mindspore 2.0.0rc1__cp38-none-any.whl → 2.2.0__cp38-none-any.whl

mindspore/common/auto_dynamic_shape.py

@@ -0,0 +1,498 @@
+# This is the Python adaptation and derivative work of Myia (https://github.com/mila-iqia/myia/).
+#
+# Copyright 2020-2023 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""Providing auto dynamic shape interface methods."""
+
+import os
+from mindspore import log as logger
+from mindspore._c_expression import GraphExecutor_, Tensor
+from mindspore.common._utils import is_shape_unknown, is_dim_unknown
+from mindspore.common.parameter import Parameter
+
+SHAPE_DIM_ANY = -1
+SHAPE_RANK_ANY = -2
+
+auto_dynamic_shepe_dict = {}
+
+class _AutoDynamicShapeManager:
+    """
+    Represents a function to manage auto identify dynamic shape.
+    """
+    def __init__(self):
+        self.real_shape_cache = []
+        self.generalize_shape_cache = []
+        self.real_phase_and_compile_args_dict = {}
+        self.generalize_phase_and_compile_args_dict = {}
+        self._graph_executor = GraphExecutor_.get_instance()
+
+
+    def __del__(self):
+        self.real_shape_cache = []
+        self.generalize_shape_cache = []
+        self.real_phase_and_compile_args_dict = {}
+        self.generalize_phase_and_compile_args_dict = {}
+
+
+    @staticmethod
+    def is_tensor_equal(input_elem, cache_elem):
+        """check two tensor or param is equal"""
+        if input_elem.shape == cache_elem.shape and input_elem.dtype == cache_elem.dtype:
+            return True
+        return False
+
+
+    @staticmethod
+    def _get_input_generalize_number(arg_list, is_shape_input):
+        """check two tensor or param is equal"""
+        count = 0
+        if is_shape_input:
+            for arg in arg_list:
+                if is_shape_unknown(arg):
+                    count = count + 1
+        else:
+            for arg in arg_list:
+                if isinstance(arg, Tensor) and is_shape_unknown(arg.shape):
+                    count = count + 1
+        return count
+
+
+    def get_real_shape_cache_number(self):
+        """get real shape cache number"""
+        return len(self.real_shape_cache)
+
+
+    def get_real_shape_cache(self):
+        """get real shape cache"""
+        return self.real_shape_cache
+
+
+    def get_generalize_shape_cache_number(self):
+        """get generalize shape cache number"""
+        return len(self.generalize_shape_cache)
+
+
+    def get_generalize_shape_cache(self):
+        """get generalize shape cache"""
+        return self.generalize_shape_cache
+
+
+    def get_cache_by_type(self, cache_type):
+        """get cache by type"""
+        if cache_type == "real":
+            shape_cache = self.real_shape_cache
+        else:
+            shape_cache = self.generalize_shape_cache
+
+        return shape_cache
+
+
+    def get_compile_args_shape_without_sink(self, input_args, res_shape):
+        """get compile args shape with out sink mode"""
+        for input in input_args:
+            if isinstance(input, Tensor):
+                res_shape.append(input.shape)
+            elif isinstance(input, (int, float)):
+                res_shape.append([])
+            elif isinstance(input, (tuple, list)):
+                tmp_shape = []
+                self.get_compile_args_shape_without_sink(input, tmp_shape)
+                res_shape.append(tmp_shape)
+
+
+    def get_compile_args_shape(self, input_args, is_sink_mode):
+        """get compile args shape"""
+        if is_sink_mode:
+            return input_args
+
+        res_shape = []
+        self.get_compile_args_shape_without_sink(input_args, res_shape)
+        return res_shape
+
+
+    def find_compile_args_in_shape_cache(self, input_args, cache_type):
+        """find compile args in real or generalize shape cache"""
+        shape_cache = self.get_cache_by_type(cache_type)
+        for cache_args in shape_cache:
+            res = self._compare_input_args_and_cache_args(input_args, cache_args)
+            if res:
+                return True
+        return False
+
+
+    def update_phase_and_compile_args(self, compile_args, phase, save_cache_number, is_sink_mode, aux=None):
+        """update compile args and phase"""
+
+        if phase in self.real_phase_and_compile_args_dict:
+            logger.debug(f'phase=%r is in real phase and compile args dict.', phase)
