fix: make parameter configurable for different architectures

AstraZeneca · Oct 24, 2024 · b560cf0 · b560cf0
1 parent ed750c3
commit b560cf0
Show file tree

Hide file tree

Showing 4 changed files with 30 additions and 20 deletions.
diff --git a/kazu/training/config.py b/kazu/training/config.py
@@ -24,5 +24,6 @@ class TrainingConfig:
     test_overfit: bool
     device: str
     workers: int
+    keys_to_use: list[str]
     architecture: str = "bert"
     epoch_completion_fraction_before_evals: float = 0.75
diff --git a/kazu/training/train_multilabel_ner.py b/kazu/training/train_multilabel_ner.py
@@ -306,6 +306,7 @@ def __init__(
         working_dir: Path,
         summary_writer: Optional[SummaryWriter] = None,
         ls_wrapper: Optional[LSManagerViewWrapper] = None,
+        keys_to_use: Optional[list[str]] = None,
     ):
 
         self.ls_wrapper = ls_wrapper
@@ -321,6 +322,9 @@ def __init__(
         self.eval_dataset = eval_dataset
         self.label_list = label_list
         self.pretrained_model_name_or_path = pretrained_model_name_or_path
+        self.keys_to_use = (
+            keys_to_use if keys_to_use else ["input_ids", "attention_mask", "token_type_ids"]
+        )
 
     def _write_to_tensorboard(
         self, global_step: int, main_tag: str, tag_scalar_dict: dict[str, NumericMetric]
@@ -441,7 +445,7 @@ def _process_docs(self, model: PreTrainedModel) -> list[Document]:
                 tokenized_word_processor=TokenizedWordProcessor(
                     labels=self.label_list, use_multilabel=True
                 ),
-                keys_to_use=["input_ids", "attention_mask", "token_type_ids"],
+                keys_to_use=self.keys_to_use,
                 device=self.training_config.device,
             )
 

diff --git a/kazu/training/train_script.py b/kazu/training/train_script.py
@@ -126,6 +126,7 @@ def run(cfg: DictConfig) -> None:
         working_dir=output_dir,
         summary_writer=SummaryWriter(log_dir=str(tb_path.absolute())),
         ls_wrapper=wrapper,
+        keys_to_use=training_config.keys_to_use,
     )
     trainer.train_model()
 

diff --git a/scripts/examples/conf/multilabel_ner_training/default.yaml b/scripts/examples/conf/multilabel_ner_training/default.yaml
@@ -20,27 +20,31 @@ training_config:
   test_overfit: false
   device: mps
   architecture: bert
+  keys_to_use: #distilbert for token classification doesn't use token_type_ids
+    - input_ids
+    - attention_mask
+    - token_type_ids
   workers: 2
 css_colors:
-  - '#000000'  # Black
-  - '#FF0000'  # Red
-  - '#00FF00'  # Lime
-  - '#0000FF'  # Blue
-  - '#FFFF00'  # Yellow
-  - '#00FFFF'  # Cyan
-  - '#FF00FF'  # Magenta
-  - '#800000'  # Maroon
-  - '#808000'  # Olive
-  - '#008000'  # Green
-  - '#800080'  # Purple
-  - '#008080'  # Teal
-  - '#FFA500'  # Orange
-  - '#A52A2A'  # Brown
-  - '#8A2BE2'  # BlueViolet
-  - '#5F9EA0'  # CadetBlue
-  - '#D2691E'  # Chocolate
-  - '#000080'  # Navy
-  - '#FFDAB9'  # PeachPuff
+  - "#000000" # Black
+  - "#FF0000" # Red
+  - "#00FF00" # Lime
+  - "#0000FF" # Blue
+  - "#FFFF00" # Yellow
+  - "#00FFFF" # Cyan
+  - "#FF00FF" # Magenta
+  - "#800000" # Maroon
+  - "#808000" # Olive
+  - "#008000" # Green
+  - "#800080" # Purple
+  - "#008080" # Teal
+  - "#FFA500" # Orange
+  - "#A52A2A" # Brown
+  - "#8A2BE2" # BlueViolet
+  - "#5F9EA0" # CadetBlue
+  - "#D2691E" # Chocolate
+  - "#000080" # Navy
+  - "#FFDAB9" # PeachPuff
 label_studio_manager:
   _target_: kazu.annotation.label_studio.LabelStudioManager
   project_name: "clean_data"