mindscience.e3nn.nn.FullyConnectedNet

class mindscience.e3nn.nn.FullyConnectedNet(h_list, act=None, out_act=False, init_method='normal', dtype=mindspore.float32)[source]

Fully-connected Neural Network with normalized activation on scalars. It stacks multiple dense layers and automatically normalizes the activation function to maintain stable signal magnitudes during forward/backward passes.

Parameters

  • h_list (list[int]) – A list of input, internal and output dimensions for dense layers.

  • act (Func, optional) – Activation function which will be automatically normalized. Default: None.

  • out_act (bool, optional) – Whether to apply the activation function on the output. Default: False.

  • init_method (Union[str, mindspore.common.initializer], optional) – Method to initialize parameters. Default: 'normal'.

  • dtype (mindspore.dtype, optional) – The data type of the input tensor. Default: mindspore.float32.

Inputs:

  • input (Tensor) - The shape of Tensor is \((h\_list[0])\).

Outputs:

  • output (Tensor) - The shape of Tensor is \((h\_list[-1])\).

Raises

TypeError – If the elements h_list are not int.

Examples

>>> import mindspore as ms
>>> from mindscience.e3nn.nn import FullyConnectedNet
>>> fc = FullyConnectedNet([4,10,20,12,6], ops.tanh)
FullyConnectedNet [4, 10, 20, 12, 6]
>>> v = ms.Tensor([.1,.2,.3,.4])
>>> grad = ops.grad(fc, weights=fc.trainable_params())
>>> fc(v).shape
(6,)
>>> [x.shape for x in grad(v)[1]]
[(4, 10), (10, 20), (20, 12), (12, 6)]