diff --git a/codes/scripts/audio/preparation/phase3_generate_similarities.py b/codes/scripts/audio/preparation/phase3_generate_similarities.py
index 67e87a6a..5b4d1772 100644
--- a/codes/scripts/audio/preparation/phase3_generate_similarities.py
+++ b/codes/scripts/audio/preparation/phase3_generate_similarities.py
@@ -1,6 +1,7 @@
 import argparse
 import functools
 import os
+import sys
 from multiprocessing.pool import ThreadPool
 
 import torch
@@ -54,13 +55,17 @@ def process_subdir(subdir, options, clip_sz):
 
         clips = []
         for path in paths:
-            clip = load_audio(str(path), 22050)
-            padding = clip_sz - clip.shape[1]
-            if padding > 0:
-                clip = F.pad(clip, (0, padding))
-            elif padding < 0:
-                clip = clip[:, :clip_sz]
-            clips.append(clip)
+            try:
+                clip = load_audio(str(path), 22050)
+                padding = clip_sz - clip.shape[1]
+                if padding > 0:
+                    clip = F.pad(clip, (0, padding))
+                elif padding < 0:
+                    clip = clip[:, :clip_sz]
+                clips.append(clip)
+            except:
+                print(f"Error processing {path}. Recovering gracefully.")
+                print(sys.exc_info())
         sims = None
         while len(clips) > 0:
             stacked = torch.stack(clips[:256], dim=0).cuda()
@@ -101,7 +106,7 @@ if __name__ == '__main__':
     """
     parser = argparse.ArgumentParser()
     parser.add_argument('-o', type=str, help='Path to the options YAML file used to train the CLIP model', default='../options/train_voice_voice_clip.yml')
-    parser.add_argument('--num_workers', type=int, help='Number concurrent processes to use', default=6)
+    parser.add_argument('--num_workers', type=int, help='Number concurrent processes to use', default=2)
     parser.add_argument('--root_path', type=str, help='Root path to search for audio directories from', default='Y:\\filtered\\big_podcast')
     parser.add_argument('--clip_size', type=int, help='Amount of audio samples to pull from each file', default=22050)
     args = parser.parse_args()
diff --git a/codes/trainer/eval/flow_gaussian_nll.py b/codes/trainer/eval/flow_gaussian_nll.py
index 42d72f36..c08f0e37 100644
--- a/codes/trainer/eval/flow_gaussian_nll.py
+++ b/codes/trainer/eval/flow_gaussian_nll.py
@@ -6,7 +6,7 @@ import trainer.eval.evaluator as evaluator
 
 # Evaluate how close to true Gaussian a flow network predicts in a "normal" pass given a LQ/HQ image pair.
 from data.images.image_folder_dataset import ImageFolderDataset
-from models.image_generation.srflow import GaussianDiag
+from models.image_generation.srflow.flow import GaussianDiag
 
 
 class FlowGaussianNll(evaluator.Evaluator):
diff --git a/codes/trainer/eval/single_point_pair_contrastive_eval.py b/codes/trainer/eval/single_point_pair_contrastive_eval.py
index 2b16e509..c5b06fa2 100644
--- a/codes/trainer/eval/single_point_pair_contrastive_eval.py
+++ b/codes/trainer/eval/single_point_pair_contrastive_eval.py
@@ -6,7 +6,6 @@ from tqdm import tqdm
 import trainer.eval.evaluator as evaluator
 
 from data.images.image_pair_with_corresponding_points_dataset import ImagePairWithCorrespondingPointsDataset
-from models.segformer.segformer import Segformer
 
 
 # Uses two datasets: a "similar" and "dissimilar" dataset, each of which contains pairs of images and similar/dissimilar