+            return
+
+        if phase in self.generalize_phase_and_compile_args_dict:
+            logger.debug(f'phase=%r is in generalize phase and compile args dict.', phase)
+            return
+
+        if len(self.real_shape_cache) < 2:
+            logger.debug(f'The real shape cache number is {len(self.real_shape_cache)}, is less than 2,'
+                         f'phase=%r should be saved in real shape cache.', phase)
+            self.real_phase_and_compile_args_dict[phase] = compile_args
+            self.real_shape_cache.append(compile_args)
+            return
+
+        max_save_cache_number = save_cache_number - 2
+
+        if len(self.generalize_phase_and_compile_args_dict) >= max_save_cache_number:
+            # step1: find delete phase
+            phase_list = list(self.generalize_phase_and_compile_args_dict.keys())
+            delete_phase = phase_list[0]
+            delete_compile_args = self.generalize_phase_and_compile_args_dict.get(delete_phase)
+
+            # step2: delete phase cache
+            if is_sink_mode:
+                if hasattr(aux, '__network_manage__') and delete_phase in aux.__network_manage__:
+                    del aux.__network_manage__[delete_phase]
+                    del self.generalize_phase_and_compile_args_dict[delete_phase]
+            else:
+                delete_cache = set()
+                delete_cache.add(delete_phase)
+                self._graph_executor.del_net_res(None, delete_cache)
+                del self.generalize_phase_and_compile_args_dict[delete_phase]
+
+            # step3: delete compile args
+            self.generalize_shape_cache.remove(delete_compile_args)
+
+        # step3 save phase and compile args into cache
+        logger.info(f'The generalize shape cache number is {len(self.generalize_shape_cache)}, is less than '
+                    f'{max_save_cache_number}, phase=%r should be saved in generalize shape cache.', phase)
+        self.generalize_phase_and_compile_args_dict[phase] = compile_args
+        self.generalize_shape_cache.append(compile_args)
+
+
+    def check_tuple_of_scalar(self, input):
+        """check tuple of scalar"""
+        for elem in input:
+            if not isinstance(elem, int):
+                return False
+        return True
+
+
+    def _compare_input_args_and_cache_args(self, input_args, cache_args):
+        """compare input args and cache args"""
+        for (input, cache) in zip(input_args, cache_args):
+            if isinstance(input, Tensor) and isinstance(cache, Tensor):
+                if not self.is_tensor_equal(input, cache):
+                    return False
+            elif isinstance(input, int) and isinstance(cache, int):
+                if input != cache:
+                    return False
+            elif isinstance(input, (tuple, list)) and isinstance(cache, (tuple, list)):
+                if not self._compare_input_args_and_cache_args(input, cache):
+                    return False
+        return True
+
+
+class _AutoIdentifyDynamicShape:
+    """
+    Represents a function auto identify dynamic shape.
+    """
+    def __init__(self):
+        self.all_shape_cache = {}
+        self.is_sink_mode = False
+        self.is_enable_auto_dynamic_shape = True
+        self.save_cache_number = 3
+        self.enable_auto_identify = os.getenv('MS_AUTO_DYNAMIC_SHAPE_ENABLE')
+        self.auto_dynamic_shape_manager = _AutoDynamicShapeManager()
+
+
+    def __del__(self):
+        self.all_shape_cache = {}
+        self.is_sink_mode = False
+        self.is_enable_auto_dynamic_shape = True
+        self.save_cache_number = 3
+
+
+    def _check_input_args_number(self, args_list):
+        """check input arg number"""
+        if self.auto_dynamic_shape_manager.get_real_shape_cache_number() > 0:
+            first_real_cache = self.auto_dynamic_shape_manager.get_real_shape_cache()[0]
+            if len(first_real_cache) != len(args_list):
+                return False
+        return True
+
+
+    def _check_input_tensor_type(self, args_list, cache_list):
+        """check input args type"""
+        for (input, cache) in zip(args_list, cache_list):
+            if isinstance(input, Tensor) and isinstance(cache, Tensor):
+                if input.dtype != cache.dtype:
