# MIT License
#
# Copyright (c) 2021 Soohwan Kim and Sangchun Ha and Soyoung Cho
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import torch.nn as nn
from torch import Tensor
from typing import Optional
from openspeech.modules.relative_multi_head_attention import RelativeMultiHeadAttention
from openspeech.modules.positional_encoding import PositionalEncoding
[docs]class MultiHeadedSelfAttentionModule(nn.Module):
r"""
Conformer employ multi-headed self-attention (MHSA) while integrating an important technique from Transformer-XL,
the relative sinusoidal positional encoding scheme. The relative positional encoding allows the self-attention
module to generalize better on different input length and the resulting encoders is more robust to the variance of
the utterance length. Conformer use prenorm residual units with dropout which helps training
and regularizing deeper models.
Args:
d_model (int): The dimension of model
num_heads (int): The number of attention heads.
dropout_p (float): probability of dropout
Inputs: inputs, mask
- **inputs** (batch, time, dim): Tensor containing input vector
- **mask** (batch, 1, time2) or (batch, time1, time2): Tensor containing indices to be masked
Returns:
- **outputs** (batch, time, dim): Tensor produces by relative multi headed self attention module.
"""
def __init__(
self,
d_model: int,
num_heads: int,
dropout_p: float = 0.1,
) -> None:
super(MultiHeadedSelfAttentionModule, self).__init__()
self.positional_encoding = PositionalEncoding(d_model)
self.layer_norm = nn.LayerNorm(d_model)
self.attention = RelativeMultiHeadAttention(d_model, num_heads, dropout_p)
self.dropout = nn.Dropout(p=dropout_p)
[docs] def forward(self, inputs: Tensor, mask: Optional[Tensor] = None) -> Tensor:
r"""
Forward propagate of conformer's multi-headed self attention module.
Inputs: inputs, mask
- **inputs** (batch, time, dim): Tensor containing input vector
- **mask** (batch, 1, time2) or (batch, time1, time2): Tensor containing indices to be masked
Returns:
- **outputs** (batch, time, dim): Tensor produces by relative multi headed self attention module.
"""
batch_size, seq_length, _ = inputs.size()
pos_embedding = self.positional_encoding(seq_length)
pos_embedding = pos_embedding.repeat(batch_size, 1, 1)
inputs = self.layer_norm(inputs)
outputs = self.attention(inputs, inputs, inputs, pos_embedding=pos_embedding, mask=mask)
return self.dropout(outputs)