fix attn can not train

paddlespeech/s2t/modules/attention.py

@@ -60,6 +60,7 @@ class MultiHeadedAttention(nn.Layer):
         super()._build_once(*args, **kwargs)
         # if self.self_att:
         # self.linear_kv = Linear(self.n_feat, self.n_feat*2)
+        if not self.training:
             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])
@@ -86,8 +87,10 @@ class MultiHeadedAttention(nn.Layer):
         n_batch = query.shape[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)
+        if self.training:
+            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)
+        else:
             k, v = F.linear(key, self.weight, self.bias).view(
                 n_batch, -1, 2 * self.h, self.d_k).split(
                     2, axis=2)