init

main_s2s.sh

@@ -19,7 +19,13 @@ do
   python fetch_dataset_s2s.py
 done
 # tokenize
-for i in $(seq 1 144);
+for i in $(seq 1 30);
+do
+  export DATASET_ID=${i}
+  export DIRECTION="enA-jaA"
+  python tokenize_dataset_s2s.py
+done
+for i in $(seq 30 60);
 do
   export DATASET_ID=${i}
   export DIRECTION="enA-jaA"

tokenize_dataset_s2s.py

@@ -3,6 +3,7 @@ from os.path import expanduser
 
 import shutil
 import torch
+from soundfile import LibsndfileError
 from datasets import load_dataset, DatasetDict
 from encodec_audio_tokenizer import EncodecTokenizer
 
@@ -18,6 +19,20 @@ dataset = load_dataset(f"{hf_org}/{hf_dataset}", f"subset_{dataset_id}", split="
 tokenizer = EncodecTokenizer.from_pretrained()
 
 
+def error_file(example):
+    for side in sides:
+        try:
+            example[f"{side}.audio"]
+        except LibsndfileError:
+            return False
+    return True
+
+
+print(f"Num examples: {len(dataset)}")
+dataset = dataset.filter(error_file)
+print(f"Num examples (after filtering): {len(dataset)}")
+
+
 def tokenize(example):
     for side in sides:
         wav = torch.as_tensor(example[f"{side}.audio"]["array"].reshape(1, 1, -1), dtype=torch.float32)