From 580b11b57ac4b62f7e2acda73813a4e10e8e4cd3 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 10 十月 2023 17:17:29 +0800
Subject: [PATCH] v0.8.0
---
funasr/modules/attention.py | 45 ++++++++++++++++++++++++++++++++++++++++-----
1 files changed, 40 insertions(+), 5 deletions(-)
diff --git a/funasr/modules/attention.py b/funasr/modules/attention.py
index f5430e1..b007d58 100644
--- a/funasr/modules/attention.py
+++ b/funasr/modules/attention.py
@@ -471,15 +471,21 @@
"""
q_h, k_h, v_h, v = self.forward_qkv(x)
- if chunk_size is not None and look_back > 0:
+ 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"] = k_h[:, :, -(look_back * chunk_size[1]):, :]
- cache["v"] = v_h[:, :, -(look_back * chunk_size[1]):, :]
+
+ 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[:, :, -(look_back * chunk_size[1]):, :],
- "v": v_h[:, :, -(look_back * chunk_size[1]):, :]}
+ 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)
@@ -699,6 +705,35 @@
scores = torch.matmul(q_h, k_h.transpose(-2, -1))
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.
--
Gitblit v1.9.1