Upload Attention.py

Attention.py

@@ -0,0 +1,324 @@
+# Written by Shigeki Karita, 2019
+# Published under Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+# Adapted by Florian Lux, 2021
+
+"""Multi-Head Attention layer definition."""
+
+import math
+
+import numpy
+import torch
+from torch import nn
+
+from Utility.utils import make_non_pad_mask
+
+
+class MultiHeadedAttention(nn.Module):
+    """
+    Multi-Head Attention layer.
+
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+    """
+
+    def __init__(self, n_head, n_feat, dropout_rate):
+        """
+        Construct an MultiHeadedAttention object.
+        """
+        super(MultiHeadedAttention, self).__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        self.linear_q = nn.Linear(n_feat, n_feat)
+        self.linear_k = nn.Linear(n_feat, n_feat)
+        self.linear_v = nn.Linear(n_feat, n_feat)
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.attn = None
+        self.dropout = nn.Dropout(p=dropout_rate)
+
+    def forward_qkv(self, query, key, value):
+        """
+        Transform query, key and value.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+
+        Returns:
+            torch.Tensor: Transformed query tensor (#batch, n_head, time1, d_k).
+            torch.Tensor: Transformed key tensor (#batch, n_head, time2, d_k).
+            torch.Tensor: Transformed value tensor (#batch, n_head, time2, d_k).
+        """
+        n_batch = query.size(0)
+        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
+        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
+        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
+        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
+        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
+        v = v.transpose(1, 2)  # (batch, head, time2, d_k)
+
+        return q, k, v
+
+    def forward_attention(self, value, scores, mask):
+        """
+        Compute attention context vector.
+
+        Args:
+            value (torch.Tensor): Transformed value (#batch, n_head, time2, d_k).
+            scores (torch.Tensor): Attention score (#batch, n_head, time1, time2).
+            mask (torch.Tensor): Mask (#batch, 1, time2) or (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
+        """
+        n_batch = value.size(0)
+        if mask is not None:
+            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+            min_value = float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
+            scores = scores.masked_fill(mask, min_value)
+            self.attn = torch.softmax(scores, dim=-1).masked_fill(mask, 0.0)  # (batch, head, time1, time2)
+        else:
+            self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        p_attn = self.dropout(self.attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k))  # (batch, time1, d_model)
+
+        return self.linear_out(x)  # (batch, time1, d_model)
+
+    def forward(self, query, key, value, mask):
+        """
+        Compute scaled dot product attention.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+        return self.forward_attention(v, scores, mask)
+
+
+class RelPositionMultiHeadedAttention(MultiHeadedAttention):
+    """
+    Multi-Head Attention layer with relative position encoding.
+    Details can be found in https://github.com/espnet/espnet/pull/2816.
+    Paper: https://arxiv.org/abs/1901.02860
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+        zero_triu (bool): Whether to zero the upper triangular part of attention matrix.
+    """
+
+    def __init__(self, n_head, n_feat, dropout_rate, zero_triu=False):
+        """Construct an RelPositionMultiHeadedAttention object."""
+        super().__init__(n_head, n_feat, dropout_rate)
+        self.zero_triu = zero_triu
+        # linear transformation for positional encoding
+        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        torch.nn.init.xavier_uniform_(self.pos_bias_u)
+        torch.nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def rel_shift(self, x):
+        """
+        Compute relative positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, head, time1, 2*time1-1).
+            time1 means the length of query vector.
+        Returns:
+            torch.Tensor: Output tensor.
+        """
+        zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
+        x_padded = torch.cat([zero_pad, x], dim=-1)
+
+        x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
+        x = x_padded[:, :, 1:].view_as(x)[:, :, :, : x.size(-1) // 2 + 1]  # only keep the positions from 0 to time2
+
+        if self.zero_triu:
+            ones = torch.ones((x.size(2), x.size(3)), device=x.device)
+            x = x * torch.tril(ones, x.size(3) - x.size(2))[None, None, :, :]
+
+        return x
+
+    def forward(self, query, key, value, pos_emb, mask):
+        """
+        Compute 'Scaled Dot Product Attention' with rel. positional encoding.
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            pos_emb (torch.Tensor): Positional embedding tensor
+                (#batch, 2*time1-1, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        q = q.transpose(1, 2)  # (batch, time1, head, d_k)
+
+        n_batch_pos = pos_emb.size(0)
+        p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        # (batch, head, time1, d_k)
+        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
+        # (batch, head, time1, d_k)
+        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        # (batch, head, time1, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
+
+        # compute matrix b and matrix d
+        # (batch, head, time1, 2*time1-1)
+        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k)  # (batch, head, time1, time2)
+
+        return self.forward_attention(v, scores, mask)
+
+
+class GuidedAttentionLoss(torch.nn.Module):
+    """
+    Guided attention loss function module.
