mindspore 2.6.0rc1__cp311-cp311-win_amd64.whl → 2.7.0__cp311-cp311-win_amd64.whl

mindspore/safeguard/rewrite_obfuscation.py

@@ -81,7 +81,7 @@ def _transform_target_modules(target_modules):
         obfuscate_layers = target_modules[2].split(':')
         if obfuscate_layers[1] != 'all':
             max_layers = int(obfuscate_layers[1])
-        layers = [i for i in range(0, max_layers)]
+        layers = list(range(0, max_layers))
         path_new = path.replace("blocks", "blocks/${layer}")
         network_obf_template['insert_ops'][0]['input_y'] = "obf_metadata_${layer}"
         weight_obf_template['weight_obf_ops'][0]['input_y'] = "obf_metadata_${layer}"
@@ -95,8 +95,8 @@ def _transform_target_modules(target_modules):
     obf_config['obf_metadata_config'].append(obf_medatadata)
 
     for name in target_list:
-        target_weight = path_new + '/' + name + '/weight'
-        target_bias = path_new + '/' + name + '/bias'
+        target_weight = '/'.join([path_new, name, 'weight'])
+        target_bias = '/'.join([path_new, name, 'bias'])
         weight_obf = weight_obf_template.copy()
         weight_obf['target'] = target_weight
         bias_obf = weight_obf_template.copy()
@@ -185,7 +185,7 @@ def obfuscate_ckpt(network, ckpt_files, target_modules=None, obf_config=None, sa
     def _gen_obf_metadata(config):
         name = config.get('name')
         if name is None:
-            return False
+            return
         save_metadata = config.get('save_metadata', False)
         metadata_op_name = config.get('metadata_op')
         layers = config.get('layers')
@@ -213,7 +213,6 @@ def obfuscate_ckpt(network, ckpt_files, target_modules=None, obf_config=None, sa
                         saved_obf_tensor = metadata_op(saved_obf_tensor)
                     if saved_obf_tensor is not None:
                         saved_metadata[obf_name] = saved_obf_tensor.asnumpy()
-        return True
 
     if not isinstance(network, nn.Cell):
         raise TypeError("network must be nn.Cell, but got {}.".format(type(network)))
@@ -283,13 +282,13 @@ def _obfuscate_single_ckpt(ckpt_name, obf_metadata, obf_config, saved_path):
     def _obfuscate_param(param, obf_metadata, obf_ops, layer=0):
         param_dtype = F.dtype(param)
         obf_param = param
-        for i in range(len(obf_ops)):
-            op_name = obf_ops[i].get('name')
+        for obf_op in obf_ops:
+            op_name = obf_op.get('name')
             if not isinstance(op_name, str):
                 raise TypeError('{} should be str type, but got {}'.format(op_name, type(op_name)))
             if op_name == 'mul':
                 input_x = obf_param
-                input_y_name = _get_op_input_name(obf_ops[i], 'input_y', layer)
+                input_y_name = _get_op_input_name(obf_op, 'input_y', layer)
                 input_y = obf_metadata.get(input_y_name)
                 if input_x is None or input_y is None:
                     log.error("input_x or input_y is None")
@@ -297,22 +296,22 @@ def _obfuscate_single_ckpt(ckpt_name, obf_metadata, obf_config, saved_path):
                 input_y = F.cast(input_y, param_dtype)
                 obf_param = ops.mul(input_x, input_y)
             elif op_name == 'permuate':
-                input_x_name = _get_op_input_name(obf_ops[i], 'input_x', layer)
+                input_x_name = _get_op_input_name(obf_op, 'input_x', layer)
                 p = obf_metadata.get(input_x_name, None)
                 if p is None or obf_param is None:
                     log.error("input_x or param is None")
                     return None
                 obf_param = obf_param[p]
             elif op_name == 'matmul':
-                input_x_name = _get_op_input_name(obf_ops[i], 'input_x', layer)
-                input_y_name = _get_op_input_name(obf_ops[i], 'input_y', layer)
+                input_x_name = _get_op_input_name(obf_op, 'input_x', layer)
+                input_y_name = _get_op_input_name(obf_op, 'input_y', layer)
                 input_x = _get_op_input(input_x_name, obf_param)
                 input_y = _get_op_input(input_y_name, obf_param)
                 if input_x is None or input_y is None:
                     log.error("the input_x or input_y of op: {} is None.".format(op_name))
                     return None
-                input_x = ops.transpose(input_x, (1, 0)) if obf_ops[i].get('transpose_a', False) else input_x
-                input_y = ops.transpose(input_y, (1, 0)) if obf_ops[i].get('transpose_b', False) else input_y
+                input_x = ops.transpose(input_x, (1, 0)) if obf_op.get('transpose_a', False) else input_x
+                input_y = ops.transpose(input_y, (1, 0)) if obf_op.get('transpose_b', False) else input_y
                 obf_param = ops.matmul(F.cast(input_x, param_dtype), F.cast(input_y, param_dtype))
             else:
                 log.error("unsupported op, op must be matmul or permuate or mul, but got {}."

