fix: revert model and make SOTA model available to COSTEER (#351)

* revert model and make SOTA model available to COSTEER

* fix CI

* fix CI

* fix CI

rdagent/components/coder/model_coder/CoSTEER/__init__.py

@@ -70,7 +70,7 @@ class ModelCoSTEER(Developer[ModelExperiment]):
         self.rag = ModelRAGStrategy(model_knowledge_base)
 
         # init intermediate items
-        model_experiment = ModelEvolvingItem(sub_tasks=exp.sub_tasks)
+        model_experiment = ModelEvolvingItem.from_experiment(exp)
 
         self.evolve_agent = ModelRAGEvoAgent(
             max_loop=self.max_loop,

rdagent/components/coder/model_coder/CoSTEER/evolvable_subjects.py

@@ -27,3 +27,10 @@ class ModelEvolvingItem(ModelExperiment, EvolvableSubjects):
             )
         else:
             self.sub_gt_implementations = sub_gt_implementations
+
+    @classmethod
+    def from_experiment(cls, exp: ModelExperiment) -> "ModelEvolvingItem":
+        ei = cls(sub_tasks=exp.sub_tasks)
+        ei.based_experiments = exp.based_experiments
+        ei.experiment_workspace = exp.experiment_workspace
+        return ei

rdagent/components/coder/model_coder/CoSTEER/evolving_strategy.py

@@ -30,23 +30,31 @@ class ModelCoderEvolvingStrategy(EvolvingStrategy):
         self,
         target_task: ModelTask,
         queried_knowledge: ModelQueriedKnowledge = None,
-        exp: ModelExperiment = None,  # Add this parameter
+        current_exp: ModelExperiment = None,  # Add this parameter
     ) -> str:
         model_information_str = target_task.get_task_information()
         model_type = target_task.model_type
 
-        # Get the current code from the experiment using build_from_SOTA
-        current_code = ""
-        if exp is not None:
-            self.build_from_SOTA(exp)
-            model_file_mapping = {
-                "XGBoost": "model_xgb.py",
-                "RandomForest": "model_rf.py",
-                "LightGBM": "model_lgb.py",
-                "NN": "model_nn.py",
-            }
-            if model_type in model_file_mapping:
-                current_code = exp.experiment_workspace.code_dict.get(model_file_mapping[model_type], "")
+        if len(current_exp.based_experiments) == 0:
+            current_code = None
+        else:
+            current_code = ""
+            sota_exp_code_dict = current_exp.based_experiments[-1].experiment_workspace.code_dict
+            if target_task.version == 2:
+                model_file_mapping = {
+                    "XGBoost": "model/model_xgboost.py",
+                    "RandomForest": "model/model_randomforest.py",
+                    "LightGBM": "model/model_lightgbm.py",
+                    "NN": "model/model_nn.py",
+                }
+                if model_type in model_file_mapping:
+                    current_code = sota_exp_code_dict.get(model_file_mapping[model_type], None)
+                elif "model.py" in sota_exp_code_dict:
+                    current_code = sota_exp_code_dict["model.py"]
+                else:
+                    current_code = None
+            elif target_task.version == 1:
+                current_code = sota_exp_code_dict.get("model.py", None)
 
         if queried_knowledge is not None and model_information_str in queried_knowledge.success_task_to_knowledge_dict:
             return queried_knowledge.success_task_to_knowledge_dict[model_information_str].implementation
@@ -74,7 +82,7 @@ class ModelCoderEvolvingStrategy(EvolvingStrategy):
                 .render(
                     scenario=self.scen.get_scenario_all_desc(),
                     queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
-                    current_code=current_code,  # Add this line
+                    current_code=current_code,
                 )
             )
 
@@ -87,7 +95,6 @@ class ModelCoderEvolvingStrategy(EvolvingStrategy):
                     )
                     .render(
                         model_information_str=model_information_str,
-                        model_type=model_type,  # Add model type to the prompt
                         queried_similar_successful_knowledge=queried_similar_successful_knowledge_to_render,
                         queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
                     )
@@ -124,7 +131,7 @@ class ModelCoderEvolvingStrategy(EvolvingStrategy):
         queried_knowledge: ModelQueriedKnowledge | None = None,
         **kwargs,
     ) -> ModelEvolvingItem:
-        # 1. Find the models that need to be evolved
+        # 1.找出需要evolve的model
         to_be_finished_task_index = []
         for index, target_model_task in enumerate(evo.sub_tasks):
             target_model_task_desc = target_model_task.get_task_information()
@@ -140,7 +147,7 @@ class ModelCoderEvolvingStrategy(EvolvingStrategy):
 
         result = multiprocessing_wrapper(
             [
-                (self.implement_one_model, (evo.sub_tasks[target_index], queried_knowledge))
+                (self.implement_one_model, (evo.sub_tasks[target_index], queried_knowledge, evo))
                 for target_index in to_be_finished_task_index
             ],
             n=RD_AGENT_SETTINGS.multi_proc_n,

rdagent/components/coder/model_coder/prompts.yaml

@@ -52,9 +52,14 @@ evolving_strategy_model_coder:
 
         Your must write your code based on your former latest attempt below which consists of your former code and code feedback, you should read the former attempt carefully and must not modify the right part of your former code.
 
-        {% if current_code %}
-        --------------Current code in the workspace:--------------- You need to tune the model based on this! If it is not None, do not write from scratch. 
+        {% if current_code is not none %}
+        User has write some code before. You should write the new code based on this code. Here is the latest code:
+        ```python
         {{ current_code }}
+        ```
+        Your code should be very similar to the former code which means your code should be ninety more percent same as the former code! You should not modify the right part of the code.
+        {% else %}
+        User has not write any code before. You should write the new code from scratch.
         {% endif %}
 
         {% if queried_former_failed_knowledge|length != 0 %}