+                    logger.debug((f'input tensor type = {input.dtype}, cache tensor type = {cache.dtype}, '
+                                  f'tensor types are not same.'))
+                    return False
+            elif isinstance(input, (tuple, list)) and isinstance(cache, (tuple, list)):
+                res = self._check_input_tensor_type(input, cache)
+                if not res:
+                    return False
+            elif (isinstance(input, int) and isinstance(cache, int)) or \
+                 (isinstance(input, float) and isinstance(cache, float)):
+                if input != cache:
+                    return False
+            elif (isinstance(input, Tensor) and not isinstance(cache, Tensor)) or \
+                 (isinstance(input, (int, float)) and not isinstance(cache, (int, float))) or \
+                 (isinstance(input, (tuple, list))) and not isinstance(cache, (tuple, list)):
+                return False
+        return True
+
+
+    def _check_input_number_and_type(self, args_list):
+        """check input number and type"""
+        res = self._check_input_args_number(args_list)
+        if not res:
+            return False
+
+        if self.auto_dynamic_shape_manager.get_real_shape_cache_number() > 0:
+            cache_list = self.auto_dynamic_shape_manager.get_real_shape_cache()[0]
+            res = self._check_input_tensor_type(args_list, cache_list)
+            if not res:
+                return False
+        return True
+
+
+    def _is_enable_auto_dynamic_shape(self, args_list, is_sink_mode):
+        """is enable auto identify shape"""
+        if not is_sink_mode and not args_list:
+            return False
+
+        if not self.enable_auto_identify:
+            self.enable_auto_identify = "0"
+
+        if self.enable_auto_identify == "0":
+            return False
+
+        for elem in args_list:
+            if elem is None:
+                continue
+            if not isinstance(elem, (list, tuple, Tensor, int, float)):
+                return False
+            if isinstance(elem, Tensor) and (is_shape_unknown(elem.shape) or (not elem.shape)):
+                return False
+            if not is_sink_mode and isinstance(elem, (list, tuple)):
+                return self._is_enable_auto_dynamic_shape(elem, is_sink_mode)
+        return True
+
+
+    @staticmethod
+    def _do_generalize_in_sink(input, cache, input_index, cache_index, cache_type):
+        """do generalize in sink, input rank must be 2"""
+        if not input:
+            raise ValueError("In sink mode, cell input can not be scalar.")
+
+        if input == cache:
+            return cache
+
+        shape_value = []
+        if len(input) != len(cache):
+            shape_value.append(SHAPE_RANK_ANY)
+        else:
+            for _ in input:
+                shape_value.append(SHAPE_DIM_ANY)
+        logger.info((f'In the {cache_type} cache[{cache_index}], the {input_index}th input tensor shape is {input},'
+                     f'cache shape is {cache}, not equal, need generalize to {shape_value}.'))
+        return shape_value
+
+    def update_phase_and_compile_args(self, args, phase, is_sink_mode, aux=None):
+        """save compile args and phase into dict"""
+        if not self.is_enable_auto_dynamic_shape:
+            return
+        self.auto_dynamic_shape_manager.update_phase_and_compile_args(args, phase, self.save_cache_number,
+                                                                      is_sink_mode, aux)
+
+
+
+    def auto_dynamic_generate_compile_args(self, args_list, is_sink_mode):
+        """generate compile args in auto dynamic shape"""
+        if not self._check_input_number_and_type(args_list) or \
+            not self._is_enable_auto_dynamic_shape(args_list, is_sink_mode):
+            self.is_enable_auto_dynamic_shape = False
+            return args_list
+        self.is_sink_mode = is_sink_mode
+
+        res_shape = self.auto_dynamic_shape_manager.get_compile_args_shape(args_list, is_sink_mode)
+        logger.debug((f'input args list shape = {res_shape}.'))
+
+        # step1: find cache in real_shape_cache.
+        real_cache_number = self.auto_dynamic_shape_manager.get_real_shape_cache_number()
+        if real_cache_number < 2:
+            logger.info((f'real shape cache cap is {real_cache_number}, smaller than 2, '
