7 +206,13 @@ class Patcher:
         import inspect
 
         if id(module.__dict__) not in self.visited_frames_ids:
-            for k, v in module.__dict__.items():
+            keys = (
+                getattr(module, "__all__")
+                if hasattr(module, "__all__")
+                else module.__dict__.keys()
+            )
+            for k in keys:
+                v = getattr(module, k)
                 if inspect.isfunction(v) and not k.startswith("_"):
                     self.patch_function(module.__dict__, k, self.wrap_fn)
             self.visited_frames_ids.add(id(module.__dict__))

imperative/python/megengine/experimental/traced_module/node.py

@@ -6,7 +6,7 @@
 # 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.
-from typing import Any, Dict, Tuple, Type
+from typing import Any, Dict, List, Tuple, Type
 
 import numpy
 
@@ -31,6 +31,7 @@ class Node:
 
     def __init__(self, expr: "Expr", name: str = None):
         self.expr = expr
+        self.users = []  # List[Expr]
         self._id = Node.__total_id
         Node.__total_id += 1
         self._name = name
@@ -59,11 +60,13 @@ class ModuleNode(Node):
     module_type = Module  # type: Type[Module]
     attr_type_map = None  # type: Dict[str, Type[Any]]
     argdef_graph_map = None  # type: Dict[Treedef, "InternalGraph"]
+    argdef_outdef_map = None  # type: Dict[Treedef, Treedef]
 
     def __init__(self, expr: "Expr", name: str = None):
         super().__init__(expr, name)
         self.attr_type_map = {}
         self.argdef_graph_map = {}
+        self.argdef_outdef_map = {}
 
     def __repr__(self):
         if self._name is None:

imperative/python/megengine/experimental/traced_module/pytree.py

@@ -10,6 +10,8 @@
 import collections
 from typing import Callable, NamedTuple
 
+import numpy as np
+
 SUPPORTED_TYPE = {}
 
 NodeType = NamedTuple("NodeType", [("flatten", Callable), ("unflatten", Callable)])
@@ -33,7 +35,7 @@ def _dict_unflatten(inps, aux_data):
 
 
 register_supported_type(list, lambda x: (x, None), lambda x, aux_data: list(x))
-register_supported_type(tuple, lambda x: (x, None), lambda x, aux_data: list(x))
+register_supported_type(tuple, lambda x: (x, None), lambda x, aux_data: tuple(x))
 register_supported_type(dict, _dict_flatten, _dict_unflatten)
 register_supported_type(
     slice,
@@ -52,7 +54,10 @@ def tree_flatten(
         assert is_leaf(values)
         node = LeafDef(leaf_type(values))
         if is_const_leaf(values):
-            node.const_val = values
+            if isinstance(values, np.ndarray):
+                node.const_val = str(values)
+            else:
+                node.const_val = values
         return [values,], node
 
     rst = []

imperative/python/megengine/experimental/traced_module/traced_module.py

@@ -10,8 +10,13 @@ import collections
 import copy
 import functools
 from inspect import getmembers, isclass, ismethod
-from typing import Dict, List, Type
+from typing import Callable, Dict, Iterable, List, Sequence, Type
 
+import numpy as np
+from numpy.lib.arraysetops import isin
+
+from ... import functional as F
+from ... import get_logger
 from ... import module as M
 from ...core._imperative_rt.core2 import Tensor as RawTensor
 from ...core._imperative_rt.core2 import (
@@ -19,6 +24,7 @@ from ...core._imperative_rt.core2 import (
     set_module_tracing,
     unset_module_tracing,
 )
+from ...core._trace_option import set_symbolic_shape
 from ...core.tensor.array_method import ArrayMethodMixin
 from ...module import Module
 from ...tensor import Tensor
@@ -32,6 +38,8 @@ from .module_tracer import (
 from .node import ModuleNode, Node, NodeMixin, TensorNode
 from .pytree import tree_flatten
 
+logger = get_logger(__name__)
+
 
 def _leaf_type(node):
     if isinstance(node, RawTensor):
@@ -42,6 +50,11 @@ def _leaf_type(node):
         return type(node)
 
 
+def _is_leaf(node):
+    assert isinstance(node, RawTensor), type(node)
+    return isinstance(node, RawTensor)
+
+
 def _is_const_leaf(node):
     if isinstance(node, (RawTensor, NodeMixin, Module)):
         return False
@@ -80,7 +93,13 @@ class InternalGraph:
 