mindspore/train/_utils.py

@@ -26,7 +26,7 @@ import numpy as np
 from mindspore.common.tensor import Tensor
 from mindspore._c_expression import TensorPy as Tensor_
 from mindspore._c_expression import MSContext, ms_ctx_param
-from mindspore.common.dtype import dtype_to_nptype, pytype_to_dtype
+from mindspore.common.dtype import _dtype_to_nptype, _pytype_to_dtype
 from mindspore.common import dtype as mstype
 from mindspore import context
 from mindspore import log as logger
@@ -54,7 +54,7 @@ def _convert_type(types):
     """
     ms_types = []
     for np_type in types:
-        ms_type = pytype_to_dtype(np_type)
+        ms_type = _pytype_to_dtype(np_type)  # pylint:disable=protected-access
         ms_types.append(ms_type)
     return ms_types
 
@@ -131,7 +131,7 @@ def _construct_tensor_list(types, shapes, batch_expand_num=1):
                 new_shape += (item * batch_expand_num,)
             else:
                 new_shape += (item,)
-        tensor = Tensor(np.zeros(new_shape, dtype_to_nptype(type_)), dtype=type_)
+        tensor = Tensor(np.zeros(new_shape, _dtype_to_nptype(type_)), dtype=type_)  # pylint:disable=protected-access
         tensor.virtual_flag = True
         tensor_list.append(tensor)
     return tensor_list
@@ -323,9 +323,15 @@ def parse_strategy_ckpt(file_name):
 def _get_strategy_opt_shard(param_redundancy_dict, parameter_layout_opt_shard):
     """Strategy ckpt append opt shard."""
     for key, value in parameter_layout_opt_shard.items():
-        if value[1] not in (-1, 0):
-            opt_para_num = value[1]
+        if value[1] != 0:
             param_redundancy_ranks = param_redundancy_dict.get(key)
+            if value[1] != -1:
+                opt_para_num = value[1]
+            elif param_redundancy_ranks:
+                opt_para_num = len(param_redundancy_ranks) * len(param_redundancy_ranks[0]) // value[0]
+            else:
+                raise ValueError(f"For get_parameter_redundancy, the format of the parallel communication domain for "
+                                 f"the optimizer is incorrect.")
             res = []
             for param_ranks in param_redundancy_ranks:
                 if len(param_ranks) % opt_para_num == 0:
@@ -576,7 +582,8 @@ def _progress_bar(iterable, total=None):
         print_progress_bar(i)
 
 
-def _load_and_transform(path, name_map, load_func, transform_func):
+def _load_and_transform(path, name_map, load_func, transform_func=None):
+    """use load_func to load and use transform_func to convert"""
     if load_func is not None:
         param_dict = load_func(path)
     else:
@@ -584,5 +591,8 @@ def _load_and_transform(path, name_map, load_func, transform_func):
     transform_dict = {}
     for k, v in param_dict.items():
         new_name = name_map.get(k, k) if name_map is not None else k
-        transform_dict[new_name] = transform_func(v, new_name)
+        if transform_func is not None:
+            transform_dict[new_name] = transform_func(v, new_name)
+        else:
+            transform_dict[new_name] = v
     return transform_dict

mindspore/train/amp.py

@@ -69,6 +69,9 @@ AMP_BLACK_LIST = [
 AMP_AUTO_WHITE_LIST = [
     P.Conv2D,
     P.Conv3D,
+    gen.Conv2DExt,
+    gen.Conv3DExt,
+    gen.ConvTranspose2D,
     P.Conv2DTranspose,
     P.Conv3DTranspose,
     gen.Convolution,
@@ -80,6 +83,10 @@ AMP_AUTO_WHITE_LIST = [
     P.Einsum,
     gen.Dense,
     gen.Addmm,
+    gen.Addbmm,
+    gen.Addmv,
+    gen.Baddbmm,
+    gen.Mv,
 ]
 