+                         f'compile args shape={res_shape}.'))
+            return args_list
+
+        is_real_shape_exist = self.auto_dynamic_shape_manager.find_compile_args_in_shape_cache(args_list, "real")
+        if is_real_shape_exist:
+            logger.debug((f'find compile args in real shape cache, compile args shape={res_shape}'))
+            return args_list
+
+        # step2: if can not find cache in real_shape_cache, then generate it
+        generalize_shape_args, can_generalize = self._do_generalize_shape("real", args_list, is_sink_mode)
+        if not can_generalize:
+            return args_list
+
+        if self.auto_dynamic_shape_manager.get_generalize_shape_cache_number() == 0:
+            res_shape = self.auto_dynamic_shape_manager.get_compile_args_shape(generalize_shape_args, is_sink_mode)
+            logger.info((f'generalize shape cache cap is smaller than 1, compile args shape = {res_shape}.'))
+            return generalize_shape_args
+
+        # step3: find generalize_shape in generalize_shape_cache
+        is_generalize_shape_exist = \
+            self.auto_dynamic_shape_manager.find_compile_args_in_shape_cache(generalize_shape_args, "generalize")
+
+        # step 4: if can not find cache in generalize_shape_cache, then generate it again
+        if not is_generalize_shape_exist:
+            new_generalize_shape, can_generalize = self._do_generalize_shape("generalize", generalize_shape_args,
+                                                                             is_sink_mode)
+            if not can_generalize:
+                return args_list
+
+            res_shape = self.auto_dynamic_shape_manager.get_compile_args_shape(new_generalize_shape, is_sink_mode)
+            logger.info((f'generalize with generalize shape cache, compile args shape = {res_shape}'))
+            return new_generalize_shape
+
+        res_shape = self.auto_dynamic_shape_manager.get_compile_args_shape(generalize_shape_args, is_sink_mode)
+        logger.debug((f'find compile args in generalize shape cache, compile args shape={res_shape}'))
+        return generalize_shape_args
+
+
+    def _cal_unknown_shape_count(self, generalize_shape_args, is_sink_mode):
+        """generalize shape by compare with real shape cache, and return the least generalize input."""
+        unknown_shape_count = 0
+        unknown_rank_count = 0
+        for elem in generalize_shape_args:
+            if isinstance(elem, (list, tuple)):
+                rank_count, shape_count = self._cal_unknown_shape_count(elem, is_sink_mode)
+                unknown_rank_count = unknown_rank_count + rank_count
+                unknown_shape_count = unknown_shape_count + shape_count
+            if isinstance(elem, Tensor):
+                if is_shape_unknown(elem.shape):
+                    unknown_shape_count = unknown_shape_count + 1
+                if is_dim_unknown(elem.shape):
+                    unknown_rank_count = unknown_rank_count + 1
+            if is_sink_mode and isinstance(elem, int):
+                if elem == SHAPE_DIM_ANY:
+                    unknown_shape_count = unknown_shape_count + 1
+                if elem == SHAPE_RANK_ANY:
+                    unknown_rank_count = unknown_rank_count + 1
+        return unknown_rank_count, unknown_shape_count
+
+
+    def _do_generalize_one_input_shape(self, input_args, cache_args, cache_type, index, is_sink_mode):
+        """do generalize shape one input by cache"""
+        generalize_one_shape = []
+        for i, (input, cache) in enumerate(zip(input_args, cache_args)):
+            if isinstance(input, Parameter) and isinstance(cache, Parameter):
+                if self.auto_dynamic_shape_manager.is_tensor_equal(input, cache):
+                    generalize_one_shape.append(input)
+                    continue
+                else:
+                    logger.info("In auto dynamic shape mode, parameter must be equal, it can not be generalize.")
+                    return input_args, False
+
+            if isinstance(input, Tensor) and isinstance(cache, Tensor):
+                if self.auto_dynamic_shape_manager.is_tensor_equal(input, cache):
+                    generalize_one_shape.append(input)
+                else:
+                    shape_value = []
+                    if len(input.shape) != len(cache.shape):
