mindspore 2.1.0__cp38-none-any.whl → 2.2.10__cp38-none-any.whl

mindspore/boost/boost.py

@@ -85,9 +85,6 @@ class AutoBoost:
                         "freeze_p": 0.7,
                         "total_steps": 65536
                     }
-                    "grad_accumulation": {
-                        "grad_accumulation_step": 1
-                    },
                     "dim_reduce": {
                         "rho": 0.55,
                         "gamma": 0.9,
@@ -139,10 +136,6 @@ class AutoBoost:
               - freeze_p (float): Gradient freezing probability. Default: ``0.7`` .
               - total_steps (int): Total training steps. Default: ``65536`` .
 
-            - grad_accumulation:
-
-              - grad_accumulation_step (int): Steps to accumulate gradients. Default: ``1`` .
-
             - dim_reduce:
 
               The leading principles of dim_reduce:
@@ -204,7 +197,7 @@ class AutoBoost:
         >>> import json
         >>> boost_json = "/path/boost_config.json"
         >>> with open(boost_json, 'r') as fp:
-        >>>     boost_config_dict = json.load(fp)
+        ...     boost_config_dict = json.load(fp)
         >>> boost = AutoBoost("O1", boost_config_dict)
     """
     _instance_lock = threading.Lock()

mindspore/boost/boost_cell_wrapper.py

@@ -136,7 +136,7 @@ class BoostTrainOneStepCell(TrainOneStepCell):
         >>> from mindspore import boost
         >>> from mindspore import nn
         >>> # Define the network structure of LeNet5. Refer to
-        >>> # https://gitee.com/mindspore/docs/blob/r2.1/docs/mindspore/code/lenet.py
+        >>> # https://gitee.com/mindspore/docs/blob/r2.2/docs/mindspore/code/lenet.py
         >>> net = LeNet5()
         >>> loss_fn = nn.SoftmaxCrossEntropyWithLogits()
         >>> optim = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
@@ -412,7 +412,8 @@ class BoostTrainOneStepWithLossScaleCell(BoostTrainOneStepCell):
         network (Cell): The training network. The network only supports single output.
         optimizer (Cell): Optimizer for updating the weights.
         scale_sense (Union[Tensor, Cell]): If this value is Cell type, the loss scaling update logic cell.If this value
-                                          is Tensor type, Tensor with shape :math:`()` or :math:`(1,)`.
+            is Tensor type, :func:`mindspore.nn.TrainOneStepWithLossScaleCell.set_sense_scale` can be called to update
+            loss scale factor, Tensor with shape :math:`()` or :math:`(1,)`.
 
     Inputs:
         - **\*inputs** (Tuple(Tensor)) - Tuple of input tensors with shape :math:`(N, \ldots)`.

mindspore/boost/grad_accumulation.py

@@ -71,7 +71,7 @@ class GradientAccumulation(Cell):
 
         if self._accumulation_step >= self._max_accumulation_step:
             loss = F.depend(loss, self.optimizer(self._grad_accumulation))
-            self._accumulation_step = 0
+            F.assign(self._accumulation_step, 0)
 
         if self._accumulation_step == 0:
             loss = F.depend(loss, self.hyper_map(F.partial(gradient_clear_op), self._grad_accumulation))

mindspore/boost/group_loss_scale_manager.py

@@ -27,7 +27,7 @@ __all__ = ["GroupLossScaleManager"]
 
 
 class GroupLossScaleManager(Cell):
-    """
+    r"""
     Enhanced hybrid precision algorithm supports multi-layer application of different loss scales and
     dynamic updating of loss scales.
 
@@ -41,7 +41,8 @@ class GroupLossScaleManager(Cell):
         - **layer2** (Int) - Last network layer value.
 
     Outputs:
-        - **x** (Tensor) - The output of `_DynamicLossScale` operator.
+        - **out** (Tensor) - A tensor with a group of loss scale tags that marks
+          the loss scale group number of the current tensor.
 
     Supported Platforms:
         ``Ascend``
@@ -80,10 +81,10 @@ class GroupLossScaleManager(Cell):
         >>> param_group1 = []
         >>> param_group2 = []
         >>> for param in net.trainable_params():
-        >>>     if 'conv' in param.name:
-        >>>         param_group1.append(param)
-        >>>     else:
-        >>>         param_group2.append(param)
+        ...     if 'conv' in param.name:
+        ...         param_group1.append(param)
+        ...     else:
+        ...         param_group2.append(param)
         >>> loss_scale_manager.loss_scale_groups = [param_group1, param_group2]
         >>> loss = nn.SoftmaxCrossEntropyWithLogits()
         >>> optim = nn.Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9)
@@ -93,7 +94,7 @@ class GroupLossScaleManager(Cell):
         ...                        loss_scale_manager=loss_scale_manager,
         ...                        boost_level="O1", boost_config_dict=boost_config_dict)
         >>> # Create the dataset taking MNIST as an example. Refer to
-        >>> # https://gitee.com/mindspore/docs/blob/r2.1/docs/mindspore/code/mnist.py
+        >>> # https://gitee.com/mindspore/docs/blob/r2.2/docs/mindspore/code/mnist.py
         >>> dataset = create_dataset()
         >>> model.train(2, dataset)
     """

