From 4ace5a95b052d338947fc88809a440ccd55cf6b4 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 16 十一月 2023 16:39:52 +0800
Subject: [PATCH] funasr pages

---
 funasr/modules/attention.py |  494 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 files changed, 480 insertions(+), 14 deletions(-)

diff --git a/funasr/modules/attention.py b/funasr/modules/attention.py
index e3ad56a..b007d58 100644
--- a/funasr/modules/attention.py
+++ b/funasr/modules/attention.py
@@ -11,7 +11,11 @@
 import numpy
 import torch
 from torch import nn
+from typing import Optional, Tuple
 
+import torch.nn.functional as F
+from funasr.modules.nets_utils import make_pad_mask
+import funasr.modules.lora.layers as lora
 
 class MultiHeadedAttention(nn.Module):
     """Multi-Head Attention layer.
@@ -318,7 +322,7 @@
 
     """
 
-    def __init__(self, n_head, in_feat, n_feat, dropout_rate, kernel_size, sanm_shfit=0):
+    def __init__(self, n_head, in_feat, n_feat, dropout_rate, kernel_size, sanm_shfit=0, lora_list=None, lora_rank=8, lora_alpha=16, lora_dropout=0.1):
         """Construct an MultiHeadedAttention object."""
         super(MultiHeadedAttentionSANM, self).__init__()
         assert n_feat % n_head == 0
@@ -328,8 +332,19 @@
         # 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.linear_q_k_v = nn.Linear(in_feat, n_feat * 3)
+        if lora_list is not None:
+            if "o" in lora_list:
+                self.linear_out = lora.Linear(n_feat, n_feat, r=lora_rank, lora_alpha=lora_alpha, lora_dropout=lora_dropout)
+            else:
+                self.linear_out = nn.Linear(n_feat, n_feat)
+            lora_qkv_list = ["q" in lora_list, "k" in lora_list, "v" in lora_list]
+            if lora_qkv_list == [False, False, False]:
+                self.linear_q_k_v = nn.Linear(in_feat, n_feat * 3)
+            else:
+                self.linear_q_k_v = lora.MergedLinear(in_feat, n_feat * 3, r=lora_rank, lora_alpha=lora_alpha, lora_dropout=lora_dropout, enable_lora=lora_qkv_list)
+        else:
+            self.linear_out = nn.Linear(n_feat, n_feat)
+            self.linear_q_k_v = nn.Linear(in_feat, n_feat * 3)
         self.attn = None
         self.dropout = nn.Dropout(p=dropout_rate)
 
@@ -347,15 +362,17 @@
             mask = torch.reshape(mask, (b, -1, 1))
             if mask_shfit_chunk is not None:
                 mask = mask * mask_shfit_chunk
+            inputs = inputs * mask
 
-        inputs = inputs * mask
         x = inputs.transpose(1, 2)
         x = self.pad_fn(x)
         x = self.fsmn_block(x)
         x = x.transpose(1, 2)
         x += inputs
         x = self.dropout(x)
-        return x * mask
+        if mask is not None:
+            x = x * mask
+        return x
 
     def forward_qkv(self, x):
         """Transform query, key and value.
@@ -439,6 +456,56 @@
         att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)
         return att_outs + fsmn_memory
 
