Fix typos in LLARA

research/LLARA/README.md

@@ -163,7 +163,7 @@ run.py \
 --num_train_epochs 1 \
 --per_device_train_batch_size 1 \
 --dataloader_drop_last True \
---normlized True \
+--normalized True \
 --temperature 0.01 \
 --query_max_len 64 \
 --passage_max_len 160 \

research/LLARA/finetune/arguments.py

@@ -160,6 +160,6 @@ class RetrieverTrainingArguments(TrainingArguments):
     temperature: Optional[float] = field(default=0.02)
     fix_position_embedding: bool = field(default=False, metadata={"help": "Freeze the parameters of position embeddings"})
     sentence_pooling_method: str = field(default='cls', metadata={"help": "the pooling method, should be cls or mean"})
-    normlized: bool = field(default=True)
+    normalized: bool = field(default=True)
     sub_batch_size: int = field(default=None)
     cache_chunk_size: int = field(default=-1, metadata={"help": "用于缓存每一步的执行."})

research/LLARA/finetune/modeling.py

@@ -27,7 +27,7 @@ class BiEncoderModel(nn.Module):
     def __init__(self,
                  model: AutoModel = None,
                  tokenizer: AutoTokenizer = None,
-                 normlized: bool = False,
+                 normalized: bool = False,
                  negatives_cross_device: bool = False,
                  temperature: float = 1.0,
                  sub_batch_size: int = -1
@@ -38,9 +38,9 @@ def __init__(self,
         self.tokenizer = tokenizer
         self.cross_entropy = nn.CrossEntropyLoss(reduction='mean')
 
-        self.normlized = normlized
+        self.normalized = normalized
         self.temperature = temperature
-        if not normlized:
+        if not normalized:
             self.temperature = 1.0
             logger.info("reset temperature = 1.0 due to using inner product to compute similarity")
 
@@ -80,14 +80,14 @@ def encode(self, features):
                     p_reps = torch.mean(p_reps, dim=1)
                     all_p_reps.append(p_reps)
                 all_p_reps = torch.cat(all_p_reps, 0).contiguous()
-                if self.normlized:
+                if self.normalized:
                     all_p_reps = torch.nn.functional.normalize(all_p_reps, dim=-1)
                 return all_p_reps.contiguous()
             else:
                 psg_out = self.model(**features, return_dict=True, output_hidden_states=True)
                 p_reps = psg_out.hidden_states[-1][:, -8:, :]
                 p_reps = torch.mean(p_reps, dim=1)
-                if self.normlized:
+                if self.normalized:
                     p_reps = torch.nn.functional.normalize(p_reps, dim=-1)
                 return p_reps.contiguous()
         else:
@@ -99,7 +99,7 @@ def encode(self, features):
                 p_reps = torch.mean(p_reps, dim=1)
                 all_p_reps.append(p_reps)
             all_p_reps = torch.cat(all_p_reps, 0).contiguous()
-            if self.normlized:
+            if self.normalized:
                 all_p_reps = torch.nn.functional.normalize(all_p_reps, dim=-1)
             return all_p_reps.contiguous()
 

research/LLARA/finetune/run.py

@@ -80,7 +80,7 @@ def main():
 
     model = BiEncoderModel(model=base_model,
                            tokenizer=tokenizer,
-                           normlized=training_args.normlized,
+                           normalized=training_args.normalized,
                            negatives_cross_device=training_args.negatives_cross_device,
                            temperature=training_args.temperature,
                            sub_batch_size=training_args.sub_batch_size)

research/LLARA/finetune/trainer.py

@@ -23,7 +23,7 @@ def _save(self, output_dir: Optional[str] = None, state_dict=None):
         # if self.is_world_process_zero():
         #     save_ckpt_for_sentence_transformers(output_dir,
         #                                         pooling_mode=self.args.sentence_pooling_method,
-        #                                         normlized=self.args.normlized)
+        #                                         normalized=self.args.normalized)
 
     def compute_loss(self, model, inputs, return_outputs=False):
         """