     @property
     def exprs(self):
-        return _expr_list(self)
+        return ExprFilter(_expr_iter(self))
+
+    def get_call_function(self, func: Callable = None):
+        return self.exprs.call_function(func)
+
+    def get_call_method(self, method: str = None):
+        return self.exprs.call_method(method)
 
     def add_input(self, i):
         self._inputs.append(i)
@@ -88,16 +107,131 @@ class InternalGraph:
     def add_output(self, o):
         self._outputs.append(o)
 
+    def get_dep_exprs(self, nodes: Sequence[Node]) -> List[Expr]:
+        if not isinstance(nodes, Sequence):
+            nodes = (nodes,)
+        ret = list()
+        queue = list(nodes)
+        while queue:
+            node = queue.pop()
+            expr = node.expr
+            if expr not in ret:
+                ret.append(expr)
+            for i in expr.inputs:
+                if i not in queue:
+                    queue.append(i)
+        return ret
+
+    def insert_call_function(self, func: Callable, nodes: Sequence[Node]):
+        if not isinstance(nodes, Sequence):
+            nodes = [nodes]
+        assert isinstance(func, Callable)
+        for i in nodes:
+            assert isinstance(
+                i, TensorNode
+            ), "CallFunction only accept TensorNode as inputs"
+
+        expr = CallFunction(func)
+        expr.inputs = nodes
+
+        for i in nodes:
+            i.users.append(expr)
+
+        idx = max(self._exprs.index(i.expr) for i in nodes) + 1
+        self._exprs.insert(idx, expr)
+
+        fake_inp_val = tuple(F.zeros(shape=i.shape, dtype=i.dtype) for i in nodes)
+        fake_out_val = func(*fake_inp_val)
+
+        def create_node(val: Tensor):
+            node = TensorNode(expr)
+            node.shape = val.shape
+            node.dtype = val.dtype
+            return node
+
+        out_nodes = list(create_node(i) for i in fake_out_val)
+        expr.outputs = out_nodes
+
+        return out_nodes
+
+    def insert_call_method(self, target, method, args):
+        if not isinstance(args, Sequence):
+            args = [args]
+        assert isinstance(target, (TensorNode, ModuleNode))
+        assert isinstance(method, str)
+        for i in args:
+            assert isinstance(i, TensorNode)
+
+        expr = CallMethod(method)
+        expr.inputs = [target, *args]
+
+        if isinstance(target, TensorNode):
+            fake_target_val = F.zeros(shape=target.shape, dtype=target.dtype)
+            fake_inp_val = tuple(F.zeros(shape=i.shape, dtype=i.dtype) for i in args)
+            fake_out_val = getattr(fake_target_val, method)(fake_inp_val)
+
+            def create_node(val: Tensor):
+                node = TensorNode(expr)
+                node.shape = val.shape
+                node.dtype = val.dtype
+                return node
+
+            out_nodes = list(create_node(i) for i in fake_out_val)
+            expr.outputs = out_nodes
+        else:
+            raise NotImplementedError()
+
+        return out_nodes
+
+    def replace_node(self, repl_dict: Dict[Node, Node]):
+        while repl_dict:
+            node, repl_node = repl_dict.popitem()
+            # check graph inputs and outputs
+            assert node not in self.inputs, "Cannot replace inputs"
+            for i, n in enumerate(self.outputs):
+                if n is node:
+                    self.outputs[i] = repl_node
+            # update users of node and repl_node
+            # update inputs of expr in node.users
+            dep_exprs = self.get_dep_exprs(repl_node)
+            i = 0
+            while i < len(node.users):
+                n = node.users[i]
+                if n in dep_exprs:
+                    logger.info("Find a loop: ignore this replacement once")
+                    logger.info("node: %s" % node.__repr__())
+                    logger.info("repl_node: %s" % repl_node.__repr__())
+                    i += 1
+                    continue
+                repl_node.users.append(n)
+                node.users.pop(i)
+                idx = n.inputs.index(node)
+                n.inputs[idx] = repl_node
+
+    def compile(self):
+        """
+        Delete unused expr.
+        """
+        dep_exprs = self.get_dep_exprs(self.outputs)
+        i = 0
+        while i < len(self._exprs):
+            expr = self._exprs[i]
+            if expr in dep_exprs:
+                i += 1
+                continue
+            for n in expr.inputs:
+                n.users.remove(expr)
+            self._exprs.remove(expr)
+
     def interpret(self, *inputs):
-        # TODO: support kwargs ?
-        # TODO: skip expressions which are independent and have no side effect
         node2value = {}
         for n, v in zip(self._inputs, inputs):
             node2value[n] = v
         for expr in self._exprs:
             values = expr.interpret(*list(node2value[i] for i in expr.inputs))
-            for n, v in zip(expr.outputs, values):
-                node2value[n] = v
+            if values is not None:
+                for n, v in zip(expr.outputs, values):
+                    node2value[n] = v
         return list(node2value[i] for i in self._outputs)
 