+    def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
+        """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_h, k_h, v_h, v = self.forward_qkv(x)
+        if chunk_size is not None and look_back > 0 or look_back == -1:
+            if cache is not None:
+                k_h_stride = k_h[:, :, :-(chunk_size[2]), :]
+                v_h_stride = v_h[:, :, :-(chunk_size[2]), :]
+                k_h = torch.cat((cache["k"], k_h), dim=2)
+                v_h = torch.cat((cache["v"], v_h), dim=2)
+
+                cache["k"] = torch.cat((cache["k"], k_h_stride), dim=2)
+                cache["v"] = torch.cat((cache["v"], v_h_stride), dim=2)
+                if look_back != -1:
+                    cache["k"] = cache["k"][:, :, -(look_back * chunk_size[1]):, :]
+                    cache["v"] = cache["v"][:, :, -(look_back * chunk_size[1]):, :]
+            else:
+                cache_tmp = {"k": k_h[:, :, :-(chunk_size[2]), :],
+                             "v": v_h[:, :, :-(chunk_size[2]), :]}
+                cache = cache_tmp
+        fsmn_memory = self.forward_fsmn(v, None)
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        att_outs = self.forward_attention(v_h, scores, None)
+        return att_outs + fsmn_memory, cache
+
+
+class MultiHeadedAttentionSANMwithMask(MultiHeadedAttentionSANM):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x, mask, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
+        q_h, k_h, v_h, v = self.forward_qkv(x)
+        fsmn_memory = self.forward_fsmn(v, mask[0], mask_shfit_chunk)
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        att_outs = self.forward_attention(v_h, scores, mask[1], mask_att_chunk_encoder)
+        return att_outs + fsmn_memory
+
 class MultiHeadedAttentionSANMDecoder(nn.Module):
     """Multi-Head Attention layer.
 
@@ -493,7 +560,7 @@
             # print("in fsmn, cache is None, x", x.size())
 
             x = self.pad_fn(x)
-            if not self.training and t <= 1:
+            if not self.training:
                 cache = x
         else:
             # print("in fsmn, cache is not None, x", x.size())
@@ -501,7 +568,7 @@
             # if t < self.kernel_size:
             #     x = self.pad_fn(x)
             x = torch.cat((cache[:, :, 1:], x), dim=2)
-            x = x[:, :, -self.kernel_size:]
+            x = x[:, :, -(self.kernel_size+t-1):]
             # print("in fsmn, cache is not None, x_cat", x.size())
             cache = x
         x = self.fsmn_block(x)
@@ -526,18 +593,32 @@
 
     """
 
-    def __init__(self, n_head, n_feat, dropout_rate, encoder_output_size=None):
+    def __init__(self, n_head, n_feat, dropout_rate, lora_list=None, lora_rank=8, lora_alpha=16, lora_dropout=0.1, encoder_output_size=None):
         """Construct an MultiHeadedAttention object."""
         super(MultiHeadedAttentionCrossAtt, 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_k_v = nn.Linear(n_feat if encoder_output_size is None else encoder_output_size, n_feat*2)
-        self.linear_out = nn.Linear(n_feat, n_feat)
+        if lora_list is not None:
+            if "q" in lora_list:
+                self.linear_q = lora.Linear(n_feat, n_feat, r=lora_rank, lora_alpha=lora_alpha, lora_dropout=lora_dropout)
+            else:
+                self.linear_q = nn.Linear(n_feat, n_feat)
+            lora_kv_list = ["k" in lora_list, "v" in lora_list]
+            if lora_kv_list == [False, False]:
+                self.linear_k_v = nn.Linear(n_feat if encoder_output_size is None else encoder_output_size, n_feat*2)
+            else:
+                self.linear_k_v = lora.MergedLinear(n_feat if encoder_output_size is None else encoder_output_size, n_feat * 2, 
+                                      r=lora_rank, lora_alpha=lora_alpha, lora_dropout=lora_dropout, enable_lora=lora_kv_list)
+            if "o" in lora_list:
+                self.linear_out = lora.Linear(n_feat, n_feat, r=lora_rank, lora_alpha=lora_alpha, lora_dropout=lora_dropout)
+            else:
+                self.linear_out = nn.Linear(n_feat, n_feat)
+        else:
+            self.linear_q = nn.Linear(n_feat, n_feat)
+            self.linear_k_v = nn.Linear(n_feat if encoder_output_size is None else encoder_output_size, n_feat*2)
+            self.linear_out = nn.Linear(n_feat, n_feat)
         self.attn = None
         self.dropout = nn.Dropout(p=dropout_rate)
 
