mindspore.nn.TransformerDecoderLayer

class mindspore.nn.TransformerDecoderLayer(d_model, nhead, dim_feedforward=2048, dropout=0.1, activation='relu', layer_norm_eps=1e-5, batch_first=False, norm_first=False, dtype=mstype.float32)[source]

Transformer Decoder Layer. This is an implementation of the single layer of the transformer decoder layer, including self-attention, cross attention and feedward layer.

Parameters

d_model (int) – The number of expected features in the input tensor.
nhead (int) – The number of heads in the MultiheadAttention modules.
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 (mindspore.nn.ReLU or mindspore.nn.GELU ) or a callable ( mindspore.ops.relu() or mindspore.ops.gelu() ). Default: "relu".
layer_norm_eps (float) – The epsilon value in LayerNorm modules. 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 located prior to attention and feedforward operations; if norm_first = False, layer norm is located after the attention and feedforward operations. Default: False.
dtype (mindspore.dtype) – Data type of Parameter. Default: mstype.float32 .

Inputs:

tgt (Tensor) - The sequence to the decoder layer. For unbatched input, the shape is \((T, E)\) ; otherwise if batch_first=False , the shape is \((T, N, E)\) and if batch_first=True , the shape is \((N, T, E)\), where \((T)\) is the target sequence length. Supported types: float16, float32, float64.
memory (Tensor) - The sequence from the last layer of the encoder. Supported types: float16, float32, float64.
tgt_mask (Tensor, optional) - The mask of the tgt sequence. The shape is \((T, T)\) or \((N*nhead, T, T)\). Supported types: float16, float32, float64, bool. Default: None.
memory_mask (Tensor, optional) - The mask of the memory sequence. The shape is \((T, S)\) . Supported types: float16, float32, float64, bool. Default: None.
tgt_key_padding_mask (Tensor, optional): The mask of the tgt keys per batch. The shape is \((T)\) for unbatched input, otherwise \((N, T)\) . Supported types: float16, float32, float64, bool. Default: None.
memory_key_padding_mask (Tensor, optional) - The mask of the memory keys per batch. The shape is \((S)\) for unbatched input, otherwise \((N, S)\) . Supported types: float16, float32, float64, bool. Default: None.

Outputs:

Tensor. The shape and dtype of Tensor is the same with tgt .

Raises

ValueError – If the init argument activation is not str, callable or Cell instance.
ValueError – If the init argument activation is not mindspore.nn.ReLU, mindspore.nn.GELU instance, mindspore.ops.relu(), mindspore.ops.gelu() , "relu" or "gelu" .

Supported Platforms:: Ascend GPU CPU

Examples

>>> import mindspore as ms
>>> import numpy as np
>>> decoder_layer = ms.nn.TransformerDecoderLayer(d_model=512, nhead=8)
>>> memory = ms.Tensor(np.random.rand(10, 32, 512), ms.float32)
>>> tgt = ms.Tensor(np.random.rand(20, 32, 512), ms.float32)
>>> out = decoder_layer(tgt, memory)
>>> print(out.shape)
(20, 32, 512)
>>> # Alternatively, when `batch_first` is ``True``:
>>> decoder_layer = ms.nn.TransformerDecoderLayer(d_model=512, nhead=8, batch_first=True)
>>> memory = ms.Tensor(np.random.rand(32, 10, 512), ms.float32)
>>> tgt = ms.Tensor(np.random.rand(32, 20, 512), ms.float32)
>>> out = decoder_layer(tgt, memory)
>>> print(out.shape)
(32, 20, 512)