+                        shape_value.append(SHAPE_RANK_ANY)
+                    else:
+                        for _ in range(len(input.shape)):
+                            shape_value.append(SHAPE_DIM_ANY)
+                    shape_tuple = tuple(shape_value)
+                    generalize_one_shape.append(Tensor(shape=shape_tuple, dtype=input.dtype))
+                    logger.info((f'In the {cache_type} cache[{index}], the {i}th input tensor shape is {input.shape},'
+                                 f'cache shape is {cache.shape}, not equal, need generalize to {shape_tuple}.'))
+
+            elif isinstance(input, (tuple, list)) and isinstance(cache, (tuple, list)):
+                if is_sink_mode:
+                    # when is_sink_mode=True, input must be the shape of Tensor.
+                    res = self._do_generalize_in_sink(input, cache, i, index, cache_type)
+                    generalize_one_shape.append(res)
+                else:
+                    res = self._do_generalize_one_input_shape(input, cache, cache_type, index, is_sink_mode)
+                    generalize_one_shape.append(res)
+            elif isinstance(input, int) and isinstance(cache, int):
+                # when is_sink_mode=False, the input must may be scalar, or the value of list/tuple.
+                # is_sink_mode can not be True
+                if input == cache:
+                    generalize_one_shape.append(input)
+                else:
+                    logger.info("In auto dynamic shape mode, scalar/tuple/list must be equal, it can not be " \
+                                "generalize.")
+                    return input_args, False
+            elif input is None and cache is None:
+                generalize_one_shape.append(input)
+
+        return generalize_one_shape, True
+
+
+    def _do_generalize_shape(self, cache_type, input_args, is_sink_mode):
+        """do generalize shape by cache"""
+        shape_cache = self.auto_dynamic_shape_manager.get_cache_by_type(cache_type)
+        all_generalize_shape_args = []
+        for index, cache_args in enumerate(shape_cache):
+            generalize_shape, can_generalize = self._do_generalize_one_input_shape(input_args, cache_args, cache_type,
+                                                                                   index, is_sink_mode)
+            if not can_generalize:
+                return generalize_shape, False
+            all_generalize_shape_args.append(tuple(generalize_shape))
+
+        unknown_shape_dict = {}
+        for generalize_shape_args in all_generalize_shape_args:
+            unknown_rank_count, unknown_shape_count = self._cal_unknown_shape_count(generalize_shape_args, is_sink_mode)
+            unknown_count = (unknown_rank_count, unknown_shape_count)
+            if unknown_count not in unknown_shape_dict:
+                unknown_shape_dict[unknown_count] = generalize_shape_args
+
+        keys = list(unknown_shape_dict.keys())
+        keys.sort(key=lambda x: (x[0], x[1]))
+        index = keys[0]
+        return unknown_shape_dict.get(index), True
+
+_auto_dynamic_shape = _AutoIdentifyDynamicShape()
+
+
+def get_auto_dynamic_shape_args(compile_args, key_id):
+    """get auto dynamic shape args."""
+    if key_id not in auto_dynamic_shepe_dict:
+        auto_dynamic_shepe_dict[key_id] = _AutoIdentifyDynamicShape()
+    compile_args = auto_dynamic_shepe_dict[key_id].auto_dynamic_generate_compile_args(compile_args, False)
+    return compile_args
+
+
+def update_auto_dynamic_shape_phase(compile_args, key_id, phase):
+    """update auto dynamic shape phase."""
+    if key_id in auto_dynamic_shepe_dict:
+        auto_dynamic_shepe_dict[key_id].update_phase_and_compile_args(compile_args, phase, False)
+
+
+def get_auto_dynamic_shape_args_with_check_input_signature(compile_args, key_id, input_signature):
+    """get auto dynamic shape args."""
+    if input_signature is None:
+        return get_auto_dynamic_shape_args(compile_args, key_id)
+    return compile_args
+
+
+def update_auto_dynamic_shape_phase_with_check_input_signature(compile_args, key_id, phase, input_signature):
+    """update auto dynamic shape phase."""
+    if input_signature is None:
+        if key_id in auto_dynamic_shepe_dict:
+            auto_dynamic_shepe_dict[key_id].update_phase_and_compile_args(compile_args, phase, False)

