Mar25

examples/examples_prompt/requirements.txt

@@ -0,0 +1 @@
+optuna

examples/examples_prompt/run.py

@@ -280,6 +280,7 @@ def main():
                 load_from_cache_file=not data_args.overwrite_cache,
             )
         train_dataset = concatenate_datasets(train_datasets)
+        print(f"Train dataset size {len(train_dataset)}")
    
     if training_args.do_eval:
         eval_datasets = {eval_dataset: AutoTask.get(eval_dataset, eval_dataset_config,

examples/examples_prompt/search_distributed.py

@@ -18,6 +18,7 @@ if __name__=="__main__":
     parser.add_argument("--continue_study", type=bool, default=False)
     parser.add_argument("--substudy_prefix", type=str, default="")
     parser.add_argument("--num_trials", type=int)
+    parser.add_argument("--pathbase", type=str, default="")
     args = parser.parse_args()
 
 
@@ -27,7 +28,7 @@ if __name__=="__main__":
     else:
         args.study_name += pardir
     
-    setattr(args, "output_dir", f"outputs_search/{args.study_name}")
+    setattr(args, "output_dir", f"{args.pathbase}/outputs_search/{args.study_name}")
 
     
     
@@ -68,7 +69,7 @@ if __name__=="__main__":
             else:
                 sub_n_trials = args.num_trials//tot_chunk_num + args.num_trials%tot_chunk_num
 
-            command = "nohup python search_single.py "
+            command = "python search_single.py "
             command += f"--cuda_id {cudas} "
             command += f"--model_name {args.model_name} "
             command += f"--dataset {args.dataset} "
@@ -76,7 +77,7 @@ if __name__=="__main__":
             command += f"--study_name {args.study_name} "
             command += f"--optuna_seed 10{id} "
             command += f"--num_trials {sub_n_trials} "
-            command += f">{args.output_dir}/{args.substudy_prefix}{id}.log 2>&1 &"
+            command += f"> {args.output_dir}/{args.substudy_prefix}{id}.log 2>&1 &"
             p = subprocess.Popen(command, cwd="./", shell=True)
             print("id {} on cuda:{}, pid {}\n {}\n".format(id, cudas, p.pid, command))
 

examples/examples_prompt/search_single.py

@@ -20,11 +20,14 @@ def objective_singleseed(args, unicode, search_space_sample  ):
     with open(f"{args.output_dir}/{unicode}/this_configs.json", 'w') as fout:
         json.dump(search_space_sample, fout, indent=4,sort_keys=True)
     
+
     command = "CUDA_VISIBLE_DEVICES={} ".format(args.cuda_id)
     command += "python run.py "
     command += f"{args.output_dir}/{unicode}/this_configs.json"
+    command += f" >> {args.output_dir}/{unicode}/output.log 2>&1"
     
 
+    print("======"*5+"\n"+command)
     status_code = os.system(command)
     print("status_code",status_code)
     # if status_code != 0:
@@ -61,7 +64,7 @@ def objective(trial, args=None):
     search_space_sample.update(DatasetSearchSpace(args.dataset).get_config(trial, args))
     search_space_sample.update(AllDeltaSearchSpace[args.delta_type]().get_config(trial, args))
     results = []
-    for seed in [100]:
+    for seed in [42,43,44]:
         search_space_sample.update({"seed": seed})
         unicode = random.randint(0, 100000000)
         while os.path.exists(f"{args.output_dir}/{unicode}"):
@@ -84,10 +87,11 @@ if __name__=="__main__":
     parser.add_argument("--study_name")
     parser.add_argument("--num_trials", type=int)
     parser.add_argument("--optuna_seed", type=int, default="the seed to sample suggest point")
+    parser.add_argument("--pathbase", type=str, default="")
     args = parser.parse_args()
 
         
-    setattr(args, "output_dir", f"outputs_search/{args.study_name}")
+    setattr(args, "output_dir", f"{args.pathbase}/outputs_search/{args.study_name}")
 
     study = optuna.load_study(study_name=args.study_name, storage=f'sqlite:///{args.study_name}.db', sampler=TPESampler(seed=args.optuna_seed))
     study.optimize(partial(objective, args=args), n_trials=args.num_trials)