support property process in TensorVariable

sot/infer_meta.py

@@ -19,14 +19,28 @@ def value_fn(self, *args, **kwargs):
 
 
 class MetaInfo:
-    def __init__(self, shape, dtype, stop_gradient):
+    def __init__(
+        self, shape, dtype, stop_gradient, name, persistable, type, place
+    ):
+        self.name = name
+        self.persistable = persistable
+        self.type = type
+        self.place = place
         self.shape = shape
         self.dtype = dtype
         self.stop_gradient = stop_gradient
 
     @staticmethod
     def from_tensor(tensor):
-        return MetaInfo(tensor.shape, tensor.dtype, tensor.stop_gradient)
+        return MetaInfo(
+            list(tensor.shape),
+            tensor.dtype,
+            tensor.stop_gradient,
+            tensor.name,
+            tensor.persistable,
+            tensor.type,
+            tensor.place,
+        )
 
     def is_dynamic_shape(self):
         """
@@ -128,11 +142,7 @@ def variable_to_meta_info(args):
     return map_if(
         args,
         pred=lambda x: isinstance(x, paddle.static.Variable),
-        true_fn=lambda x: MetaInfo(
-            list(x.shape),
-            x.dtype,
-            x.stop_gradient,
-        ),
+        true_fn=lambda x: MetaInfo.from_tensor(x),
         false_fn=lambda x: x,
     )
 
@@ -153,11 +163,7 @@ def infer_meta_for_layer(layer, *args, **kwargs):
     args, kwargs = convert_to_input_spec(args), convert_to_input_spec(kwargs)
     concrete_program = layer.forward.get_concrete_program(*args, **kwargs)[0]
     out = concrete_program.outputs[0]
-    out = MetaInfo(
-        list(out.shape),
-        out.dtype,
-        out.stop_gradient,
-    )
+    out = MetaInfo.from_tensor(out)
     layer.forward.rollback()
     return out
 

sot/opcode_translator/executor/variable_dispatch.py

@@ -4,6 +4,8 @@
 from functools import partial
 from typing import TYPE_CHECKING
 
+import paddle
+
 from ...utils import BreakGraphError, NotImplementException
 from ...utils.magic_methods import (
     BINARY_OPS,
@@ -97,6 +99,37 @@
     lambda var: var.bool(),
 )
 
+# TensorVariable
+Dispatcher.register(
+    paddle.is_tensor,
+    ("TensorVariable",),
+    {},
+    lambda var: var.is_tensor(),
+)
+Dispatcher.register(
+    paddle.is_complex,
+    ("TensorVariable",),
+    {},
+    lambda var: var.is_complex(),
+)
+Dispatcher.register(
+    paddle.is_integer,
+    ("TensorVariable",),
+    {},
+    lambda var: var.is_integer(),
+)
+Dispatcher.register(
+    paddle.is_floating_point,
+    ("TensorVariable",),
+    {},
+    lambda var: var.is_floating_point(),
+)
+Dispatcher.register(
+    paddle.rank,
+    ("TensorVariable",),
+    {},
+    lambda var: var.ndim,
+)
 
 # VariableBase
 Dispatcher.register(

sot/opcode_translator/executor/variables/basic.py

@@ -33,21 +33,34 @@
 if TYPE_CHECKING:
     from ..function_graph import FunctionGraph
 
-DTYPE_ABBRS = {
+
+FP_DTYPE_ABBRS = {
     paddle.bfloat16: 'bfloat16',
     paddle.float64: 'float64',
     paddle.float32: 'float32',
     paddle.float16: 'float16',
+}
+
+CP_DTYPE_ABBRS = {
     paddle.complex64: 'complex64',
     paddle.complex128: 'complex128',
+}
+
+INT_DTYPE_ABBRS = {
     paddle.int8: 'int8',
     paddle.int16: 'int16',
     paddle.int32: 'int32',
     paddle.int64: 'int64',
-    paddle.bool: 'bool',
     paddle.uint8: 'uint8',
 }
 
+DTYPE_ABBRS = {
+    **FP_DTYPE_ABBRS,
+    **CP_DTYPE_ABBRS,
+    **INT_DTYPE_ABBRS,
+    paddle.bool: 'bool',
+}
+
 
 class ConstantVariable(VariableBase):
     def __init__(
@@ -95,6 +108,14 @@ def wrap_literal(value: Any) -> ConstantVariable:
         return ConstantVariable(value, ConstTracker(value))
 
