core\refac: #98 oxford pet prefetch optimization

- optimized dataset prefetching and properly display best hps in optuna
  tuning

Author: blotero

core/seg_tgce/data/oxford_pet/oxford_pet.py

@@ -77,7 +77,7 @@ def get_data_multiple_annotators(
         )
         for data, labeler_manager in (
             (
-                train_dataset.cache().shuffle(1000).repeat(2).prefetch(tf.data.AUTOTUNE),
+                train_dataset.shuffle(1000).prefetch(tf.data.AUTOTUNE),
                 train_labeler_manager,
             ),
             (val_dataset.prefetch(tf.data.AUTOTUNE), None),

core/seg_tgce/experiments/pets/scalar.py

@@ -13,15 +13,15 @@
 from seg_tgce.models.builders import build_scalar_model_from_hparams
 from seg_tgce.models.ma_model import ScalarVisualizationCallback
 
-TARGET_SHAPE = (64, 64)
+TARGET_SHAPE = (128, 128)
 BATCH_SIZE = 16
 NUM_CLASSES = 3
 NOISE_LEVELS = [-20.0, 10.0]
 NUM_SCORERS = len(NOISE_LEVELS)
 TRAIN_EPOCHS = 50
-TUNER_EPOCHS = 10
+TUNER_EPOCHS = 5
 LABELING_RATE = 0.5
-TUNER_MAX_TRIALS = 10
+TUNER_MAX_TRIALS = 3
 STUDY_NAME = "pets_scalar_tuning"
 OBJECTIVE = "val_segmentation_output_dice_coefficient"
 DEFAULT_HPARAMS = {
@@ -78,8 +78,8 @@ def build_model_from_trial(trial: HpTunerTrial | None) -> Model:
     )
 
     model = handle_training_optuna(
-        train,
-        val,
+        train.take(10).cache(),
+        val.take(10).cache(),
         model_builder=build_model_from_trial,
         use_tuner=args.use_tuner,
         tuner_epochs=TUNER_EPOCHS,

core/seg_tgce/experiments/utils.py

@@ -36,7 +36,7 @@ def train_with_tuner(train_gen: tf.data.Dataset, val_gen: tf.data.Dataset) -> Mo
 
         print("Starting hyperparameter search...")
         tuner.search(
-            train_gen.take(10),
+            train_gen,
             epochs=tuner_epochs,
             validation_data=val_gen,
         )
@@ -134,7 +134,7 @@ def _objective(trial: optuna.Trial) -> float:
 
         best_hps = study.best_trial
         print("\nBest hyperparameters:")
-        for param, value in best_hps.values.items():
+        for param, value in best_hps.params.items():
             print(f"{param}: {value}")
 
         create_importance_visualizations(study)