mindspore 2.1.0__cp38-cp38-manylinux1_x86_64.whl → 2.2.0__cp38-cp38-manylinux1_x86_64.whl

mindspore/dataset/vision/utils.py

@@ -303,20 +303,19 @@ class ImageReadMode(IntEnum):
 
 class Inter(IntEnum):
     """
-    Interpolation Modes.
-
-    Possible enumeration values are: ``Inter.NEAREST``, ``Inter.ANTIALIAS``, ``Inter.LINEAR``, ``Inter.BILINEAR``,
-    ``Inter.CUBIC``, ``Inter.BICUBIC``, ``Inter.AREA``, ``Inter.PILCUBIC``.
-
-    - ``Inter.NEAREST`` : means interpolation method is nearest-neighbor interpolation.
-    - ``Inter.ANTIALIAS`` : means the interpolation method is antialias interpolation.
-    - ``Inter.LINEAR`` : means interpolation method is bilinear interpolation, here is the same as ``Inter.BILINEAR``.
-    - ``Inter.BILINEA`` : means interpolation method is bilinear interpolation.
-    - ``Inter.CUBIC`` : means the interpolation method is bicubic interpolation, here is the same as ``Inter.BICUBIC``.
-    - ``Inter.BICUBIC`` : means the interpolation method is bicubic interpolation.
-    - ``Inter.AREA`` : means interpolation method is pixel area interpolation.
-    - ``Inter.PILCUBIC`` : means interpolation method is bicubic interpolation like implemented in pillow, input
-      should be in 3 channels format.
+    Interpolation methods.
+
+    Available values are as follows:
+
+    - ``Inter.NEAREST`` : Nearest neighbor interpolation.
+    - ``Inter.ANTIALIAS`` : Antialias interpolation. Supported only when the input is PIL.Image.Image.
+    - ``Inter.LINEAR`` : Linear interpolation, the same as ``Inter.BILINEAR``.
+    - ``Inter.BILINEAR`` : Bilinear interpolation.
+    - ``Inter.CUBIC`` : Cubic interpolation, the same as ``Inter.BICUBIC``.
+    - ``Inter.BICUBIC`` : Bicubic interpolation.
+    - ``Inter.AREA`` : Pixel area interpolation. Supported only when the input is numpy.ndarray.
+    - ``Inter.PILCUBIC`` : Pillow implementation of bicubic interpolation. Supported only when the input
+      is numpy.ndarray.
     """
     NEAREST = 0
     ANTIALIAS = 1

mindspore/dataset/vision/validators.py

@@ -24,7 +24,7 @@ from mindspore.dataset.core.validator_helpers import check_value, check_uint8, F
     check_pos_float32, check_float32, check_2tuple, check_range, check_positive, INT32_MAX, INT32_MIN, \
     parse_user_args, type_check, type_check_list, check_c_tensor_op, UINT8_MAX, UINT8_MIN, check_value_normalize_std, \
     check_value_cutoff, check_value_ratio, check_odd, check_non_negative_float32, check_non_negative_int32, \
-    check_pos_int32, check_int32, check_tensor_op, deprecator_factory
+    check_pos_int32, check_int32, check_tensor_op, deprecator_factory, check_valid_str
 from mindspore.dataset.transforms.validators import check_transform_op_type
 from .utils import Inter, Border, ImageBatchFormat, ConvertMode, SliceMode, AutoAugmentPolicy
 
@@ -339,6 +339,16 @@ def check_resize_interpolation(method):
 
     return new_method
 
+def check_device_target(method):
+    """A wrapper that wraps a parameter checker"""
+
+    @wraps(method)
+    def new_method(self, *args, **kwargs):
+        [device_target], _ = parse_user_args(method, *args, **kwargs)
+        check_valid_str(device_target, ["CPU", "Ascend"], "device_target")
+        return method(self, *args, **kwargs)
+    return new_method
+
 
 def check_resized_crop(method):
     """A wrapper that wraps a parameter checker around the original function(ResizedCrop operation)."""

mindspore/experimental/map_parameter.py

@@ -265,6 +265,20 @@ class MapParameter(Parameter):
         """
         return self._map_tensor.export_data(incremental)
 
+    def export_bytes(self, incremental=False):
+        """
+        Export bytes from this map parameter.
+
+        Args:
+            incremental (bool): False for full export, otherwise for incremental export. Default: False.
+            When exporting data incrementally, the value_array does not contain unchanged data. The length
+            of the key_array and the length of the status_array are consistent.
+
+        Returns:
+            Tuple(bytes, bytes, bytes), The exported bytes as a tuple.
+        """
+        return self._map_tensor.export_bytes(incremental)
+
     def import_data(self, data):
         """
         Import this map parameter from exported data.