mindspore/common/dtype.py

@@ -38,11 +38,12 @@ __dtype__ = [
     "bool_", "float_",
     "list_", "tuple_",
     "int_", "uint",
-    "number", "tensor",
+    "number", "tensor_type",
     "string", "type_none",
-    "tensor_type", "_null",
+    "TensorType", "_null",
     "Type", "Int",
-    "complex64", "complex128"
+    "complex64", "complex128",
+    "bfloat16"
 ]
 
 __method__ = [
@@ -81,6 +82,7 @@ float32 = typing.Float(32)
 single = float32
 float64 = typing.Float(64)
 double = float64
+bfloat16 = typing.BFloat(16)
 complex64 = typing.Complex(64)
 complex128 = typing.Complex(128)
 
@@ -94,7 +96,7 @@ tuple_ = typing.Tuple()
 type_none = typing.TypeNone()
 _null = typing.TypeNull()
 
-tensor = typing.TensorType()
+tensor_type = typing.TensorType()
 index_slices = typing.RowTensorType()
 coo_tensor = typing.COOTensorType()
 csr_tensor = typing.CSRTensorType()
@@ -114,16 +116,16 @@ List = typing.List
 Tuple = typing.Tuple
 Dict = typing.Dict
 Slice = typing.Slice
-function_type = typing.Function
+FunctionType = typing.Function
 Ellipsis_ = typing.TypeEllipsis
 MsClassType = typing.TypeMsClassType
-none_type = typing.TypeNone
-env_type_type = typing.EnvType
-tensor_type = typing.TensorType
-csr_tensor_type = typing.CSRTensorType
-anything_type = typing.TypeAnything
-ref_type = typing.RefType
-_null_type = typing.TypeNull
+NoneType = typing.TypeNone
+EnvType = typing.EnvType
+TensorType = typing.TensorType
+CSRTensorType = typing.CSRTensorType
+AnythingType = typing.TypeAny
+RefType = typing.RefType
+_NullType = typing.TypeNull
 
 number_type = (int8,
                int16,
@@ -136,6 +138,7 @@ number_type = (int8,
                float16,
                float32,
                float64,
+               bfloat16,
                complex64,
                complex128,)
 
@@ -144,9 +147,9 @@ uint_type = (uint8, uint16, uint32, uint64,)
 float_type = (float16, float32, float64,)
 signed_type = (int8, byte, int16, short, int32, intc, int64,
                intp, float16, half, float32, single, float64,
-               double, complex64, complex128)
+               double, bfloat16, complex64, complex128)
 complex_type = (complex64, complex128,)
-all_types = (bool_, int8, uint8, int16, int32, int64, float16, float32, float64, complex64, complex128)
+all_types = (bool_, int8, uint8, int16, int32, int64, float16, float32, float64, bfloat16, complex64, complex128)
 implicit_conversion_seq = {t: idx for idx, t in enumerate(all_types)}
 