     def __repr__(self):
@@ -109,7 +243,8 @@ class InternalGraph:
 
 
 def _get_meth_name(obj, func):
-    for cls in type(obj).mro():
+    tp = obj if isinstance(obj, type) else type(obj)
+    for cls in tp.mro():
         for k, v in cls.__dict__.items():
             if v == func:
                 return k
@@ -131,15 +266,31 @@ def _wrapped_function(orig_func):
             meth_name = _get_meth_name(args[0], wrapped_fn)
             if meth_name:
                 self = inputs[0]
-                call_node = CallMethod.make(NodeMixin.get(self), meth_name)
+                if meth_name == "__new__":
+                    if all([not isinstance(i, RawTensor) for i in inputs]):
+                        # only trace Tensor.__new__() when there are tensors in args
+                        set_module_tracing()
+                        return orig_func(*args, **kwargs)
+                    if isinstance(args[1], RawTensor):
+                        node = NodeMixin.get(inputs[1])
+                        inputs[1] = copy.copy(inputs[1])
+                        # copy inputs[1] to avoid tensor and Tensor(tensor) share same m_tensor, which will cause they have same _NodeMixin__node in tracing.
+                        NodeMixin.wrap_safe(inputs[1], node)
+                        args, kwargs = tree_def.unflatten(inputs)
+                    call_node = CallMethod.make(self, meth_name)
+                else:
+                    call_node = CallMethod.make(NodeMixin.get(self), meth_name)
+                call_node.add_inputs(inputs[1:])
             else:
                 call_node = CallFunction.make(orig_func)
-
-            call_node.add_inputs(inputs)
+                call_node.add_inputs(inputs)
 
             call_node.arg_def = tree_def
             outputs = orig_func(*args, **kwargs)
-            call_node.add_outputs(outputs)
+            if meth_name == "__new__":
+                call_node.add_outputs(outputs, False)
+            else:
+                call_node.add_outputs(outputs)
             set_module_tracing()
             return outputs
         return orig_func(*args, **kwargs)
@@ -197,13 +348,14 @@ class TracedModuleBuilder(NodeMixin):
             mark_constant(i)
         callnode = CallMethod.make(NodeMixin.get(self))
 
-        callnode.add_inputs(inputs)
+        callnode.add_inputs(inputs[1:])
 
         callnode.arg_def = tree_def
 
         if self._is_builtin:
             unset_module_tracing()
-            outputs = self._mod(*args, **kwargs)
+            rst = self._mod(*args, **kwargs)
+            outputs, out_def = tree_flatten(rst, leaf_type=_leaf_type, is_leaf=_is_leaf)
             set_module_tracing()
             if self._is_builtin:
                 self._body = None
@@ -215,14 +367,13 @@ class TracedModuleBuilder(NodeMixin):
             NodeMixin.wrap_safe(
                 self, Input.make("self", NodeMixin.get_wrapped_type(self))
             )
+            origin_inp_node = [NodeMixin.get(i, None) for i in inputs[1:]]
             # prepare args and kwargs for inner graph
             def wrap(x):
-                wrapped = copy.copy(x)  # FIXME
                 NodeMixin.wrap(
-                    wrapped,
-                    lambda: Input.make(type=NodeMixin.get_wrapped_type(wrapped)),
+                    x, lambda: Input.make(type=NodeMixin.get_wrapped_type(x)),
                 )
-                return wrapped
+                return x
 
             args = [self]
             for i in inputs[1:]:
@@ -231,21 +382,25 @@ class TracedModuleBuilder(NodeMixin):
             active_module_tracer().patcher.auto_patch(
                 getattr(getattr(self._mod, "forward", self._mod), "__globals__", {})
             )
-            outputs = type(self._mod).forward(*args, **kwargs)
-
+            rst = type(self._mod).forward(*args, **kwargs)
+            outputs, out_def = tree_flatten(rst, leaf_type=_leaf_type, is_leaf=_is_leaf)
             for i in (
                 outputs if isinstance(outputs, collections.abc.Sequence) else (outputs,)
             ):
                 active_module_tracer().current_scope().add_output(NodeMixin.get(i))
 
