mindchemistry.cell.FCNet

View Source On Gitee
class mindchemistry.cell.FCNet(channels, weight_init='normal', has_bias=True, bias_init='zeros', has_dropout=False, dropout_rate=0.5, has_layernorm=False, layernorm_epsilon=1e-7, has_activation=True, act='relu')[source]

The Fully Connected Network. Applies a series of fully connected layers to the incoming data.

Parameters

  • channels (List) – the list of numbers of channel of each fully connected layers.

  • weight_init (Union[str, float, mindspore.common.initializer, List]) – initialize layer weights. If weight_init was List, each element corresponds to each layer. Default: 'normal' .

  • has_bias (Union[bool, List]) – The switch for whether the dense layers has bias. If has_bias was List, each element corresponds to each dense layer. Default: True .

  • bias_init (Union[str, float, mindspore.common.initializer, List]) – The initializer of the bias of dense layer. If bias_init was List, each element corresponds to each dense layer. Default: 'zeros' .

  • has_dropout (Union[bool, List]) – The switch for whether linear block has a dropout layer. If has_dropout was List, each element corresponds to each layer. Default: False .

  • dropout_rate (float) – The dropout rate for dropout layer, the dropout rate must be a float in range (0, 1] If dropout_rate was List, each element corresponds to each dropout layer. Default: 0.5 .

  • has_layernorm (Union[bool, List]) – The switch for whether linear block has a layer normalization layer. If has_layernorm was List, each element corresponds to each layer. Default: False .

  • layernorm_epsilon (float) – The hyper parameter epsilon for layer normalization layer. If layernorm_epsilon was List, each element corresponds to each layer normalization layer. Default: 1e-7 .

  • has_activation (Union[bool, List]) – The switch for whether linear block has an activation layer. If has_activation was List, each element corresponds to each layer. Default: True .

  • act (Union[str, None, List]) – The activation function in linear block. If act was List, each element corresponds to each activation layer. Default: 'relu' .

Inputs:

  • input (Tensor) - The shape of Tensor is \((*, channels[0])\).

Outputs:

  • output (Tensor) - The shape of Tensor is \((*, channels[-1])\).

Supported Platforms:

Ascend

Examples

>>> import numpy as np
>>> from mindchemistry.cell import FCNet
>>> from mindspore import Tensor
>>> inputs = Tensor(np.array([[180, 234, 154], [244, 48, 247]], np.float32))
>>> net = FCNet([3, 16, 32, 16, 8])
>>> output = net(inputs)
>>> print(output.shape)
(2, 8)