# Copyright 2021 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.
# ============================================================================
"""
derivative
"""
from mindspore import nn, ops, jacrev
from mindspore.ops import constexpr
from mindspore import dtype as mstype
class SimplifiedGradient(nn.Cell):
"""Simplify the results of the input network."""
def __init__(self, net, order=1):
super().__init__()
if not isinstance(order, int):
raise TypeError("The type of order should be int, but got {}".format(type(order)))
self.net = net
self.axis = order - 1
self.cast = ops.Cast()
def construct(self, x):
return self.cast(self.net(x).sum(axis=self.axis), mstype.float32)
# generate api by del decorator.
[文档]def batched_jacobian(model):
"""
Calculate Jacobian matrix of network model.
Args:
model (mindspore.nn.Cell): a network with the input dimension is in_channels and output dimension is
out_channels.
Returns:
jacobian(Tensor), jacobi of the model. With the input dimension is [batch_size, in_channels], output dimension
is [out_channels, batch_size, in_channels].
Note:
The version of MindSpore should be >= 2.0.0 for using `mindspore.jacrev`.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> import numpy as np
>>> from mindspore import nn, ops, Tensor
>>> from mindspore import dtype as mstype
>>> from mindflow.operators import batched_jacobian
>>> np.random.seed(123456)
>>> class Net(nn.Cell):
... def __init__(self, cin=2, cout=1, hidden=10):
... super().__init__()
... self.fc1 = nn.Dense(cin, hidden)
... self.fc2 = nn.Dense(hidden, hidden)
... self.fcout = nn.Dense(hidden, cout)
... self.act = ops.Tanh()
...
... def construct(self, x):
... x = self.act(self.fc1(x))
... x = self.act(self.fc2(x))
... x = self.fcout(x)
... return x
>>> model = Net()
>>> jacobian = batched_jacobian(model)
>>> inputs = np.random.random(size=(3, 2))
>>> res = jacobian(Tensor(inputs, mstype.float32))
>>> print(res.shape)
(1, 3, 2)
"""
return jacrev(SimplifiedGradient(model, 1))
# generate api by del decorator.
[文档]def batched_hessian(model):
"""
Calculate Hessian matrix of network model.
Args:
model (mindspore.nn.Cell): a network with the input dimension is in_channels and output dimension is
out_channels.
Returns:
hessian(Tensor), hessian of the model. With the input dimension is [batch_size, in_channels], output dimension
is [out_channels, in_channels, batch_size, in_channels].
Note:
The version of MindSpore should be >= 2.0.0 for using `mindspore.jacrev`.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> import numpy as np
>>> from mindspore import nn, ops, Tensor
>>> from mindspore import dtype as mstype
>>> from mindflow.operators import batched_hessian
>>> np.random.seed(123456)
>>> class Net(nn.Cell):
... def __init__(self, cin=2, cout=1, hidden=10):
... super().__init__()
... self.fc1 = nn.Dense(cin, hidden)
... self.fc2 = nn.Dense(hidden, hidden)
... self.fcout = nn.Dense(hidden, cout)
... self.act = ops.Tanh()
...
... def construct(self, x):
... x = self.act(self.fc1(x))
... x = self.act(self.fc2(x))
... x = self.fcout(x)
... return x
>>> model = Net()
>>> hessian = batched_hessian(model)
>>> inputs = np.random.random(size=(3, 2))
>>> res = hessian(Tensor(inputs, mstype.float32))
>>> print(res.shape)
(1, 2, 3, 2)
"""
return jacrev(SimplifiedGradient(batched_jacobian(model), 2))