             NodeMixin.wrap_safe(self, orig_self)
+            for arg, node in zip(inputs[1:], origin_inp_node):
+                if node:
+                    NodeMixin.wrap_safe(arg, node)
             active_module_tracer().pop_scope()
 
         # rebind output to outer graph
         callnode.add_outputs(outputs)
         self_node = NodeMixin.get(self)
         self_node.argdef_graph_map[callnode.arg_def] = self._body
-        return outputs
+        self_node.argdef_outdef_map[callnode.arg_def] = out_def
+        return rst
 
     def __getattr__(self, name):
         if name not in self._mod.__dict__:
@@ -268,20 +423,29 @@ class TracedModuleBuilder(NodeMixin):
             return super().__getattribute__(name)
         else:
             wrapped = super().__getattribute__(name)
-            if name in self._mod.__dict__ and not NodeMixin.get(wrapped, None):
-                assert not self._is_builtin
-                NodeMixin.wrap(
-                    wrapped,
-                    lambda: GetAttr.make(
+            if name in self._mod.__dict__:
+                if not NodeMixin.get(wrapped, None):
+                    assert not self._is_builtin
+                    NodeMixin.wrap(
+                        wrapped,
+                        lambda: GetAttr.make(
+                            NodeMixin.get(self),
+                            name,
+                            type=NodeMixin.get_wrapped_type(wrapped),
+                        ),
+                    )
+                else:
+                    node = NodeMixin.get(wrapped)
+                    expr = GetAttr.make(
                         NodeMixin.get(self),
                         name,
                         type=NodeMixin.get_wrapped_type(wrapped),
-                    ),
-                )
+                    ).expr
+                    expr.outputs[0] = node
             return wrapped
 
 
-class _expr_list:
+class _expr_iter:
     def __init__(self, graph: InternalGraph):
         self.graph = graph
 
@@ -295,6 +459,59 @@ class _expr_list:
                 yield expr
 
 
+class ExprFilter:
+    def __init__(self, expr_iter: Iterable):
+        self._iter = expr_iter
+
+    def __iter__(self):
+        return iter(self._iter)
+
+    def call_function(self, func):
+        return ExprFilterCallFunction(self, func)
+
+    def call_method(self, method):
+        return ExprFilterCallMethod(self, method)
+
+    def as_list(self):
+        return list(self)
+
+    def as_dict(self):
+        raise NotImplementedError("need key")
+
+    def as_unique(self):
+        (expr,) = self
+        return expr
+
+    def as_count(self):
+        return sum(1 for _ in self)
+
+
+class ExprFilterCallFunction(ExprFilter):
+    def __init__(self, expr_iter, func: Callable = None):
+        super().__init__(expr_iter)
+        self.func = func
+
+    def __iter__(self):
+        for i in self._iter:
+            if not isinstance(i, CallFunction):
+                continue
+            if self.func is None or i.func == self.func:
+                yield i
+
+
+class ExprFilterCallMethod(ExprFilter):
+    def __init__(self, expr_iter, method: str = None):
+        super().__init__(expr_iter)
+        self.method = method
+
+    def __iter__(self):
+        for i in self._iter:
+            if not isinstance(i, CallMethod):
+                continue
+            if self.method is None or i.method == self.method:
+                yield i
+
+
 class TracedModule(Module):
     """
     `TracedModule` is the Module created by tracing normal module. It owns a ModuleNode(m_node). `TracedModule` can not be called directly. It can be 
@@ -312,10 +529,12 @@ class TracedModule(Module):
             ((self, *args), kwargs), _leaf_type, is_const_leaf=_is_const_leaf
         )
         assert treedef in self.m_node.argdef_graph_map
-        inputs = [i for i in inputs if isinstance(i, (Module, RawTensor))]
+        inputs = filter(
+            lambda i: isinstance(i, (Module, TracedModuleBuilder, RawTensor)), inputs
+        )  # allow TracedModuleBuilder for retrace.
         outputs = self.m_node.argdef_graph_map[treedef].interpret(*inputs)
-        if len(outputs) == 1:
-            return outputs[0]
+        out_def = self.m_node.argdef_outdef_map[treedef]
+        outputs = out_def.unflatten(outputs)
         return outputs
 
