debugv3

Author: jingyuanlm

rdagent/scenarios/data_science/proposal/exp_gen/prompts_v2.yaml

@@ -147,7 +147,7 @@ scenario_description: |-
   ====== Time Limit ======
   Your code's execution is limited to {{ time_limit }}. After this time limit, your code will be terminated. However, remember that your primary objective is to achieve the best possible performance, and you're allowed to revise your code multiple times. So be bold — make full use of the entire time limit, and don’t be too conservative.
   During this period, you have full access to computational resources (CPUs and GPUs). Please take advantage of them: choose powerful models, use large batch sizes, and enable parallelism (e.g., large-scale data sampling or multi-GPU training) where applicable.
-  If your code involves ensemble training, note that the total time allowed for ensemble runs is {{ ensemble_limit }}. Make sure to plan your ensemble strategy wisely within this limit.
+  If your code involves ensemble training, note that the total time allowed for ensemble runs is {{ ensemble_limit }} seconds. Make sure to plan your ensemble strategy wisely within this limit.
   {% endif %}
 
 hypothesis_gen:

rdagent/scenarios/data_science/proposal/exp_gen/proposal.py

@@ -731,15 +731,16 @@ def hypothesis_select_with_llm(self,
         # res_time = 12*3600 - time_use_current
         res_time = RD_Agent_TIMER_wrapper.timer.remain_time()
         total_time = RD_Agent_TIMER_wrapper.timer.all_duration
-        use_time = total_time.seconds - res_time.seconds
-        use_ratio = 100* use_time // total_time.seconds 
+        use_time =  round(total_time.total_seconds(),2) -  round(res_time.total_seconds(),2)
+        use_ratio = 100* use_time / round(total_time.total_seconds(),2)
+        use_ratio = round(use_ratio, 2)
 
         ensemble_timeout = DS_RD_SETTING.ensemble_timeout
         hypothesis_candidates =  str(json.dumps(hypothesis_candidates, indent=2))
 
         sys_prompt = T(".prompts_v2:hypothesis_select.system").r(
                 hypothesis_candidates = hypothesis_candidates,
-                res_time = res_time.seconds,
+                res_time = round(res_time.total_seconds(),2),
                 ensemble_timeout = ensemble_timeout,
                 use_ratio = use_ratio,
                 hypothesis_output_format = T(".prompts_v2:output_format.hypothesis_select_format").r(hypothesis_candidates = hypothesis_candidates)

rdagent/scenarios/data_science/scen/__init__.py

@@ -140,6 +140,7 @@ def get_competition_full_desc(self) -> str:
             metric_direction=self.metric_direction,
             raw_description=self.raw_description,
             use_raw_description=DS_RD_SETTING.use_raw_description,
+            ensemble_limit = DS_RD_SETTING.ensemble_timeout,
             time_limit=None,
             eda_output=None,
         )
@@ -157,6 +158,7 @@ def get_scenario_all_desc(self, eda_output=None) -> str:
             raw_description=self.raw_description,
             use_raw_description=DS_RD_SETTING.use_raw_description,
             time_limit=f"{DS_RD_SETTING.full_timeout / 60 / 60 : .2f} hours",
+            ensemble_limit = DS_RD_SETTING.ensemble_timeout,
             eda_output=eda_output,
         )
 

rdagent/scenarios/data_science/scen/prompts.yaml

@@ -23,6 +23,7 @@ scenario_description: |-
   {% if time_limit %}------The time limit to your code------
   You code running is limit to {{ time_limit }}, after this time limit, your code will be terminated. But remember your main target is to achieve the best performance and you have several times to modify your code. So please be bold to make the best use of all the time limit and don't be too conservative.
   During this time limit, you have all the resources available to you. Please fully leverage all the computational resources(CPUs and GPUs) to achieve the best performance like choose a powerful model, use a large batch size, enable data sampler with big parallel.
+  If your code involves ensemble training, note that the total time allowed for ensemble runs is {{ ensemble_limit }} seconds. Make sure to plan your ensemble strategy wisely within this limit.
   {% endif %}
   {% if evaluation is not none %}------Evaluation------
   {{ evaluation }}