+
+    This module calculates the guided attention loss described
+    in `Efficiently Trainable Text-to-Speech System Based
+    on Deep Convolutional Networks with Guided Attention`_,
+    which forces the attention to be diagonal.
+
+    .. _`Efficiently Trainable Text-to-Speech System
+        Based on Deep Convolutional Networks with Guided Attention`:
+        https://arxiv.org/abs/1710.08969
+    """
+
+    def __init__(self, sigma=0.4, alpha=1.0):
+        """
+        Initialize guided attention loss module.
+
+        Args:
+            sigma (float, optional): Standard deviation to control
+                how close attention to a diagonal.
+            alpha (float, optional): Scaling coefficient (lambda).
+            reset_always (bool, optional): Whether to always reset masks.
+        """
+        super(GuidedAttentionLoss, self).__init__()
+        self.sigma = sigma
+        self.alpha = alpha
+        self.guided_attn_masks = None
+        self.masks = None
+
+    def _reset_masks(self):
+        self.guided_attn_masks = None
+        self.masks = None
+
+    def forward(self, att_ws, ilens, olens):
+        """
+        Calculate forward propagation.
+
+        Args:
+            att_ws (Tensor): Batch of attention weights (B, T_max_out, T_max_in).
+            ilens (LongTensor): Batch of input lenghts (B,).
+            olens (LongTensor): Batch of output lenghts (B,).
+
+        Returns:
+            Tensor: Guided attention loss value.
+        """
+        self._reset_masks()
+        self.guided_attn_masks = self._make_guided_attention_masks(ilens, olens).to(att_ws.device)
+        self.masks = self._make_masks(ilens, olens).to(att_ws.device)
+        losses = self.guided_attn_masks * att_ws
+        loss = torch.mean(losses.masked_select(self.masks))
+        self._reset_masks()
+        return self.alpha * loss
+
+    def _make_guided_attention_masks(self, ilens, olens):
+        n_batches = len(ilens)
+        max_ilen = max(ilens)
+        max_olen = max(olens)
+        guided_attn_masks = torch.zeros((n_batches, max_olen, max_ilen), device=ilens.device)
+        for idx, (ilen, olen) in enumerate(zip(ilens, olens)):
+            guided_attn_masks[idx, :olen, :ilen] = self._make_guided_attention_mask(ilen, olen, self.sigma)
+        return guided_attn_masks
+
+    @staticmethod
+    def _make_guided_attention_mask(ilen, olen, sigma):
+        """
+        Make guided attention mask.
+        """
+        grid_x, grid_y = torch.meshgrid(torch.arange(olen, device=olen.device).float(), torch.arange(ilen, device=ilen.device).float())
+        return 1.0 - torch.exp(-((grid_y / ilen - grid_x / olen) ** 2) / (2 * (sigma ** 2)))
+
+    @staticmethod
+    def _make_masks(ilens, olens):
+        """
+        Make masks indicating non-padded part.
+
+        Args:
+            ilens (LongTensor or List): Batch of lengths (B,).
+            olens (LongTensor or List): Batch of lengths (B,).
+
+        Returns:
+            Tensor: Mask tensor indicating non-padded part.
+                    dtype=torch.uint8 in PyTorch 1.2-
+                    dtype=torch.bool in PyTorch 1.2+ (including 1.2)
+        """
+        in_masks = make_non_pad_mask(ilens, device=ilens.device)  # (B, T_in)
+        out_masks = make_non_pad_mask(olens, device=olens.device)  # (B, T_out)
+        return out_masks.unsqueeze(-1) & in_masks.unsqueeze(-2)  # (B, T_out, T_in)
+
+
+class GuidedMultiHeadAttentionLoss(GuidedAttentionLoss):
+    """
+    Guided attention loss function module for multi head attention.
+
+    Args:
+        sigma (float, optional): Standard deviation to control
+        how close attention to a diagonal.
+        alpha (float, optional): Scaling coefficient (lambda).
+        reset_always (bool, optional): Whether to always reset masks.
+    """
+
+    def forward(self, att_ws, ilens, olens):
+        """
+        Calculate forward propagation.
+
+        Args:
+            att_ws (Tensor):
+                Batch of multi head attention weights (B, H, T_max_out, T_max_in).
+            ilens (LongTensor): Batch of input lenghts (B,).
+            olens (LongTensor): Batch of output lenghts (B,).
+
+        Returns:
+            Tensor: Guided attention loss value.
+        """
+        if self.guided_attn_masks is None:
+            self.guided_attn_masks = (self._make_guided_attention_masks(ilens, olens).to(att_ws.device).unsqueeze(1))
+        if self.masks is None:
+            self.masks = self._make_masks(ilens, olens).to(att_ws.device).unsqueeze(1)
+        losses = self.guided_attn_masks * att_ws
+        loss = torch.mean(losses.masked_select(self.masks))
+        if self.reset_always:
+            self._reset_masks()
+
+        return self.alpha * loss