 AMP_AUTO_BLACK_LIST = [
@@ -90,8 +97,10 @@ AMP_AUTO_BLACK_LIST = [
     P.Erfinv,
     P.Exp,
     P.Expm1,
-    P.Log,
-    P.Log1p,
+    gen.Log,
+    gen.Log10,
+    gen.Log1p,
+    gen.Log2,
     P.Reciprocal,
     P.Rsqrt,
     P.Sinh,
@@ -103,6 +112,7 @@ AMP_AUTO_BLACK_LIST = [
     P.BatchNorm,
     gen.BatchNormExt,
     gen.GroupNorm,
+    gen.Norm,
     P.KLDivLoss,
     P.SmoothL1Loss,
     P.MultilabelMarginLoss,
@@ -113,7 +123,19 @@ AMP_AUTO_BLACK_LIST = [
     P.Pdist,
     P.Cdist,
     P.Renorm,
+    gen.ReduceProd,
+    gen.Softmax,
+    gen.LogSoftmax,
+    gen.LogSoftmaxExt,
+    gen.CumProd,
+    gen.CumSum,
+    gen.CumsumExt,
+    gen.ProdExt,
+    gen.SumExt,
+    gen.L1LossExt,
     gen.MSELossExt,
+    gen.NLLLoss,
+    gen.NLLLoss2d,
 ]
 
 # Indicates which inputs of primitives need to be converted
@@ -358,7 +380,7 @@ def _auto_black_list(network, black_list, dtype):
     return network
 
 
-class amp_decorator:
+class AmpDecorator:
     """
     Auto mixed precision decorator.
     Type of lists: List[Tuple[str, List[int]]]
@@ -384,7 +406,7 @@ def _set_amp_decorator(obj, amp_level, amp_dtype, white_list, black_list):
     if inspect.isfunction(obj) or inspect.ismethod(obj):
         @functools.wraps(obj)
         def wrapper(*args, **kwargs):
-            with amp_decorator(amp_level, amp_dtype, white_list, black_list):
+            with AmpDecorator(amp_level, amp_dtype, white_list, black_list):
                 return obj(*args, **kwargs)
         return wrapper
     if isinstance(obj, nn.Cell):
@@ -423,26 +445,24 @@ def auto_mixed_precision(network, amp_level="O0", dtype=mstype.float16):
 
     Operators in `auto_whitelist` are:
 
-    ``Conv2D``, ``Conv3D``, ``Conv2DTranspose``, ``Conv3DTranspose``, ``Convolution``, ``MatMul``, ``MatMulExt``,
-    ``BatchMatMul``, ``BatchMatMulExt``, ``PReLU``, ``Einsum``, ``Dense``, ``Addmm``
+    ``Conv2D``, ``Conv2DExt``, ``Conv3D``, ``Conv3DExt``, ``Conv2DTranspose``, ``ConvTranspose2D``,
+    ``Conv3DTranspose``, ``Convolution``, ``MatMul``, ``MatMulExt``, ``BatchMatMul``, ``BatchMatMulExt``, ``PReLU``,
+    ``Einsum``, ``Dense``, ``Addmm``, ``Addbmm``, ``Addmv``, ``Baddbmm``, ``Mv``
 
     Operators in `auto_blacklist` are:
 
