Added config option to hpo script, styling (#235)

use-cases/eurac/hpo.py

@@ -1,6 +1,5 @@
 import argparse
 import os
-from pathlib import Path
 from typing import Dict
 
 import matplotlib.pyplot as plt

use-cases/eurac/slurm_ray.sh

@@ -3,7 +3,7 @@
 # Job configuration
 #SBATCH --job-name=ray_tune_hpo    
 #SBATCH --account=intertwin       
-#SBATCH --time 01:00:00
+#SBATCH --time 02:30:00
 
 # Resources allocation
 #SBATCH --cpus-per-task=24
@@ -88,7 +88,7 @@ echo All Ray workers started.
 # Run the Python script using Ray
 echo 'Starting HPO.'
 
-python hpo.py --num_samples 8 --max_iterations 2 --ngpus $num_gpus --ncpus $num_cpus
+python hpo.py --num_samples 4 --max_iterations 2 --ngpus $num_gpus --ncpus $num_cpus --pipeline_name rnn_training_pipeline # NOTE: conv_training_pipeline has not been tested
 
 # Shutdown Ray after completion
 ray stop

use-cases/eurac/trainer.py

@@ -213,6 +213,7 @@ def train(self):
             self.lr_scheduler.step(avg_val_loss)
             loss_history["train"].append(train_loss)
             loss_history["val"].append(avg_val_loss)
+
             self.log(
                 item=train_loss.item(),
                 identifier="train_loss_per_epoch",
@@ -496,6 +497,12 @@ def train(self):
                 best_loss = avg_val_loss
                 # self.model.load_state_dict(best_model_weights)
 
+            # Report training metrics of last epoch to Ray
+            train.report(
+                {"loss": avg_val_loss.item(),
+                 "train_loss": train_loss.item()}
+            )
+
         return loss_history, metric_history
 
     def create_dataloaders(self, train_dataset, validation_dataset, test_dataset):