mindspore/{nn/optim_ex → experimental/optim}/__init__.py

@@ -1,29 +1,30 @@
-# Copyright 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.
-# ============================================================================
-"""
-Optimizer.
-
-Provide common optimizers for training, such as SGD, ADAM, Momentum.
-The optimizer is used to calculate and update the gradients.
-"""
-from __future__ import absolute_import
-
-from mindspore.nn.optim_ex.optimizer import Optimizer
-from mindspore.nn.optim_ex.adamw import AdamW
-from mindspore.nn.optim_ex.sgd import SGD
-from mindspore.nn.optim_ex.adam import Adam
-
-
-__all__ = ['Optimizer', 'AdamW', 'SGD', 'Adam']
+# Copyright 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.
+# ============================================================================
+"""
+Optimizer.
+
+Provide common optimizers for training, such as SGD, ADAM, Momentum.
+The optimizer is used to calculate and update the gradients.
+"""
+from __future__ import absolute_import
+
+from mindspore.experimental.optim.optimizer import Optimizer
+from mindspore.experimental.optim.adamw import AdamW
+from mindspore.experimental.optim.sgd import SGD
+from mindspore.experimental.optim.adam import Adam
+from mindspore.experimental.optim import lr_scheduler
+
+
+__all__ = ['Optimizer', 'AdamW', 'SGD', 'Adam']

mindspore/{nn/optim_ex → experimental/optim}/adam.py

@@ -15,12 +15,30 @@
 """adam"""
 from __future__ import absolute_import
 
-from mindspore.ops import functional as F, operations as P
-from mindspore.common.parameter import Parameter, ParameterTuple
+from mindspore.ops import functional as F, composite as C, operations as P
+from mindspore.common.parameter import Parameter
 from mindspore.common.tensor import Tensor
 import mindspore.common.dtype as mstype
-from mindspore.nn.optim_ex.optimizer import Optimizer
-from mindspore import ops
+from mindspore.experimental.optim.optimizer import Optimizer
+
+_adam_opt = C.MultitypeFuncGraph("adam_opt")
+
+
+@_adam_opt.register("Function", "Tensor", "Tensor", "Float", "Float", "Float", "Tensor",
+                    "Tensor", "Tensor", "Tensor", "Tensor")
+def _run_adam_opt(opt, beta1_power, beta2_power, beta1, beta2, eps, lr, gradient, param, moment1, moment2):
+    """Apply adam optimizer to the weight parameter."""
+    success = True
+    success = F.depend(success, opt(param, moment1, moment2, beta1_power, beta2_power, lr, beta1, beta2, eps, gradient))
+    return success
+
+
+@_adam_opt.register("Function", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor")
+def _run_adam_with_amsgrad_opt(opt, beta1_power, beta2_power, lr, gradient, param, moment1, moment2, vhat):
+    """Apply adam optimizer to the weight parameter with amsgrad."""
+    success = True
+    success = F.depend(success, opt(param, moment1, moment2, vhat, beta1_power, beta2_power, lr, gradient))
+    return success
 
 
 class Adam(Optimizer):
@@ -62,7 +80,7 @@ class Adam(Optimizer):
     .. warning::
         This is an experimental optimizer API that is subject to change.
         This module must be used with lr scheduler module in `LRScheduler Class
-        <https://www.mindspore.cn/docs/en/r2.1/api_python/mindspore.nn.html#lrscheduler>`_ .
+        <https://www.mindspore.cn/docs/en/r2.2/api_python/mindspore.nn.html#learningrateschedule-class>`_ .
 
     Args:
         params (Union[list(Parameter), list(dict)]): list of parameters to optimize or dicts defining
@@ -95,11 +113,12 @@ class Adam(Optimizer):
     Examples:
         >>> import mindspore
         >>> from mindspore import nn
+        >>> from mindspore.experimental import optim
         >>> # 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(sparse=True)
-        >>> optimizer = nn.optim_ex.Adam(net.trainable_params(), lr=0.1)
+        >>> optimizer = optim.Adam(net.trainable_params(), lr=0.1)
         >>> def forward_fn(data, label):
         ...     logits = net(data)
         ...     loss = loss_fn(logits, label)
@@ -131,69 +150,43 @@ class Adam(Optimizer):
         self.exp_avg = self.parameters.clone(prefix="exp_avg", init='zeros')
         self.exp_avg_sq = self.parameters.clone(prefix="exp_avg_sq", init='zeros')
         self.max_exp_avg_sq = self.parameters.clone(prefix="max_exp_avg_sq", init='zeros')
-        self.state_step = ParameterTuple(Parameter(
-            Tensor(0, mstype.int32), "step_"+str(i)) for i in range(len(self.parameters)))
+        self.state_step = Parameter(Tensor(0, mstype.int32), "state_step")
         self.increase_tensor = Tensor(1, mstype.int32)
-        self.op_mul = P.Mul()
         self.assignadd = P.AssignAdd()
+        self.op_add = P.AddN()
+        self.op_mul = P.Mul()
         self.op_pow = P.Pow()
-        self.op_sqrt = P.Sqrt()
-        self.op_maximum = P.Maximum()
+        self.adam_opt = P.Adam(False, False)
+        self.op_cast = P.Cast()
 
