tune tuning script

nnaakkaaii · Jun 16, 2024 · deabae5 · deabae5
1 parent 92510b5
commit deabae5
Show file tree

Hide file tree

Showing 2 changed files with 24 additions and 18 deletions.
diff --git a/.gitignore b/.gitignore
@@ -2,6 +2,11 @@ outputs/
 results/
 data/
 playground/
+logs/
+scripts/
+
+*.out
+*.d000*
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

diff --git a/tuning/mmnist.py b/tuning/mmnist.py
@@ -62,7 +62,7 @@ def objective(trial):
     }
 
     dataloader_option = BasicDataLoaderOption(
-        batch_size=96,
+        batch_size=256,
         train_val_ratio=0.8,
         dataset=dataset_option,
         transform_order_train=["min_max_normalization"],
@@ -94,22 +94,22 @@ def objective(trial):
         "rnn1d",
         "tsn1d",
     ]:
-        if phase == "all":
-            pred_diff = False
-        else:
-            pred_diff = trial.suggest_categorical("pred_diff", [True, False])
+        # if phase == "all":
+        #     pred_diff = False
+        # else:
+        #     pred_diff = trial.suggest_categorical("pred_diff", [True, False])
         motion_encoder_name = args.motion_encoder_name
     elif phase == "all":
-        pred_diff = False
+        # pred_diff = False
         motion_encoder_name = trial.suggest_categorical(
             "motion_encoder_all", ["conv2d", "normal1d", "rnn1d"]
         )
     else:
-        pred_diff = trial.suggest_categorical("pred_diff", [True, False])
+        # pred_diff = trial.suggest_categorical("pred_diff", [True, False])
         motion_encoder_name = trial.suggest_categorical(
             "motion_encoder", ["conv2d", "guided1d", "normal1d", "rnn1d", "tsn1d"]
         )
-    motion_encoder_num_layers = trial.suggest_int("motion_encoder_num_layers", 0, 3)
+    motion_encoder_num_layers = trial.suggest_int("motion_encoder_num_layers", 0, 6)
     if motion_encoder_name == "rnn1d":
         if args.rnn_name in [
             "conv_lstm1d",
@@ -122,18 +122,19 @@ def objective(trial):
                 "rnn", ["conv_lstm1d", "gru1d", "tcn1d"]
             )
         motion_encoder_name = f"{motion_encoder_name}/{rnn_name}"
+        rnn_num_layers = trial.suggest_int("rnn_num_layers", 1, 5)
         if rnn_name == "conv_lstm1d":
             rnn_option = ConvLSTM1dOption(
-                num_layers=trial.suggest_int("rnn_num_layers", 2, 4),
+                num_layers=rnn_num_layers,
             )
         elif rnn_name == "gru1d":
             rnn_option = GRU1dOption(
-                num_layers=trial.suggest_int("rnn_num_layers", 2, 4),
+                num_layers=rnn_num_layers,
                 image_size=8,
             )
         elif rnn_name == "tcn1d":
             rnn_option = TCN1dOption(
-                num_layers=trial.suggest_int("rnn_num_layers", 2, 4),
+                num_layers=rnn_num_layers,
                 image_size=8,
                 kernel_size=3,
                 dropout=0.1,
@@ -187,9 +188,9 @@ def objective(trial):
         )
     else:
         raise RuntimeError("unreachable")
-    latent_dim = trial.suggest_int("latent_dim", 16, 64, step=8)
+    latent_dim = trial.suggest_int("latent_dim", 16, 96, step=8)
     content_encoder_num_layers = trial.suggest_int(
-        "content_encoder_encoder_num_layers", 0, 3
+        "content_encoder_num_layers", 0, 4
     )
     if network_name == "rae2d":
         network_option = RAE2dOption(
@@ -234,17 +235,17 @@ def objective(trial):
         raise RuntimeError("unreachable")
 
     optimizer_option = AdamOptimizerOption(
-        lr=trial.suggest_float("lr", 1e-5, 1e-2, log=True),
+        lr=trial.suggest_float("lr", 1e-5, 5e-2, log=True),
     )
 
     scheduler_option = OneCycleLRSchedulerOption(
-        max_lr=trial.suggest_float("max_lr", 1e-3, 1e-2, log=True),
+        max_lr=trial.suggest_float("max_lr", 1e-3, 5e-2, log=True),
     )
 
     model_option = VRModelOption(
         loss_coef={"wmse": 1.0},
         phase=phase,
-        pred_diff=pred_diff,
+        pred_diff=False,
         loss=loss_option,
         network=network_option,
         optimizer=optimizer_option,
@@ -258,7 +259,7 @@ def objective(trial):
         result_dir=result_dir,
         dataloader=dataloader_option,
         model=model_option,
-        n_epoch=50,
+        n_epoch=150,
     )
     result_dir.mkdir(parents=True, exist_ok=True)
     with open(result_dir / "config.json", "w") as f:
@@ -288,7 +289,7 @@ def objective(trial):
     import optuna
 
     parser = argparse.ArgumentParser()
-    parser.add_argument("--network_name", type=str)
+    parser.add_argument("--network_name", type=str, default="hrdae2d")
     parser.add_argument("--motion_encoder_name", type=str)
     parser.add_argument("--rnn_name", type=str)
     args = parser.parse_args()