-    ``Pow``, ``ACos``, ``Asin``, ``Cosh``, ``Erfinv``, ``Exp``, ``Expm1``, ``Log``, ``Log1p``, ``Reciprocal``,
-    ``Rsqrt``, ``Sinh``, ``Tan``, ``Softplus``, ``SoftplusExt``, ``LayerNorm``, ``LayerNormExt``, ``BatchNorm``,
-    ``BatchNormExt``, ``GroupNorm``, ``KLDivLoss``, ``SmoothL1Loss``, ``MultilabelMarginLoss``, ``SoftMarginLoss``,
-    ``TripletMarginLoss``, ``MultiMarginLoss``, ``BCEWithLogitsLoss``, ``Pdist``, ``Cdist``, ``Renorm``,
-    ``ReduceProd``, ``Softmax``, ``LogSoftmax``, ``CumProd``, ``CumSum``, ``CumsumExt``, ``ProdExt``, ``SumExt``,
-    ``Norm``, ``MSELossExt``
+    ``Pow``, ``ACos``, ``Asin``, ``Cosh``, ``Erfinv``, ``Exp``, ``Expm1``, ``Log``, ``Log10``, ``Log1p``, ``Log2``,
+    ``Reciprocal``, ``Rsqrt``, ``Sinh``, ``Tan``, ``Softplus``, ``SoftplusExt``, ``LayerNorm``, ``LayerNormExt``,
+    ``BatchNorm``, ``BatchNormExt``, ``GroupNorm``, ``KLDivLoss``, ``SmoothL1Loss``, ``MultilabelMarginLoss``,
+    ``SoftMarginLoss``, ``TripletMarginLoss``, ``MultiMarginLoss``, ``BCEWithLogitsLoss``, ``Pdist``, ``Cdist``,
+    ``Renorm``, ``ReduceProd``, ``Softmax``, ``LogSoftmax``, ``LogSoftmaxExt``, ``CumProd``, ``CumSum``,
+    ``CumsumExt``, ``ProdExt``, ``SumExt``, ``Norm``, ``L1LossExt``, ``MSELossExt``, ``NLLLoss``, ``NLLLoss2d``
 
     Operators in `promote_list` are:
 
     ``Addcdiv``, ``Addcmul``, ``Cross``, ``_PyboostCrossPrim``, ``Dot``, ``GridSampler2D``, ``GridSampler3D``,
     ``BiasAdd``, ``AddN``, ``Concat``
 
-    For details on automatic mixed precision, refer to
-    `Automatic Mix Precision <https://www.mindspore.cn/tutorials/en/master/beginner/mixed_precision.html>`_ .
-
     Note:
         - Repeatedly calling mixed-precision interfaces, such as `custom_mixed_precision` and `auto_mixed_precision`,
           can result in a larger network hierarchy and slower performance.
@@ -638,7 +658,7 @@ def _add_loss_network(network, loss_fn, cast_model_type):
 
 
 def _is_grad_accumulation(mcell):
-    if mcell.cls_name == "GradAccumulationCell" or mcell.cls_name == "GradAccumulation":
+    if mcell.cls_name in {"GradAccumulationCell", "GradAccumulation"}:
         return True
     for cell in mcell.cells():
         if _is_grad_accumulation(cell):
@@ -675,23 +695,23 @@ def build_train_network(network, optimizer, loss_fn=None, level='O0', boost_leve
     Build the mixed precision training cell automatically.
 
     Note:
-        - After using `custom_mixed_precision` or `auto_mixed_precision` for precision conversion, it is not supported
-          to perform the precision conversion again. If  `build_train_network` is used to train a converted network,
-          `level` need to be configured to ``O0`` to avoid the duplicated accuracy conversion.
+        After using `custom_mixed_precision` or `auto_mixed_precision` for precision conversion, it is not supported
+        to perform the precision conversion again. If  `build_train_network` is used to train a converted network,
+        `level` need to be configured to ``O0`` to avoid the duplicated accuracy conversion.
 
     Args:
         network (Cell): Definition of the network.
         optimizer (:class:`mindspore.nn.Optimizer`): Define the optimizer to update the Parameter.
-        loss_fn (Union[None, Cell]): Define the loss function. If None, the `network` should have the loss inside.
-            Default: ``None`` .
-        level (str): Supports ['O0', 'O1', 'O2', 'O3', 'auto']. Default: ``'O0'`` .
+        loss_fn (Union[None, Cell], optional): Define the loss function. If None,
+            the `network` should have the loss inside. Default: ``None`` .
+        level (str, optional): Supports ['O0', 'O1', 'O2', 'O3', 'auto']. Default: ``'O0'`` .
 
