From 8f26a9acc2461ce0c77eacc3d36d3cef3457f520 Mon Sep 17 00:00:00 2001
From: speech_asr <wangjiaming.wjm@alibaba-inc.com>
Date: 星期三, 29 三月 2023 15:49:46 +0800
Subject: [PATCH] Merge branch 'dev_wjm' of https://github.com/alibaba/FunASR into dev_wjm
---
funasr/modules/embedding.py | 13 +++++++++++--
1 files changed, 11 insertions(+), 2 deletions(-)
diff --git a/funasr/modules/embedding.py b/funasr/modules/embedding.py
index e4f9bff..79ca0b2 100644
--- a/funasr/modules/embedding.py
+++ b/funasr/modules/embedding.py
@@ -8,7 +8,7 @@
import math
import torch
-
+import torch.nn.functional as F
def _pre_hook(
state_dict,
@@ -409,9 +409,18 @@
def forward_chunk(self, x, cache=None):
start_idx = 0
+ pad_left = 0
+ pad_right = 0
batch_size, timesteps, input_dim = x.size()
if cache is not None:
start_idx = cache["start_idx"]
+ pad_left = cache["left"]
+ pad_right = cache["right"]
positions = torch.arange(1, timesteps+start_idx+1)[None, :]
position_encoding = self.encode(positions, input_dim, x.dtype).to(x.device)
- return x + position_encoding[:, start_idx: start_idx + timesteps]
+ outputs = x + position_encoding[:, start_idx: start_idx + timesteps]
+ outputs = outputs.transpose(1,2)
+ outputs = F.pad(outputs, (pad_left, pad_right))
+ outputs = outputs.transpose(1,2)
+ return outputs
+
--
Gitblit v1.9.1