making LoRA layer config changable

stanza/models/coref/config.py

@@ -4,7 +4,7 @@
 """
 
 from dataclasses import dataclass
-from typing import Dict
+from typing import Dict, List
 
 
 @dataclass
@@ -40,6 +40,9 @@ class Config:  # pylint: disable=too-many-instance-attributes, too-few-public-me
     lora_rank: int
     lora_dropout: float
 
+    lora_targets: List[str]
+    lora_fully_tune: List[str]
+
     bert_finetune: bool
     dropout_rate: float
     learning_rate: float

stanza/models/coref/coref_config.toml

@@ -71,6 +71,9 @@ lora = false
 lora_alpha = 128
 lora_dropout = 0.1
 lora_rank = 64
+lora_targets = []
+lora_fully_tune = []
+
 
 # Training settings ==================
 
@@ -118,6 +121,8 @@ bert_model = "roberta-large"
 bert_model = "roberta-large"
 bert_learning_rate = 0.00005
 lora = true
+lora_targets = [ "query", "value", "output.dense", "intermediate.dense" ]
+lora_fully_tune = [ "pooler" ]
 
 [roberta_no_finetune]
 bert_model = "roberta-large"

stanza/models/coref/model.py

@@ -80,11 +80,10 @@ def __init__(self,
             logger.debug("Creating lora adapter with rank %d", self.config.lora_rank)
             self.__peft_config = LoraConfig(inference_mode=False,
                                             r=self.config.lora_rank,
-                                            target_modules=["query", "value",
-                                                            "output.dense", "intermediate.dense"],
+                                            target_modules=self.config.lora_targets,
                                             lora_alpha=self.config.lora_alpha,
                                             lora_dropout=self.config.lora_dropout,
-                                            modules_to_save=["pooler"],
+                                            modules_to_save=self.config.lora_fully_tune,
                                             bias="none")
 
             self.bert = get_peft_model(self.bert, self.__peft_config)