     @property
@@ -339,9 +558,8 @@ class TracedModule(Module):
             if graph is None:
                 assert not isinstance(module, TracedModule)
                 const = Constant(module)
-                modulenode = const.outputs[0]
-                modulenode.module_type = type(module)
-                call.inputs[0] = modulenode
+                const.outputs[0] = call.inputs[0]
+                const.outputs[0].expr = const
                 return [const, call]
             exprs = []
             for expr in graph._exprs:
@@ -350,30 +568,41 @@ class TracedModule(Module):
                     if call and inp in graph._inputs:
                         inp_idx = graph._inputs.index(inp)
                         expr.inputs[idx] = call.inputs[inp_idx]
+                        call.inputs[inp_idx].users.append(expr)
                 # replace outputs for submodule's expr
                 for idx, outp in enumerate(expr.outputs):
                     if call and outp in graph._outputs:
                         oup_idx = graph._outputs.index(outp)
                         expr.outputs[idx] = call.outputs[oup_idx]
+                        call.outputs[oup_idx].expr = expr
 
                 if isinstance(expr, GetAttr):
                     # replace GetAttr with Constant
                     if isinstance(expr.outputs[0], TensorNode):
                         const = Constant(getattr(module, expr.name))
                         const.outputs = expr.outputs
+                        const.outputs[0].expr = const
                         exprs.append(const)
 
                 elif isinstance(expr, CallMethod):
                     obj_node = expr.inputs[0]
                     if isinstance(obj_node, ModuleNode):
-                        assert isinstance(expr.inputs[0].expr, GetAttr)
-                        (obj,) = expr.inputs[0].expr.interpret(module)
-                        exprs.extend(_flatten_subgraph(expr.graph, obj, expr))
+                        pre_expr = expr.inputs[0].expr
+                        if isinstance(pre_expr, GetAttr):
+                            (obj,) = expr.inputs[0].expr.interpret(module)
+                            exprs.extend(_flatten_subgraph(expr.graph, obj, expr))
+                        else:
+                            # module has been replaced.
+                            assert isinstance(pre_expr, Constant)
                     else:
                         exprs.append(expr)
                 else:
                     exprs.append(expr)
 
+            if call is not None:
+                for i in call.inputs:
+                    i.users.remove(call)
+
             return exprs
 
         new_module.graph._exprs = _flatten_subgraph(new_module.graph, new_module)
@@ -422,22 +651,26 @@ def trace_module(mod: Module, *args: Tensor, **kwargs: Tensor) -> TracedModule:
     """
     assert active_module_tracer() is None
     try:
+        use_sym_shape = set_symbolic_shape(True)
         set_module_tracing()
         set_active_module_tracer(module_tracer(_wrapped_function))
+
         with active_module_tracer().patcher:
             global_scope = InternalGraph()
             active_module_tracer().push_scope(global_scope)
 
             builder = TracedModuleBuilder(mod, True)
             NodeMixin.wrap_safe(builder, Input.make("TopModule", ModuleNode))
-            inputs, _ = tree_flatten((args, kwargs))
+            inputs, _ = tree_flatten((args, kwargs), is_const_leaf=_is_const_leaf)
             for _, i in enumerate(inputs):
-                NodeMixin.wrap_safe(
-                    i, Input.make("arg_{}".format(_), NodeMixin.get_wrapped_type(i))
-                )
+                if isinstance(i, RawTensor):
+                    NodeMixin.wrap_safe(
+                        i, Input.make("arg_{}".format(_), NodeMixin.get_wrapped_type(i))
+                    )
             builder(*args, **kwargs)
             active_module_tracer().pop_scope()
             return builder.build()
     finally:
+        set_symbolic_shape(use_sym_shape)
         set_active_module_tracer(None)
         unset_module_tracing()

