correction

src/conf/model.yaml

@@ -14,7 +14,7 @@ models:
     model_name: "bert-base-uncased"
     num_labels: 2
     learning_rate: 5e-5
-    epochs: 3
+    epochs: 8
     batch_size: 12
     warmup_steps: 0
     weight_decay: 0.0
@@ -25,7 +25,7 @@ models:
     model_name: "roberta-base"
     num_labels: 2
     learning_rate: 5e-5
-    epochs: 3
+    epochs: 8
     batch_size: 12
     warmup_steps: 0
     weight_decay: 0.0

src/mlflow_integration/mlflow_decorator.py

@@ -49,14 +49,6 @@ class MLflowDecorator:
         Gets the relevant parameters from the trainer's config.
         This replaces the singledispatch logic previously in parameter_logging.py
         """
-        # In this specific project structure, all trainers seem to store
-        # the relevant hyperparameters directly in trainer_instance.config.model.params.
-        # If specific trainers needed different logic, we could add isinstance checks here.
-        # from trainers.cuml.svm_trainer import SvmTrainer # Example import if needed
-        # if isinstance(trainer_instance, SvmTrainer):
-        #    return specific_logic_for_svm(trainer_instance)
-        
-        # Default logic: return the model params from the config
         if hasattr(trainer_instance, 'config') and hasattr(trainer_instance.config, 'model') and hasattr(trainer_instance.config.model, 'params'):
              return trainer_instance.config.model.params
         return {}

src/trainers/huggingface/huggingface_transformer_trainer.py

@@ -143,7 +143,8 @@ class HuggingFaceTransformerTrainer(BaseTrainer):
             train_dataset=self.train_dataset, # Sera défini dans train()
             eval_dataset=self.eval_dataset,   # Sera défini dans train()
             compute_metrics=compute_metrics,
-            tokenizer=self.tokenizer # Ajout du tokenizer pour le padding dynamique si besoin
+            tokenizer=self.tokenizer, # Ajout du tokenizer pour le padding dynamique si besoin
+            callbacks=[]
         )
 
     def train(self) -> None:
@@ -264,11 +265,11 @@ class HuggingFaceTransformerTrainer(BaseTrainer):
             learning_rate=float(params.get("learning_rate")),
             warmup_steps=int(params.get("warmup_steps")),
             weight_decay=float(params.get("weight_decay")),
-            evaluation_strategy="epoch",
-            save_strategy="epoch",
+            save_steps=50,
             logging_dir="./logs",
-            logging_steps=10
-            # Ne pas désactiver les rapports MLflow intégrés
+            logging_strategy="no",
+            save_strategy="epoch",
+            report_to="mlflow"
         )
         
     def optimize_if_needed(self) -> None: