mindspore.nn.Transformer

class mindspore.nn.Transformer(d_model: int = 512, nhead: int = 8, num_encoder_layers: int = 6, num_decoder_layers: int = 6, dim_feedforward: int = 2048, dropout: float = 0.1, activation: Union[str, Cell, callable] = 'relu', custom_encoder: Optional[Cell] = None, custom_decoder: Optional[Cell] = None, layer_norm_eps: float = 1e-05, batch_first: bool = False, norm_first: bool = False)[source]

Transformer module including encoder and decoder. The difference with the original implements is the module use the residual addition before the layer normalization. And the default hidden act is gelu. The details can be found in Attention is all you need.

Parameters

  • d_model (int) – The number of expected features in the inputs tensor. Default: 512.

  • nhead (int) – The number of heads in the MultiheadAttention modules. Default: 8.

  • num_encoder_layers (int) – The number of encoder-layers in the encoder. Default: 6.

  • num_decoder_layers (int) – The number of decoder-layers in the decoder. Default: 6.

  • dim_feedforward (int) – The dimension of the feedforward layer. Default: 2048.

  • dropout (float) – The dropout value. Default: 0.1.

  • activation (Union[str, callable, Cell]) – The activation function of the intermediate layer, can be a string (“relu” or “gelu”), Cell instance (nn.ReLU() or nn.GELU()) or a callable (ops.relu or ops.gelu). Default: "relu"

  • custom_encoder (Cell) – Custom encoder. Default: None.

  • custom_decoder (Cell) – Custom decoder. Default: None.

  • layer_norm_eps (float) – the epsilion value in layer normalization module. Default: 1e-5.

  • batch_first (bool) – If batch_first = True, then the shape of input and output tensors is \((batch, seq, feature)\) , otherwise the shape is \((seq, batch, feature)\) . Default: False.

  • norm_first (bool) – If norm_first = True, layer norm is done prior to attention and feedforward operations, respectively. Default: False.

Inputs:

  • src (Tensor): The source sequence to the encoder.

  • tgt (Tensor): The target sequence to the decoder.

  • src_mask (Tensor, optional): The mask of the src sequence. Default: None.

  • tgt_mask (Tensor, optional): The mask of the tgt sequence. Default: None.

  • memory_mask (Tensor, optional): The additive mask of the encoder output. Default: None.

  • src_key_padding_mask (Tensor, optional): The mask of src keys per batch. Default: None.

  • tgt_key_padding_mask (Tensor, optional): The mask of tgt keys per batch. Default: None.

  • memory_key_padding_mask (Tensor, optional): The mask of memory keys per batch. Default: None.

Outputs:

Tensor.

Supported Platforms:

Ascend GPU CPU

Examples

>>> transformer_model = nn.Transformer(nhead=16, num_encoder_layers=12)
>>> src = Tensor(np.random.rand(10, 32, 512), mindspore.float32)
>>> tgt = Tensor(np.random.rand(20, 32, 512), mindspore.float32)
>>> out = transformer_model(src, tgt)
>>> print(out.shape)
(20, 32, 512)