imperative/python/test/unit/traced_module/test_haoruitao.py

@@ -0,0 +1,90 @@
+import io
+import pickle
+
+import numpy as np
+
+import megengine.functional as F
+import megengine.module as M
+import megengine.utils.comp_graph_tools as cgtools
+from megengine.core._trace_option import set_symbolic_shape
+from megengine.experimental.traced_module import trace_module
+from megengine.jit import trace
+
+set_symbolic_shape(True)
+
+
+class Main(M.Module):
+    def forward(self, x):
+        return x
+
+
+class PreProcess(M.Module):
+    def __init__(self):
+        super().__init__()
+        self.I = F.ones((1,))
+        self.M = F.zeros((1,))
+
+    def forward(self, data, idx, roi):
+        N, H, W, C = data.shape
+        xmax = roi[:, 1, 0]
+        xmin = roi[:, 0, 0]
+        ymax = roi[:, 1, 1]
+        ymin = roi[:, 0, 1]
+        scale = F.maximum((xmax - xmin) / W, (ymax - ymin) / H)
+        I = F.broadcast_to(self.I, (N,))
+        M = F.broadcast_to(self.M, (N, 3, 3))
+        M[:, 0, 0] = scale
+        M[:, 0, 2] = xmin
+        M[:, 1, 1] = scale
+        M[:, 1, 2] = ymin
+        M[:, 2, 2] = I
+        resized = (
+            F.warp_perspective(
+                data, M, (H, W), mat_idx=idx, border_mode="CONSTANT", format="NHWC"
+            )
+            .transpose(0, 3, 1, 2)
+            .astype(np.float32)
+        )
+        return resized
+
+
+class Net(M.Module):
+    def __init__(self, traced_module):
+        super().__init__()
+        self.pre_process = PreProcess()
+        self.traced_module = traced_module
+
+    def forward(self, data, idx, roi):
+        x = self.pre_process(data, idx, roi)
+        x = self.traced_module(x)
+        return x
+
+
+def test_preprocess():
+    module = Main()
+    data = F.ones((1, 14, 8, 8), dtype=np.uint8)
+    traced_module = trace_module(module, data)
+    obj = pickle.dumps(traced_module)
+    traced_module = pickle.loads(obj)
+    module = Net(traced_module)
+    module.eval()
+    idx = F.zeros((1,), dtype=np.int32)
+    roi = F.ones((1, 2, 2), dtype=np.float32)
+    y = module(data, idx, roi)
+    traced_module = trace_module(module, data, idx, roi)
+    np.testing.assert_array_equal(traced_module(data, idx, roi), y)
+    func = trace(traced_module, capture_as_const=True)
+    np.testing.assert_array_equal(func(data, idx, roi), y)
+    model = io.BytesIO()
+    func.dump(model, arg_names=("data", "idx", "roi"))
+    model.seek(0)
+    infer_cg = cgtools.GraphInference(model)
+    np.testing.assert_allclose(
+        list(
+            infer_cg.run(
+                inp_dict={"data": data.numpy(), "idx": idx.numpy(), "roi": roi.numpy()}
+            ).values()
+        )[0],
+        y,
+        atol=1e-6,
+    )