@@ -622,4 +703,389 @@
         q_h, k_h, v_h = self.forward_qkv(x, memory)
         q_h = q_h * self.d_k ** (-0.5)
         scores = torch.matmul(q_h, k_h.transpose(-2, -1))
-        return self.forward_attention(v_h, scores, memory_mask)
\ No newline at end of file
+        return self.forward_attention(v_h, scores, memory_mask)
+
+    def forward_chunk(self, x, memory, cache=None, chunk_size=None, look_back=0):
+        """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_h, k_h, v_h = self.forward_qkv(x, memory)
+        if chunk_size is not None and look_back > 0:
+            if cache is not None:
+                k_h = torch.cat((cache["k"], k_h), dim=2)
+                v_h = torch.cat((cache["v"], v_h), dim=2)
+                cache["k"] = k_h[:, :, -(look_back * chunk_size[1]):, :]
+                cache["v"] = v_h[:, :, -(look_back * chunk_size[1]):, :]
+            else:
+                cache_tmp = {"k": k_h[:, :, -(look_back * chunk_size[1]):, :],
+                             "v": v_h[:, :, -(look_back * chunk_size[1]):, :]}
+                cache = cache_tmp
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        return self.forward_attention(v_h, scores, None), cache
+
+
+class MultiHeadSelfAttention(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, in_feat, n_feat, dropout_rate):
+        """Construct an MultiHeadedAttention object."""
+        super(MultiHeadSelfAttention, 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_out = nn.Linear(n_feat, n_feat)
+        self.linear_q_k_v = nn.Linear(in_feat, n_feat * 3)
+        self.attn = None
+        self.dropout = nn.Dropout(p=dropout_rate)
+
+    def forward_qkv(self, x):
+        """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).
+
+        """
+        b, t, d = x.size()
+        q_k_v = self.linear_q_k_v(x)
+        q, k, v = torch.split(q_k_v, int(self.h * self.d_k), dim=-1)
+        q_h = torch.reshape(q, (b, t, self.h, self.d_k)).transpose(1, 2)  # (batch, head, time1, d_k)
+        k_h = torch.reshape(k, (b, t, self.h, self.d_k)).transpose(1, 2)  # (batch, head, time2, d_k)
+        v_h = torch.reshape(v, (b, t, self.h, self.d_k)).transpose(1, 2)  # (batch, head, time2, d_k)
+
+        return q_h, k_h, v_h, v
+
+    def forward_attention(self, value, scores, mask, mask_att_chunk_encoder=None):
+        """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:
+            if mask_att_chunk_encoder is not None:
+                mask = mask * mask_att_chunk_encoder
+
+            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, x, mask, mask_att_chunk_encoder=None):
+        """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_h, k_h, v_h, v = self.forward_qkv(x)
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)
+        return att_outs
+
+class RelPositionMultiHeadedAttentionChunk(torch.nn.Module):
+    """RelPositionMultiHeadedAttention definition.
+    Args:
+        num_heads: Number of attention heads.
+        embed_size: Embedding size.
+        dropout_rate: Dropout rate.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        embed_size: int,
+        dropout_rate: float = 0.0,
+        simplified_attention_score: bool = False,
+    ) -> None:
+        """Construct an MultiHeadedAttention object."""
+        super().__init__()
+
+        self.d_k = embed_size // num_heads
+        self.num_heads = num_heads
+
+        assert self.d_k * num_heads == embed_size, (
+            "embed_size (%d) must be divisible by num_heads (%d)",
+            (embed_size, num_heads),
+        )
+
+        self.linear_q = torch.nn.Linear(embed_size, embed_size)
+        self.linear_k = torch.nn.Linear(embed_size, embed_size)
+        self.linear_v = torch.nn.Linear(embed_size, embed_size)
+
+        self.linear_out = torch.nn.Linear(embed_size, embed_size)
+
+        if simplified_attention_score:
+            self.linear_pos = torch.nn.Linear(embed_size, num_heads)
+
