pull/2425/head
Hui Zhang 2 years ago
parent 2a75405e9a
commit 925abcca23

@ -19,8 +19,8 @@ from typing import Tuple
import paddle
from paddle import nn
from paddle.nn import initializer as I
from paddle.nn import functional as F
from paddle.nn import initializer as I
from paddlespeech.s2t.modules.align import Linear
from paddlespeech.s2t.utils.log import Log
@ -56,12 +56,12 @@ class MultiHeadedAttention(nn.Layer):
self.linear_out = Linear(n_feat, n_feat)
self.dropout = nn.Dropout(p=dropout_rate)
def _build_once(self, *args, **kwargs):
super()._build_once(*args, **kwargs)
# if self.self_att:
# self.linear_kv = Linear(self.n_feat, self.n_feat*2)
self.weight = paddle.concat([self.linear_k.weight, self.linear_v.weight], axis=-1)
self.weight = paddle.concat(
[self.linear_k.weight, self.linear_v.weight], axis=-1)
self.bias = paddle.concat([self.linear_k.bias, self.linear_v.bias])
self._built = True
@ -88,7 +88,9 @@ class MultiHeadedAttention(nn.Layer):
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)
k, v = F.linear(key, self.weight, self.bias).view(n_batch, -1, 2 * self.h, self.d_k).split(2, axis=2)
k, v = F.linear(key, self.weight, self.bias).view(
n_batch, -1, 2 * self.h, self.d_k).split(
2, axis=2)
q = q.transpose([0, 2, 1, 3]) # (batch, head, time1, d_k)
k = k.transpose([0, 2, 1, 3]) # (batch, head, time2, d_k)

Loading…
Cancel
Save