imperative/python/test/unit/traced_module/test_modification.py

@@ -0,0 +1,113 @@
+# 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 numpy as np
+
+import megengine.functional as F
+import megengine.module as M
+from megengine.experimental.traced_module import trace_module
+from megengine.experimental.traced_module.expr import CallFunction, GetAttr
+
+
+class MyBlock(M.Module):
+    def __init__(self, in_channels=3, channels=3):
+        super(MyBlock, self).__init__()
+        self.conv1 = M.Conv2d(in_channels, channels, 3, 1, padding=1, bias=False)
+        self.bn1 = M.BatchNorm2d(channels)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = F.relu(x) + 1
+        return x
+
+
+class MyModule(M.Module):
+    def __init__(self):
+        super(MyModule, self).__init__()
+        self.block0 = MyBlock()
+        self.block1 = MyBlock()
+
+    def forward(self, x):
+        x = self.block0(x)
+        x = self.block1(x)
+        return x
+
+
+def _init_cls(cls):
+    module = cls()
+    x = F.ones((1, 3, 3, 3))
+    y = module(x)
+    traced_module = trace_module(module, x)
+    return traced_module, x, y
+
+
+def _init_block():
+    return _init_cls(MyBlock)
+
+
+def _init_module():
+    return _init_cls(MyModule)
+
+
+def test_search():
+    traced_module, *_ = _init_block()
+    graph = traced_module.graph
+    relu_expr = graph.get_call_function(F.relu).as_unique()
+    assert isinstance(relu_expr, CallFunction) and relu_expr.func == F.relu
+
+
+def test_insert():
+    traced_module, x, expect = _init_block()
+    graph = traced_module.graph
+    relu_node = graph.get_call_function(F.relu).as_unique().outputs
+    neg_node = graph.insert_call_function(F.neg, relu_node)
+    graph.replace_node({relu_node[0]: neg_node[0]})
+    graph.compile()
+    np.testing.assert_allclose(expect - 1, 1 - traced_module(x), atol=1e-6)
+
+
+def test_delete():
+    traced_module, x, expect = _init_block()
+    graph = traced_module.graph
+    relu_expr = graph.get_call_function(F.relu).as_unique()
+    node = relu_expr.outputs
+    repl_node = relu_expr.inputs
+    graph.replace_node({node[0]: repl_node[0]})
+    graph.compile()
+    np.testing.assert_allclose(expect - 1, F.relu(traced_module(x) - 1), atol=1e-6)
+
+
+def test_flatten():
+    traced_module, x, expect = _init_module()
+    traced_module = traced_module.flatten()
+    traced_module.graph.compile()
+    assert all(not isinstance(i, GetAttr) for i in traced_module.graph._exprs)
+    assert len(traced_module.graph._exprs) == 12
+
+
+def test_extra_block():
+    class PostProcess(M.Module):
+        def forward(self, x):
+            return x * 2
+
+    class Net(M.Module):
+        def __init__(self, traced_module):
+            super().__init__()
+            self.post_process = PostProcess()
+            self.traced_module = traced_module
+
+        def forward(self, x):
+            x = self.traced_module(x)
+            x = self.post_process(x)
+            return x
+
+    traced_module, x, expect = _init_block()
+    module = Net(traced_module)
+    np.testing.assert_allclose(2 * expect, module(x), atol=1e-6)
+    traced_module = trace_module(module, x)
+    np.testing.assert_allclose(2 * expect, traced_module(x), atol=1e-6)

imperative/python/test/unit/traced_module/test_serialization.py

@@ -0,0 +1,52 @@
+# 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 pickle
+
+import numpy as np
+
+import megengine.functional as F
+import megengine.module as M
+from megengine import Tensor
+from megengine.experimental.traced_module import trace_module
+from megengine.module import Module
+
+
+class MyBlock(Module):
+    def __init__(self, in_channels, channels):
+        super(MyBlock, self).__init__()
+        self.conv1 = M.Conv2d(in_channels, channels, 3, 1, padding=1, bias=False)
+        self.bn1 = M.BatchNorm2d(channels)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = F.relu(x) + 1
+        return x
+
+
+class MyModule(Module):
+    def __init__(self):
+        super(MyModule, self).__init__()
+        self.block0 = MyBlock(8, 4)
+        self.block1 = MyBlock(4, 2)
+
+    def forward(self, x):
+        x = self.block0(x)
+        x = self.block1(x)
+        return x
+
+
+def test_dump_and_load():
+    module = MyModule()
+    x = Tensor(np.ones((1, 8, 14, 14)))
+    expect = module(x)
+    traced_module = trace_module(module, x)
+    np.testing.assert_array_equal(expect, traced_module(x))
+    obj = pickle.dumps(traced_module)
+    pickle.loads(obj)
+    np.testing.assert_array_equal(expect, traced_module(x))

