feat(traced_module): support trace custom qat module

GitOrigin-RevId: 49f70a5f46

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

@@ -17,12 +17,14 @@ 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, tree_flatten
+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):
@@ -76,6 +78,7 @@ class Expr:
                 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))
 
@@ -154,6 +157,11 @@ class Expr:
             return self._top_graph()
         return None
 
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state.pop("_top_graph", None)
+        return state
+
 
 # expr: None (i.e. fake expression which is used to mark input)
 class Input(Expr):
@@ -321,14 +329,36 @@ class Apply(Expr):
             ", ".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)
+
     @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))
-        apply_node = cls.make(opdef)
-        apply_node.add_inputs(inputs)
+
+        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()
@@ -387,7 +417,7 @@ class Constant(Expr):
         super().__init__()
         assert isinstance(c, (RawTensor, Module))
         if isinstance(c, Module):
-            assert module_tracer.is_builtin(c)
+            assert module_tracer.is_builtin(c) or c.is_qat
         self.value = c
         self.name = name
         self.inputs = []
@@ -395,6 +425,7 @@ class Constant(Expr):
         self.outputs = [
             node_cls(self, name=name),
         ]
+        self.outputs[0]._name = name if name else "const_" + str(self._id)
 
     @classmethod
     def make(cls, *args, **kwargs):
@@ -422,7 +453,7 @@ class Constant(Expr):
         )
 
     def __getstate__(self):
-        state = self.__dict__.copy()
+        state = super().__getstate__()
         if isinstance(self.value, RawTensor):
             state["value"] = Tensor(self.value)
         return state

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

@@ -0,0 +1,48 @@
+from copy import deepcopy
+from typing import Union
+
+from ...core.tensor.dtype import QuantDtypeMeta
+from ...quantization.fake_quant import QParamsModuleMixin, _FakeQuantize
+from ...quantization.utils import QParams, QuantMode, fake_quant_tensor
+
+
+class FakeQuantize(_FakeQuantize, QParamsModuleMixin):
+    def __init__(
+        self, dtype: Union[str, QuantDtypeMeta], enable: bool = True, **kwargs
+    ):
+        super().__init__(dtype, enable, **kwargs)
+        self.qparams = None
+
+    def fake_quant_forward(self, inp, qparams: QParams = None):
+        if qparams is None:
+            qparams = self.get_qparams()
+        assert (
+            qparams.dtype_meta is self.dtype
+        ), "input qparams' dtype is not equal to self.dtype.\nqparams.dtype_meta={}\nself.dtype={}".format(
+            qparams.dtype_meta, self.dtype
+        )
+        return fake_quant_tensor(inp, qparams)
+
+    def get_qparams(self):
+        return self.qparams
+
+    def set_qparams(self, qparams: QParams):
+        """
+        :param qparams: used to set initial scale.
+        """
+        if qparams.scale is None:
+            raise AssertionError("Can not get an initialized scale")
+        scale = qparams.scale
+        if qparams.dtype_meta is None:
+            qparams.dtype_meta = self.dtype
+        else:
+            assert (
+                qparams.dtype_meta is self.dtype
+            ), "input qparams' dtype is not equal to self.dtype.\nqparams.dtype_meta={}\nself.dtype={}".format(
+                qparams.dtype_meta, self.dtype
+            )
+        dtype_meta = qparams.dtype_meta
+        zero_point = qparams.zero_point
+        mode = qparams.mode
+
+        self.qparams = QParams(mode, dtype_meta, scale, zero_point)

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

@@ -12,6 +12,7 @@ from ... import Tensor
 from ... import functional as F
 from ...core.tensor.array_method import ArrayMethodMixin
 from ...module import Module
+from ...module.qat import QATModule
 
 _active_module_tracer = None
 
@@ -68,7 +69,7 @@ BUILTIN_ARRAY_METHOD = [
     "__iand__",
     "__ior__",
     "__ixor__",
-    "T",
+    "transpose",
     "astype",
     "reshape",
     "_broadcast",
@@ -180,6 +181,7 @@ class Patcher:
             self.patch_method(ArrayMethodMixin, meth, self.wrap_fn)
         self.patch_method(Tensor, "detach", self.wrap_fn)
         self.patch_method(Tensor, "__new__", self.wrap_fn)
+        self.patch_method(QATModule, "_apply_fakequant_with_observer", 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)

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

@@ -127,7 +127,7 @@ class TensorNode(Node):
 
     shape = None  # type: Tuple[int]
     dtype = None  # type: numpy.dtype
-    qparam = None
+    qparams = None
     device = None
 
     def __getstate__(self):
@@ -135,7 +135,7 @@ class TensorNode(Node):
             "expr": self.expr,
             "users": self.users,
             "_id": self._id,
-            "qparam": self.qparam,
+            "qparams": self.qparams,
             "shape": self.shape,
             "dtype": self.dtype,
             "device": self.device,