+            self.compute_att_score = self.compute_simplified_attention_score
+        else:
+            self.linear_pos = torch.nn.Linear(embed_size, embed_size, bias=False)
+
+            self.pos_bias_u = torch.nn.Parameter(torch.Tensor(num_heads, self.d_k))
+            self.pos_bias_v = torch.nn.Parameter(torch.Tensor(num_heads, self.d_k))
+            torch.nn.init.xavier_uniform_(self.pos_bias_u)
+            torch.nn.init.xavier_uniform_(self.pos_bias_v)
+
+            self.compute_att_score = self.compute_attention_score
+
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.attn = None
+
+    def rel_shift(self, x: torch.Tensor, left_context: int = 0) -> torch.Tensor:
+        """Compute relative positional encoding.
+        Args:
+            x: Input sequence. (B, H, T_1, 2 * T_1 - 1)
+            left_context: Number of frames in left context.
+        Returns:
+            x: Output sequence. (B, H, T_1, T_2)
+        """
+        batch_size, n_heads, time1, n = x.shape
+        time2 = time1 + left_context
+
+        batch_stride, n_heads_stride, time1_stride, n_stride = x.stride()
+
+        return x.as_strided(
+            (batch_size, n_heads, time1, time2),
+            (batch_stride, n_heads_stride, time1_stride - n_stride, n_stride),
+            storage_offset=(n_stride * (time1 - 1)),
+        )
+
+    def compute_simplified_attention_score(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        pos_enc: torch.Tensor,
+        left_context: int = 0,
+    ) -> torch.Tensor:
+        """Simplified attention score computation.
+        Reference: https://github.com/k2-fsa/icefall/pull/458
+        Args:
+            query: Transformed query tensor. (B, H, T_1, d_k)
+            key: Transformed key tensor. (B, H, T_2, d_k)
+            pos_enc: Positional embedding tensor. (B, 2 * T_1 - 1, size)
+            left_context: Number of frames in left context.
+        Returns:
+            : Attention score. (B, H, T_1, T_2)
+        """
+        pos_enc = self.linear_pos(pos_enc)
+
+        matrix_ac = torch.matmul(query, key.transpose(2, 3))
+
+        matrix_bd = self.rel_shift(
+            pos_enc.transpose(1, 2).unsqueeze(2).repeat(1, 1, query.size(2), 1),
+            left_context=left_context,
+        )
+
+        return (matrix_ac + matrix_bd) / math.sqrt(self.d_k)
+
+    def compute_attention_score(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        pos_enc: torch.Tensor,
+        left_context: int = 0,
+    ) -> torch.Tensor:
+        """Attention score computation.
+        Args:
+            query: Transformed query tensor. (B, H, T_1, d_k)
+            key: Transformed key tensor. (B, H, T_2, d_k)
+            pos_enc: Positional embedding tensor. (B, 2 * T_1 - 1, size)
+            left_context: Number of frames in left context.
+        Returns:
+            : Attention score. (B, H, T_1, T_2)
+        """
+        p = self.linear_pos(pos_enc).view(pos_enc.size(0), -1, self.num_heads, self.d_k)
+
+        query = query.transpose(1, 2)
+        q_with_bias_u = (query + self.pos_bias_u).transpose(1, 2)
+        q_with_bias_v = (query + self.pos_bias_v).transpose(1, 2)
+
+        matrix_ac = torch.matmul(q_with_bias_u, key.transpose(-2, -1))
+
+        matrix_bd = torch.matmul(q_with_bias_v, p.permute(0, 2, 3, 1))
+        matrix_bd = self.rel_shift(matrix_bd, left_context=left_context)
+
+        return (matrix_ac + matrix_bd) / math.sqrt(self.d_k)
+
+    def forward_qkv(
+        self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Transform query, key and value.
+        Args:
+            query: Query tensor. (B, T_1, size)
+            key: Key tensor. (B, T_2, size)
+            v: Value tensor. (B, T_2, size)
+        Returns:
+            q: Transformed query tensor. (B, H, T_1, d_k)
+            k: Transformed key tensor. (B, H, T_2, d_k)
+            v: Transformed value tensor. (B, H, T_2, d_k)
+        """
+        n_batch = query.size(0)
+