             For details on amp level, refer to :func:`mindspore.amp.auto_mixed_precision`.
 
             Property of `keep_batchnorm_fp32`, `cast_model_type` and `loss_scale_manager` determined by `level`
             setting may be overwritten by settings in `kwargs`.
 
-        boost_level (str): Option for argument `level` in `mindspore.boost` , level for boost mode
+        boost_level (str, optional): Option for argument `level` in `mindspore.boost` , level for boost mode
             training. Supports ['O0', 'O1', 'O2']. Default: ``'O0'`` .
 
             - 'O0': Do not change.

mindspore/train/callback/__init__.py

@@ -15,6 +15,11 @@
 """Callback related classes and functions."""
 from __future__ import absolute_import
 
+__all__ = ["Callback", "LossMonitor", "TimeMonitor", "ModelCheckpoint", "FlopsUtilizationCollector",
+           "SummaryCollector", "CheckpointConfig", "RunContext", "LearningRateScheduler", "SummaryLandscape",
+           "History", "LambdaCallback", "ReduceLROnPlateau", "EarlyStopping", "OnRequestExit", "BackupAndRestore",
+           "TrainFaultTolerance"]
+
 from mindspore.train.callback._callback import Callback
 from mindspore.train.callback._callback import CallbackManager as _CallbackManager
 from mindspore.train.callback._callback import InternalCallbackParam as _InternalCallbackParam
@@ -37,8 +42,3 @@ from mindspore.train.callback._on_request_exit import OnRequestExit
 from mindspore.train.callback._backup_and_restore import BackupAndRestore
 from mindspore.train.callback._flops_collector import FlopsUtilizationCollector
 from mindspore.train.callback._train_fault_tolerance import TrainFaultTolerance
-
-__all__ = ["Callback", "LossMonitor", "TimeMonitor", "ModelCheckpoint", "FlopsUtilizationCollector",
-           "SummaryCollector", "CheckpointConfig", "RunContext", "LearningRateScheduler", "SummaryLandscape",
-           "History", "LambdaCallback", "ReduceLROnPlateau", "EarlyStopping", "OnRequestExit", "BackupAndRestore",
-           "TrainFaultTolerance"]

mindspore/train/callback/_callback.py

@@ -60,7 +60,8 @@ def _fill_param_into_net(net, parameter_list):
         if np_val.shape == (1,):
             parameter_dict[param_name] = Parameter(np_val, name=param_name)
         elif np_val.shape == ():
-            parameter_dict[param_name] = Parameter(Tensor(np_val.tolist(), mstype.pytype_to_dtype(np_val.dtype)),
+            # pylint:disable=protected-access
+            parameter_dict[param_name] = Parameter(Tensor(np_val.tolist(), mstype._pytype_to_dtype(np_val.dtype)),
                                                    name=param_name)
         else:
             parameter_dict[param_name] = Parameter(Tensor(np_val), name=param_name)

mindspore/train/callback/_checkpoint.py

@@ -411,8 +411,6 @@ class CheckpointConfig:
             handle_append_info["epoch_num"] = 0
         if "step_num" in append_info:
             handle_append_info["step_num"] = 0
-        if "random_op" in append_info:
-            handle_append_info["random_op"] = 0
         dict_num = 0
         for element in append_info:
             if not isinstance(element, str) and not isinstance(element, dict):
@@ -588,8 +586,6 @@ class ModelCheckpoint(Callback):
         # save graph (only once)
         if not self._graph_saved:
             graph_file_name = os.path.join(self._directory, self._prefix + '-graph.meta')
-            if os.path.isfile(graph_file_name) and context.get_context("mode") == context.GRAPH_MODE:
-                os.remove(graph_file_name)
             _save_graph(cb_params.train_network, graph_file_name)
             self._graph_saved = True
         self._save_ckpt(cb_params)
@@ -632,13 +628,6 @@ class ModelCheckpoint(Callback):
         if "step_num" in self._append_dict:
             self._append_dict["step_num"] = self._append_step_num + step_num
 
