update

egs/librilight/SSL/local/extract_kmeans.py

@@ -18,6 +18,7 @@
 import argparse
 import logging
 import math
+import os
 from pathlib import Path
 from typing import Optional
 
@@ -36,6 +37,8 @@ from tqdm import tqdm
 torch.set_num_threads(1)
 torch.set_num_interop_threads(1)
 
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:512"
+
 
 class ApplyKmeans(object):
     def __init__(self, km_path):

egs/librilight/SSL/prepare.sh

@@ -64,7 +64,7 @@ fi
 
 if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   log "Stage 3: Split medium and large subset into pieces"
-  num_per_split=2500
+  num_per_split=10000
   split_dir=data/kmeans/medium_split
   if [ ! -f $split_dir/.split_completed ]; then
     lhotse split-lazy ./data/kmeans/librilight_cuts_medium_raw.jsonl.gz $split_dir $num_per_split