imperative/python/test/unit/traced_module/test_trace_module.py

@@ -0,0 +1,42 @@
+import numpy as np
+
+from megengine import Tensor
+from megengine.experimental.traced_module import trace_module
+from megengine.module import Module as M
+
+
+class MyModule1(M):
+    def forward(self, x):
+        y = Tensor(x)
+        y += 1
+        x = x + 2
+        return x, y
+
+
+class MyModule2(M):
+    def forward(self, x):
+        y = Tensor([1, x, 1])
+        y += 1
+        x = x + 2
+        return x, y
+
+
+def test_trace_module():
+
+    x = Tensor(1)
+    m1 = MyModule1()
+    tm1 = trace_module(m1, x)
+
+    m2 = MyModule2()
+    tm2 = trace_module(m2, x)
+    inp = Tensor(2)
+    gt = m1(inp)
+    output = tm1(inp)
+    for a, b in zip(output, gt):
+        np.testing.assert_equal(a.numpy(), b.numpy())
+
+    gt1 = m2(inp)
+    output1 = tm2(inp)
+
+    for a, b in zip(output1, gt1):
+        np.testing.assert_equal(a.numpy(), b.numpy())

imperative/python/test/unit/traced_module/test_wujianan.py

@@ -0,0 +1,94 @@
+import io
+import pickle
+
+import numpy as np
+
+import megengine as mge
+import megengine.functional as F
+import megengine.module as M
+import megengine.utils.comp_graph_tools as cgtools
+from megengine.core._trace_option import set_symbolic_shape
+from megengine.experimental.traced_module import trace_module
+from megengine.jit import trace
+
+set_symbolic_shape(True)
+
+
+class Main(M.Module):
+    def forward(self, x):
+        return x["data"]
+
+
+class PreProcess(M.Module):
+    def __init__(self):
+        super().__init__()
+        self.A = F.zeros((1,))
+        self.I = F.ones((1,))
+        self.bb_out = mge.tensor(
+            np.array([[[0, 0], [160, 0], [160, 48], [0, 48]]], dtype="float32")
+        )
+
+    def forward(self, data, quad):
+        """
+        data: (1, 3, 48, 160)
+        quad: (1, 4, 2)
+        """
+        N = quad.shape[0]
+        dst = F.repeat(self.bb_out, N, axis=0).reshape(-1, 4, 2)
+        I = F.broadcast_to(self.I, quad.shape)
+        A = F.broadcast_to(self.A, (N, 8, 8))
+        A[:, 0:4, 0:2] = quad
+        A[:, 4:8, 5:6] = I[:, :, 0:1]
+        A[:, 0:4, 6:8] = -quad * dst[:, :, 0:1]
+        A[:, 4:8, 3:5] = quad
+        A[:, 0:4, 2:3] = I[:, :, 0:1]
+        A[:, 4:8, 6:8] = -quad * dst[:, :, 1:2]
+        B = dst.transpose(0, 2, 1).reshape(-1, 8, 1)
+        M = F.concat([F.matmul(F.matinv(A), B)[:, :, 0], I[:, 0:1, 0]], axis=1).reshape(
+            -1, 3, 3
+        )
+        new_data = F.warp_perspective(data, M, (48, 160))  # (N, 3, 48, 160)
+        return {"data": new_data}
+
+
+class Net(M.Module):
+    def __init__(self, traced_module):
+        super().__init__()
+        self.pre_process = PreProcess()
+        self.traced_module = traced_module
+
+    def forward(self, data, quad):
+        x = self.pre_process(data, quad)
+        x = self.traced_module(x)
+        return x
+
+
+def test_preprocess():
+    batch_size = 2
+    module = Main()
+    data = mge.tensor(
+        np.random.randint(0, 256, size=(batch_size, 3, 48, 160)), dtype=np.float32
+    )
+    traced_module = trace_module(module, {"data": data})
+    obj = pickle.dumps(traced_module)
+    traced_module = pickle.loads(obj)
+    module = Net(traced_module)
+    module.eval()
+    quad = mge.tensor(np.random.normal(size=(batch_size, 4, 2)), dtype=np.float32)
+    expect = module(data, quad)
+    traced_module = trace_module(module, data, quad)
+    actual = traced_module(data, quad)
+    for i, j in zip(expect, actual):
+        np.testing.assert_array_equal(i, j)
+    func = trace(traced_module, capture_as_const=True)
+    actual = func(data, quad)
+    for i, j in zip(expect, actual):
+        np.testing.assert_array_equal(i, j)
+    model = io.BytesIO()
+    func.dump(model, arg_names=("data", "quad"))
+    model.seek(0)
+    infer_cg = cgtools.GraphInference(model)
+    actual = list(
+        infer_cg.run(inp_dict={"data": data.numpy(), "quad": quad.numpy()}).values()
+    )[0]
+    np.testing.assert_allclose(expect, actual)