     def construct(self, gradients):
+        self.assignadd(self.state_step, self.increase_tensor)
         for group_id, group in enumerate(self.param_groups):
-            params = []
-            grads = []
-            exp_avgs = []
-            exp_avg_sqs = []
-            max_exp_avg_sqs = []
-            state_steps = []
-            params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps = \
-                self._init_group(group, gradients, params, grads, exp_avgs, exp_avg_sqs,
-                                 max_exp_avg_sqs, state_steps, group_id)
-            beta1, beta2 = group['betas']
-            self.apply_adam(params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps,
-                            group['amsgrad'], beta1, beta2, group['lr'], group['weight_decay'], group['eps'],
-                            group["maximize"], group["grad_centralization"])
-
-    def apply_adam(self, params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps,
-                   amsgrad, beta1, beta2, lr, weight_decay, eps, maximize, grad_centralization):
-        grads = self._decay_weight(weight_decay, params, grads)
-        grads = self._gradients_centralization(grad_centralization, grads)
-
-        for i, param in enumerate(params):
-            grad = grads[i] if not maximize else -grads[i]
-            exp_avg = exp_avgs[i]
-            exp_avg_sq = exp_avg_sqs[i]
-            step_t = state_steps[i]
-
-            F.assign(exp_avg, self.op_mul(exp_avg, beta1) + self.op_mul(grad, 1-beta1))
-            F.assign(exp_avg_sq, ops.addcmul(self.op_mul(exp_avg_sq, beta2), grad, grad.conj(), 1-beta2))
-            step_t = F.depend(step_t, self.assignadd(step_t, self.increase_tensor))
-
-            bias_correction1 = F.tuple_to_array((1.0,)) - self.op_pow(beta1, step_t)
-            bias_correction2 = F.tuple_to_array((1.0,)) - self.op_pow(beta2, step_t)
-            step_size = lr / bias_correction1
-            bias_correction2_sqrt = self.op_sqrt(bias_correction2)
-
-            if amsgrad:
-                F.assign(max_exp_avg_sqs[i], self.op_maximum(max_exp_avg_sqs[i], exp_avg_sq))
-                denom = self.op_sqrt(max_exp_avg_sqs[i]) / bias_correction2_sqrt + eps
+            start_id = self.group_start_id[group_id]
+            end_id = self.group_start_id[group_id+1]
+
+            lr = group.get("lr")
+            weight_decay = group.get("weight_decay")
+            beta1, beta2 = group.get("betas")
+            maximize = group.get("maximize")
+            eps = group.get("eps")
+
+            beta1_power = self.op_pow(beta1, self.state_step)
+            beta2_power = self.op_pow(beta2, self.state_step)
+            adam_with_amsgrad_opt = P.ApplyAdamWithAmsgrad(beta1, beta2, eps, False)
+            params = self.parameters[start_id: end_id]
+            grads = gradients[start_id: end_id]
+            grads = grads if not maximize else -grads
+            grads = self._decay_weight(weight_decay, params, grads)
+            if isinstance(lr, float):
+                lr = self.op_cast(group.get("lr"), mstype.float32)
+            if group.get("amsgrad"):
+                self.hyper_map(F.partial(_adam_opt, adam_with_amsgrad_opt, beta1_power, beta2_power, lr),
+                               grads, params,
+                               self.exp_avg[start_id: end_id], self.exp_avg_sq[start_id: end_id],
+                               self.max_exp_avg_sq[start_id: end_id])
             else:
-                denom = self.op_sqrt(exp_avg_sq) / bias_correction2_sqrt + eps
-            param = F.depend(param, denom)
-            next_param = param - self.op_mul(exp_avg / denom, step_size)
-            F.assign(param, next_param)
-
-    def _init_group(self, group, gradients, params, grads, exp_avgs, exp_avg_sqs,
-                    max_exp_avg_sqs, state_steps, group_id):
-        """ Initialize group params. """
-        p_id = self.group_start_id[group_id]
-        for i, param in enumerate(group["params"]):
-            params.append(param)
-            grads.append(gradients[p_id+i])
-            exp_avgs.append(self.exp_avg[p_id+i])
-            exp_avg_sqs.append(self.exp_avg_sq[p_id+i])
-            max_exp_avg_sqs.append(self.max_exp_avg_sq[p_id+i])
-            state_steps.append(self.state_step[p_id+i])
-        return params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps
+                self.hyper_map(F.partial(_adam_opt, self.adam_opt, beta1_power, beta2_power, beta1, beta2, eps, lr),
+                               grads, params,
+                               self.exp_avg[start_id: end_id], self.exp_avg_sq[start_id: end_id])
+        return True