+        q = (
+            self.linear_q(query)
+            .view(n_batch, -1, self.num_heads, self.d_k)
+            .transpose(1, 2)
+        )
+        k = (
+            self.linear_k(key)
+            .view(n_batch, -1, self.num_heads, self.d_k)
+            .transpose(1, 2)
+        )
+        v = (
+            self.linear_v(value)
+            .view(n_batch, -1, self.num_heads, self.d_k)
+            .transpose(1, 2)
+        )
+
+        return q, k, v
+
+    def forward_attention(
+        self,
+        value: torch.Tensor,
+        scores: torch.Tensor,
+        mask: torch.Tensor,
+        chunk_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Compute attention context vector.
+        Args:
+            value: Transformed value. (B, H, T_2, d_k)
+            scores: Attention score. (B, H, T_1, T_2)
+            mask: Source mask. (B, T_2)
+            chunk_mask: Chunk mask. (T_1, T_1)
+        Returns:
+           attn_output: Transformed value weighted by attention score. (B, T_1, H * d_k)
+        """
+        batch_size = scores.size(0)
+        mask = mask.unsqueeze(1).unsqueeze(2)
+        if chunk_mask is not None:
+            mask = chunk_mask.unsqueeze(0).unsqueeze(1) | mask
+        scores = scores.masked_fill(mask, float("-inf"))
+        self.attn = torch.softmax(scores, dim=-1).masked_fill(mask, 0.0)
+
+        attn_output = self.dropout(self.attn)
+        attn_output = torch.matmul(attn_output, value)
+
+        attn_output = self.linear_out(
+            attn_output.transpose(1, 2)
+            .contiguous()
+            .view(batch_size, -1, self.num_heads * self.d_k)
+        )
+
+        return attn_output
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        pos_enc: torch.Tensor,
+        mask: torch.Tensor,
+        chunk_mask: Optional[torch.Tensor] = None,
+        left_context: int = 0,
+    ) -> torch.Tensor:
+        """Compute scaled dot product attention with rel. positional encoding.
+        Args:
+            query: Query tensor. (B, T_1, size)
+            key: Key tensor. (B, T_2, size)
+            value: Value tensor. (B, T_2, size)
+            pos_enc: Positional embedding tensor. (B, 2 * T_1 - 1, size)
+            mask: Source mask. (B, T_2)
+            chunk_mask: Chunk mask. (T_1, T_1)
+            left_context: Number of frames in left context.
+        Returns:
+            : Output tensor. (B, T_1, H * d_k)
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        scores = self.compute_att_score(q, k, pos_enc, left_context=left_context)
+        return self.forward_attention(v, scores, mask, chunk_mask=chunk_mask)
+
+
+class CosineDistanceAttention(nn.Module):
+    """ Compute Cosine Distance between spk decoder output and speaker profile 
+    Args:
+        profile_path: speaker profile file path (.npy file)
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, spk_decoder_out, profile, profile_lens=None):
+        """
+        Args:
+            spk_decoder_out(torch.Tensor):(B, L, D)
+            spk_profiles(torch.Tensor):(B, N, D)
+        """
+        x = spk_decoder_out.unsqueeze(2)  # (B, L, 1, D)
+        if profile_lens is not None:
+            
+            mask = (make_pad_mask(profile_lens)[:, None, :]).to(profile.device)
+            min_value = float(
+                numpy.finfo(torch.tensor(0, dtype=x.dtype).numpy().dtype).min
+            )
+            weights_not_softmax=F.cosine_similarity(x, profile.unsqueeze(1), dim=-1).masked_fill(mask, min_value)
+            weights = self.softmax(weights_not_softmax).masked_fill(mask, 0.0)  # (B, L, N)
+        else:
+            x = x[:, -1:, :, :]
+            weights_not_softmax=F.cosine_similarity(x, profile.unsqueeze(1).to(x.device), dim=-1)
+            weights = self.softmax(weights_not_softmax)  # (B, 1, N)
+        spk_embedding = torch.matmul(weights, profile.to(weights.device))  # (B, L, D)
+
+        return spk_embedding, weights

--
Gitblit v1.9.1