 
+implemented_property = set()
+
+
+def tensor_property(func):
+    implemented_property.add(func.__name__)
+    return property(func)
+
+
 class TensorVariable(VariableBase):
     var_name_generator = NameGenerator("var_")
 
@@ -186,7 +207,7 @@ def __setitem__(self, key, value):
             value,
         )
 
-    @property
+    @tensor_property
     def T(self):
         perm = list(range(len(self.meta.shape) - 1, -1, -1))
         perm_var = VariableFactory.from_value(
@@ -195,40 +216,75 @@ def T(self):
         out = self.graph.call_paddle_api(paddle.transpose, self, perm_var)
         return out
 
-    @property
+    @tensor_property
     def ndim(self):
         return ConstantVariable.wrap_literal(len(self.meta.shape))
 
-    @property
-    def shape(self):
+    @tensor_property
+    def size(self):
+        # TODO: maybe break graph.
         if self.meta.is_dynamic_shape():
             raise BreakGraphError(
                 f"Getting size for a dynamic shape tensor causes graph break. shape = {self.meta.shape}"
             )
+        elements = reduce(operator.mul, self.meta.shape, 1)
+        return ConstantVariable.wrap_literal(elements)
+
+    @tensor_property
+    def shape(self):
+        if self.meta.is_dynamic_shape():
+            raise BreakGraphError(
+                f"Getting shape for a dynamic shape tensor causes graph break. shape = {self.meta.shape}"
+            )
         self.graph.add_global_guarded_variable(self)
         return VariableFactory.from_value(
             self.meta.shape, self.graph, tracker=ConstTracker(self.meta.shape)
         )
 
-    @property
-    def size(self):
-        # TODO: maybe break graph.
-        if self.meta.is_dynamic_shape():
-            raise BreakGraphError(
-                f"Getting size for a dynamic shape tensor causes graph break. shape = {self.meta.shape}"
-            )
-        elements = reduce(operator.mul, self.meta.shape, 1)
-        return ConstantVariable.wrap_literal(elements)
+    def is_tensor(self):
+        return ConstantVariable.wrap_literal(True)
+
+    def is_complex(self):
+        dtype = self.meta.dtype
+        is_cp_dtype = dtype in CP_DTYPE_ABBRS
+        return ConstantVariable.wrap_literal(is_cp_dtype)
+
+    def is_integer(self):
+        dtype = self.meta.dtype
+        is_int_dtype = dtype in INT_DTYPE_ABBRS
+        return ConstantVariable.wrap_literal(is_int_dtype)
+
+    def is_floating_point(self):
+        dtype = self.meta.dtype
+        is_fp_dtype = dtype in FP_DTYPE_ABBRS
+        return ConstantVariable.wrap_literal(is_fp_dtype)
 
     def getattr(self, name: str):
-        if name in ["shape", "dtype", "stop_gradient"]:
+        method_name_to_builtin_fn = {
+            "dim": paddle.rank,
+            "ndimension": paddle.rank,
+            "is_tensor": paddle.is_tensor,
+            "is_complex": paddle.is_complex,
+            "is_integer": paddle.is_integer,
+            "is_floating_point": paddle.is_floating_point,
+        }
+        if name in ["dtype", "type", "persistable", "name", "stop_gradient"]:
             return VariableFactory.from_value(
                 getattr(self.meta, name),
                 self.graph,
                 tracker=GetAttrTracker(self, name),
             )
-        elif name in ["T", "ndim", "size"]:
+        elif name in implemented_property:
             return getattr(self, name)
+        elif name in method_name_to_builtin_fn:
+            # TODO: backward, gradient
+            from .callable import BuiltinVariable
+
+            builtin_fn = method_name_to_builtin_fn[name]
+
+            return BuiltinVariable(
+                builtin_fn, self.graph, DanglingTracker()
+            ).bind(self, name)
         elif name in paddle_tensor_methods:
             from .callable import TensorFunctionVariable
 

tests/test_18_tensor_method.py

@@ -26,7 +26,16 @@ def tensor_method_passed_by_user(a: paddle.Tensor, func: paddle.Tensor):
 
 
 def tensor_method_property(a: paddle.Tensor, b: paddle.Tensor):
-    return a @ b.T + len(a.shape) + b.size + a.ndim
+    return (
+        a.name,
+        str(a.place),
+        a.persistable,
+        a.dtype,
+        a.type,
+        a.is_tensor(),
+        a.clear_gradient(),
+        a @ b.T + len(a.shape) + b.size + a.ndim + a.dim() + a.rank(),
+    )
 
 
 class TestTensorMethod(TestCaseBase):