From 7ac54b3c97491ee9ac8a8ebbb7033240864f805a Mon Sep 17 00:00:00 2001
From: haoneng.lhn <haoneng.lhn@alibaba-inc.com>
Date: 星期二, 18 七月 2023 19:26:01 +0800
Subject: [PATCH] add lora finetune code
---
funasr/models/decoder/sanm_decoder.py | 6 +++++-
1 files changed, 5 insertions(+), 1 deletions(-)
diff --git a/funasr/models/decoder/sanm_decoder.py b/funasr/models/decoder/sanm_decoder.py
index d83f89f..c12e098 100644
--- a/funasr/models/decoder/sanm_decoder.py
+++ b/funasr/models/decoder/sanm_decoder.py
@@ -833,6 +833,10 @@
att_layer_num: int = 6,
kernel_size: int = 21,
sanm_shfit: int = 0,
+ lora_list: List[str] = None,
+ lora_rank: int = 8,
+ lora_alpha: int = 16,
+ lora_dropout: float = 0.1,
tf2torch_tensor_name_prefix_torch: str = "decoder",
tf2torch_tensor_name_prefix_tf: str = "seq2seq/decoder",
):
@@ -885,7 +889,7 @@
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
),
MultiHeadedAttentionCrossAtt(
- attention_heads, attention_dim, src_attention_dropout_rate
+ attention_heads, attention_dim, src_attention_dropout_rate, lora_list, lora_rank, lora_alpha, lora_dropout
),
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
dropout_rate,
--
Gitblit v1.9.1