mindspore/common/__init__.py

@@ -18,7 +18,7 @@ from mindspore.common import dtype
 from mindspore.common.api import ms_function, ms_memory_recycle, ms_class, jit, jit_class, _no_grad
 from mindspore.common.dtype import Type, int8, byte, int16, short, int32, intc, int64, intp, \
     uint8, ubyte, uint16, ushort, uint32, uintc, uint64, uintp, float16, half, \
-    float32, single, float64, double, bool_, float_, list_, tuple_, int_, \
+    float32, single, float64, bfloat16, double, bool_, float_, list_, tuple_, int_, \
     uint, number, tensor_type, string, type_none, TensorType, Int, \
     complex64, complex128, dtype_to_nptype, _null, _NullType, \
     dtype_to_pytype, pytype_to_dtype, get_py_obj_dtype, QuantDtype
@@ -30,6 +30,7 @@ from mindspore.common.sparse_tensor import RowTensor, RowTensorInner, SparseTens
 from mindspore.common.mutable import mutable
 from mindspore.common.jit_config import JitConfig
 from mindspore.common.lazy_inline import lazy_inline
+from mindspore.common.mindir_util import load_mindir, save_mindir
 
 # symbols from dtype
 __all__ = [
@@ -55,7 +56,8 @@ __all__ = [
     "complex64", "complex128",
     # __method__ from dtype
     "dtype_to_nptype", "dtype_to_pytype",
-    "pytype_to_dtype", "get_py_obj_dtype"
+    "pytype_to_dtype", "get_py_obj_dtype",
+    "bfloat16",
 ]
 
 __all__.extend([
@@ -67,5 +69,5 @@ __all__.extend([
     "set_dump",
     "ms_memory_recycle",
     "mutable", "JitConfig",
-    "lazy_inline"
+    "lazy_inline", "load_mindir", "save_mindir"
 ])

mindspore/common/_jit_fallback_utils.py

@@ -103,3 +103,9 @@ def list_inplace_clear(list_obj):
         list_obj = []
     list_obj.clear()
     return list_obj
+
+
+def dict_inplace_setitem(dict_obj, key, target):
+    """Inplace dictionary setitem operation for dict_obj."""
+    dict_obj[key] = target
+    return dict_obj

mindspore/common/_register_for_adapter.py

@@ -20,6 +20,7 @@ class Registry:
     """Registry class for ms adapter."""
 
     def __init__(self):
+        self.is_registered = False
         self._tensor = None
         self._parameter = None
         self._convert_map = {}
@@ -48,6 +49,7 @@ class Registry:
         if self._tensor is not None:
             raise ValueError("Repeated registration of tensor in ms adapter config.")
         self._tensor = value
+        self.is_registered = True
 
     def register_parameter(self, value):
         """Register the parameter of ms adapter."""

mindspore/common/_register_for_tensor.py

@@ -17,7 +17,7 @@
 
 from __future__ import absolute_import
 from collections import UserDict
-from mindspore import context
+from mindspore._c_expression import Tensor as Tensor_
 
 
 class Registry(UserDict):
@@ -31,7 +31,7 @@ class Registry(UserDict):
         """Get the value by str."""
         if not isinstance(obj_str, str):
             raise TypeError("key for tensor registry must be string.")
-        if context.get_context("enable_ge"):
+        if Tensor_._is_test_stub() is True:  # pylint: disable=W0212
             def wrap(*args):
                 new_args = list(args)
                 new_args.append(obj_str)

mindspore/common/_stub_tensor.py

@@ -144,6 +144,8 @@ class StubTensor:
     slice_num_of_persistent_data = _stub_method(Tensor.slice_num_of_persistent_data)
     slice_shape_of_persistent_data = _stub_method(Tensor.slice_shape_of_persistent_data)
     flush_from_cache = _stub_method(Tensor.flush_from_cache)
+    contiguous = _stub_method(Tensor.contiguous)
+    is_contiguous = _stub_method(Tensor.is_contiguous)
 
     def stub_sync(self):
         """sync real tensor."""
@@ -156,6 +158,17 @@ class StubTensor:
             self.stub = None
         return self.tensor
 
+    def __getstate__(self):
+        state = {}
+        value = self.stub.get_value() if self.stub else self.tensor.__getstate__()
+        state["value"] = value
+        return state
+
+    def __setstate__(self, state):
+        value = state.pop("value")
+        self.stub = None
+        self.tensor = Tensor(value, internal=True)
+
 
 def _init_stub_tensor_api():
     """adapt to python tensor and cpp tensor api"""

mindspore/common/_utils.py

@@ -18,9 +18,11 @@
 
 import os
 import math
+import ctypes
 import functools
 
 import mindspore
+from mindspore import log as logger
 from mindspore.common import dtype as mstype
 from mindspore.parallel._ps_context import _is_ps_mode, _is_role_pserver, _is_role_sched
 
@@ -103,3 +105,14 @@ def ones_like(x):
 def zeros_like(x):
     """Implement `zeroslike`."""
     return mindspore.ops.composite.zeros_like(x)
+
+
+def load_lib(lib_path):
+    """load specified library."""
+    try:
+        ctypes.CDLL(lib_path)
+    # pylint: disable=broad-except
+    except Exception:
+        logger.warning(f'Loading {lib_path} lib error.')
+        return False
+    return True