mindscience.common.batched_jacobian

mindscience.common.batched_jacobian(model)

Calculate Jacobian matrix of network model.

Note

This function internally relies on mindspore.jacrev to compute Jacobian matrices. Therefore, MindSpore version >= 2.0.0 is required.

Parameters

model (mindspore.nn.Cell) – A network with the input dimension is in_channels and output dimension is out_channels.

Returns

Tensor, jacobi of the model. With the input dimension is \([batch_size, in_channels]\), output dimension is \([out_channels, batch_size, in_channels]\).

Examples

>>> import numpy as np
>>> from mindspore import nn, ops, Tensor
>>> from mindspore import dtype as mstype
>>> from mindscience.common 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)