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

@@ -155,10 +155,7 @@ def tree_flatten(
         assert is_leaf(values), values
         node = LeafDef(leaf_type(values))
         if is_const_leaf(values):
-            if isinstance(values, np.ndarray):
-                node.const_val = str(values)
-            else:
-                node.const_val = values
+            node.const_val = values
         return [values,], node
 
     rst = []
@@ -232,9 +229,13 @@ class LeafDef(TreeDef):
         return leaves[0]
 
     def __eq__(self, other):
+        if isinstance(self.const_val, np.ndarray):
+            return self.type == other.type and (self.const_val == other.const_val).all()
         return self.type == other.type and self.const_val == other.const_val
 
     def __hash__(self):
+        if isinstance(self.const_val, np.ndarray):
+            return hash(tuple([self.type, str(self.const_val)]))
         return hash(tuple([self.type, self.const_val]))
 
     def __repr__(self):

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

@@ -29,14 +29,20 @@ from ...core._imperative_rt.core2 import (
 from ...core._trace_option import set_symbolic_shape
 from ...core.tensor.array_method import ArrayMethodMixin
 from ...module import Module
-from ...quantization.fake_quant import LSQ, TQT, FakeQuantize
+from ...module.qat import QATModule
+from ...quantization.fake_quant import LSQ, TQT, FakeQuantize, _FakeQuantize
 from ...quantization.observer import (
     ExponentialMovingAverageObserver,
+    HistogramObserver,
     MinMaxObserver,
+    Observer,
+    PassiveObserver,
+    SyncExponentialMovingAverageObserver,
     SyncMinMaxObserver,
 )
 from ...tensor import Tensor
 from .expr import Apply, CallFunction, CallMethod, Constant, Expr, GetAttr, Input
+from .fake_quant import FakeQuantize as TM_FakeQuant
 from .module_tracer import (
     Patcher,
     active_module_tracer,
@@ -613,7 +619,8 @@ def _wrapped_function(orig_func):
                     if isinstance(i, (RawTensor, NodeMixin)):
                         NodeMixin.wrap_safe(i, Constant.make(i))
             meth_name = _get_meth_name(args[0], wrapped_fn) if args else None
-            if meth_name:
+            arg_type = args[0] if isinstance(args[0], type) else type(args[0])
+            if meth_name and issubclass(arg_type, RawTensor):
                 self = inputs[0]
                 if meth_name == "__new__":
                     if all([not isinstance(i, RawTensor) for i in inputs]):
@@ -680,7 +687,15 @@ class TracedModuleBuilder(NodeMixin):
         self._mod = mod
         self._body = None
         self._is_top = is_top_module
-        self._is_builtin = module_tracer.is_builtin(mod)
+        self._is_builtin = (
+            True
+            if isinstance(mod, (Observer, _FakeQuantize))
+            else module_tracer.is_builtin(mod)
+        )
+        if isinstance(self._mod, QATModule):
+            unset_module_tracing()
+            self._check_qat_module(self._mod)
+            set_module_tracing()
         self._argdef_graph_map = {}
         self._argdef_outdef_map = {}
 
@@ -693,15 +708,65 @@ class TracedModuleBuilder(NodeMixin):
             dict(TracedModuleBuilder.__dict__),
         )
 
+    def _check_qat_module(self, qat_module):
+        def isbuiltin(m):
+            return m is None or module_tracer.is_builtin(m)
+
+        if qat_module.with_act:
+            act_observer = qat_module.act_observer
+            act_fakequant = qat_module.act_fake_quant
+            if not isbuiltin(act_observer) or not isbuiltin(act_fakequant):
+                qparams = (
+                    act_observer.get_qparams()
+                    if hasattr(act_observer, "get_qparams")
+                    else act_fakequant.get_qparams()
+                )
+                dtype = (
+                    act_observer.dtype
+                    if hasattr(act_observer, "dtype")
+                    else act_fakequant.dtype
+                )
+                qat_module.act_observer = None
+                qat_module.act_fake_quant = TM_FakeQuant(dtype)
+                qat_module.act_fake_quant.set_qparams(qparams)
+
+        if qat_module.with_weight:
+            weight_observer = qat_module.weight_observer
+            weight_fakequant = qat_module.weight_fake_quant
+            if not isbuiltin(weight_observer) or not isbuiltin(weight_fakequant):
+                qparams = (
+                    weight_observer.get_qparams()
+                    if hasattr(weight_observer, "get_qparams")
+                    else weight_fakequant.get_qparams()
+                )
+                dtype = (
+                    weight_observer.dtype
+                    if hasattr(weight_observer, "dtype")
+                    else weight_fakequant.dtype
+                )
+                qat_module.weight_observer = None
+                qat_module.weight_fake_quant = TM_FakeQuant(dtype)
+                qat_module.weight_fake_quant.set_qparams(qparams)
+
     def build(self):
         if self._is_builtin or isinstance(self._mod, TracedModule):
+            if module_tracer.is_builtin(self._mod) or isinstance(
+                self._mod, TracedModule
+            ):
+                mod_type = type(self._mod)
+            else:
+                assert isinstance(self._mod, (Observer, _FakeQuantize))
+                mod_type = (
+                    Observer if isinstance(self._mod, Observer) else _FakeQuantize
+                )
             for node in self.nodes:
-                node.module_type = type(self._mod)
-                # node._owner = weakref.ref(self._mod)
+                node.module_type = mod_type
+
             return self._mod
         else:
+            is_qat = isinstance(self._mod, QATModule)
             traced_module = TracedModule(
-                self._is_top, self._argdef_graph_map, self._argdef_outdef_map
+                self._is_top, self._argdef_graph_map, self._argdef_outdef_map, is_qat
             )
             for _, g in self._argdef_graph_map.items():
                 g.compile()
@@ -712,6 +777,20 @@ class TracedModuleBuilder(NodeMixin):
                         v = v.build()
                     setattr(traced_module, k, v)
 