 _simple_types = {
@@ -186,6 +189,12 @@ def pytype_to_dtype(obj):
 
     Raises:
         NotImplementedError: If the python type cannot be converted to MindSpore type.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> out = ms.pytype_to_dtype(bool)
+        >>> print(out)
+        Bool
     """
 
     if isinstance(obj, np.dtype):
@@ -209,10 +218,15 @@ def get_py_obj_dtype(obj):
 
     Returns:
         Type of MindSpore type.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> ms.get_py_obj_dtype(1)
+        mindspore.int64
     """
     # Tensor
     if hasattr(obj, 'shape') and hasattr(obj, 'dtype') and isinstance(obj.dtype, typing.Type):
-        return tensor_type(obj.dtype)
+        return TensorType(obj.dtype)
     # Primitive or Cell
     if hasattr(obj, '__primitive_flag__') or hasattr(obj, 'construct'):
         return function
@@ -230,14 +244,19 @@ def get_py_obj_dtype(obj):
 
 
 def dtype_to_nptype(type_):
-    """
+    r"""
     Convert MindSpore dtype to numpy data type.
 
     Args:
-        type_ (:class:`mindspore.dtype`): MindSpore's dtype.
+        type\_ (:class:`mindspore.dtype`): MindSpore's dtype.
 
     Returns:
         The data type of numpy.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> ms.dtype_to_nptype(ms.int8)
+        <class 'numpy.int8'>
     """
 
     return {
@@ -255,18 +274,25 @@ def dtype_to_nptype(type_):
         float64: np.float64,
         complex64: np.complex64,
         complex128: np.complex128,
+        bfloat16: np.float32,
     }[type_]
 
 
 def dtype_to_pytype(type_):
-    """
+    r"""
     Convert MindSpore dtype to python data type.
 
     Args:
-        type_ (:class:`mindspore.dtype`): MindSpore's dtype.
+        type\_ (:class:`mindspore.dtype`): MindSpore's dtype.
 
     Returns:
         Type of python.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> out = ms.dtype_to_pytype(ms.bool_)
+        >>> print(out)
+        <class 'bool'>
     """
 
     return {
@@ -284,6 +310,7 @@ def dtype_to_pytype(type_):
         float16: float,
         float32: float,
         float64: float,
+        bfloat16: float,
         list_: list,
         tuple_: tuple,
         string: str,
@@ -321,12 +348,17 @@ class QuantDtype(enum.Enum):
     """
     An enum for quant datatype, contains `INT1` ~ `INT16`, `UINT1` ~ `UINT16`.
 
-    `QuantDtype` is defined in `mindspore/common/dtype.py`, use command below to import:
+    `QuantDtype` is defined in `dtype.py <https://gitee.com/mindspore/mindspore/tree/r2.2/mindspore/python/mindspore/
+    common/dtype.py>`_, use command below to import:
 
     .. code-block::
 
         from mindspore import QuantDtype
 
+    Tutorial Examples:
+        - `Quantization algorithm in Golden Stick
+          <https://www.mindspore.cn/golden_stick/docs/en/r0.3/quantization/slb.html
+          #applying-the-quantization-algorithm>`_
     """
     INT1 = 0
     INT2 = 1
@@ -367,9 +399,17 @@ class QuantDtype(enum.Enum):
 
     def value(self) -> int:
         """
-        Return value of `QuantDtype`.
+        Return value of `QuantDtype`. This interface is currently used to serialize or deserialize `QuantDtype`
+        primarily.
 
         Returns:
             An int as value of `QuantDtype`.
+
+        Examples:
+            >>> from mindspore import QuantDtype
+            >>> print(QuantDtype.INT8.value())
+            7
+            >>> print(QuantDtype.UINT16.value())
+            115
         """
         return self._value_