-    def _update_save_step(self, cb_params):
-        """update step if used async d2h copy"""
-        step_num_in_epoch = int((cb_params.cur_step_num - 1) % cb_params.batch_num + 1)
-        if self._d2h_async and self._run_mode == context.GRAPH_MODE:
-            step_num_in_epoch -= 1
-        return step_num_in_epoch
-
     def _save_ckpt(self, cb_params, force_to_save=False):
         """Save checkpoint files."""
         if cb_params.cur_step_num == self._last_triggered_step:
@@ -649,7 +638,7 @@ class ModelCheckpoint(Callback):
             self._flush_from_cache(cb_params)
 
         save_ckpt = self._check_save_ckpt(cb_params, force_to_save)
-        step_num_in_epoch = self._update_save_step(cb_params)
+        step_num_in_epoch = int((cb_params.cur_step_num - 1) % cb_params.batch_num + 1)
 
         if save_ckpt:
 
@@ -713,12 +702,13 @@ class ModelCheckpoint(Callback):
                 save_checkpoint(network, cur_file, False, self._config.async_save,
                                 self._append_dict, self._config.enc_key, self._config.enc_mode,
                                 crc_check=self._config.crc_check, format=self._config.format,
-                                incremental=self._map_param_inc, choice_func=choice_func)
+                                incremental=self._map_param_inc, choice_func=choice_func,
+                                remove_redundancy=self._config.remove_redundancy)
             else:
                 save_checkpoint(network, cur_file, self._config.integrated_save, self._config.async_save,
                                 self._append_dict, self._config.enc_key, self._config.enc_mode,
                                 crc_check=self._config.crc_check, format=self._config.format,
-                                incremental=self._map_param_inc)
+                                incremental=self._map_param_inc, remove_redundancy=self._config.remove_redundancy)
 
             self._latest_ckpt_file_name = cur_file
 

mindspore/train/callback/_flops_collector.py

@@ -31,7 +31,6 @@ from mindspore.communication.management import (create_group, get_group_size,
 from mindspore.parallel._auto_parallel_context import auto_parallel_context
 from mindspore.ops import operations as P
 from mindspore.common import Tensor
-from mindspore import context
 import mindspore.nn as nn
 
 
@@ -53,7 +52,7 @@ class FlopsUtilizationCollector(Callback):
     The FlopsUtilizationCollector interface counts the model utilization information MFU
     and the hardware utilization information HFU.
     Currently, the API counts only the forward and backward flops of MatMul,
-    BatchMatMul, FlashAttentionScore, and Conv2D operators.
+    BatchMatMul, flash_attention_score, and Conv2D operators.
     Only used in graph mode with static shape.
 
     Args:
@@ -152,16 +151,21 @@ class FlopsUtilizationCollector(Callback):
         """
         Check whether FlopsUtilizationCollector is working in the current environment
         """
-        if context.get_context("mode") != context.GRAPH_MODE:
-            if self.verbose:
-                raise ValueError("FlopsUtilizationCollector now only support graph mode.")
-            logger.info("FlopsUtilizationCollector now only support graph mode.")
-            return False
         cb_params = run_context.original_args()
         if cb_params.mode == 'train':
             network = cb_params.train_network
+            if not network.compiled:
+                if self.verbose:
+                    raise ValueError("FlopsUtilizationCollector now only support graph mode.")
+                logger.info("FlopsUtilizationCollector now only support graph mode.")
+                return False
         elif cb_params.mode == 'eval':
             network = cb_params.eval_network
+            if not network.compiled:
+                if self.verbose:
+                    raise ValueError("FlopsUtilizationCollector now only support graph mode.")
+                logger.info("FlopsUtilizationCollector now only support graph mode.")
+                return False
         else:
             if self.verbose:
                 raise ValueError('FlopsUtilizationCollector only support train and eval mode!')

mindspore/train/callback/_landscape.py

@@ -404,7 +404,6 @@ class SummaryLandscape:
     def _set_context(device_id):
         """Set context."""
         context.set_context(device_id=device_id)
-        context.set_context(mode=context.GRAPH_MODE)
 
     def _create_landscape_by_pca(self, epochs, proz, landscape_size, device_ids=None, callback_fn=None, executor=None):
         """Create landscape by PCA."""