+            if isinstance(self._mod, QATModule):
+                unset_module_tracing()
+                traced_module.with_act = self._mod.with_act
+                traced_module.with_weight = self._mod.with_weight
+                if not hasattr(traced_module, "act_fake_quant"):
+                    traced_module.act_fakequant = None
+                if not hasattr(traced_module, "act_observer"):
+                    traced_module.act_observer = None
+                if not hasattr(traced_module, "weight_fake_quant"):
+                    traced_module.weight_fakequant = None
+                if not hasattr(traced_module, "weight_observer"):
+                    traced_module.weight_observer = None
+                set_module_tracing()
+
             return traced_module
 
     def _record_wrapped_nodes(self, node):
@@ -846,7 +925,8 @@ class TracedModuleBuilder(NodeMixin):
             attr = getattr(self._mod, name)
             if isinstance(attr, Module):
                 attr = TracedModuleBuilder(attr)
-            setattr(self, name, attr)
+            if isinstance(attr, (Module, RawTensor)):
+                setattr(self, name, attr)
             NodeMixin.wrap(
                 attr,
                 lambda: GetAttr.make(
@@ -1066,7 +1146,7 @@ class TracedModule(Module):
     argdef_graph_map = None
     argdef_outdef_map = None
 
-    def __init__(self, is_top, argdef_graph_map, argdef_outdef_map):
+    def __init__(self, is_top, argdef_graph_map, argdef_outdef_map, is_qat=False):
         super(TracedModule, self).__init__()
         self.argdef_graph_map = argdef_graph_map
         self.argdef_outdef_map = argdef_outdef_map
@@ -1074,6 +1154,7 @@ class TracedModule(Module):
         self.watch_points = []
         self.watch_node_value = {}
         self.end_points = []
+        self.is_qat = is_qat
 
     def forward(self, *args, **kwargs):
         inputs, treedef = tree_flatten(((self, *args), kwargs))
@@ -1195,8 +1276,8 @@ class TracedModule(Module):
         ):
             if graph is not None and prefix_name and prefix_name[-1] != "_":
                 prefix_name += "_"
-            if graph is None:
-                assert not isinstance(module, TracedModule)
+            if graph is None or module.is_qat:
+                assert not isinstance(module, TracedModule) or module.is_qat
                 const = Constant(module, "self.%s" % module2name[id(module)])
                 m_node = call.inputs[0]
                 if m_node.top_graph != active_module_tracer().current_scope():
@@ -1326,9 +1407,23 @@ def _register_all_builtin_module():
                 isclass(m[1])
                 and issubclass(m[1], M.Module)
                 and m[1] is not M.Sequential
+                and m[1] is not M.ModuleList
             ):
                 module_tracer.register_as_builtin(m[1])
 
+    module_tracer.register_as_builtin(Observer)
+    module_tracer.register_as_builtin(MinMaxObserver)
+    module_tracer.register_as_builtin(SyncMinMaxObserver)
+    module_tracer.register_as_builtin(ExponentialMovingAverageObserver)
+    module_tracer.register_as_builtin(SyncExponentialMovingAverageObserver)
+    module_tracer.register_as_builtin(HistogramObserver)
+    module_tracer.register_as_builtin(PassiveObserver)
+
+    module_tracer.register_as_builtin(LSQ)
+    module_tracer.register_as_builtin(TQT)
+    module_tracer.register_as_builtin(FakeQuantize)
+    module_tracer.register_as_builtin(TM_FakeQuant)
+
 
 def trace_module(mod: Module, *args: Tensor, **kwargs: Tensor) -> TracedModule:
     """