每隔 100 epoch 保存最好的 1 个 checkpoint

examples/asr/asr_ctc/speech_to_text_ctc_bpe.py

@@ -98,10 +98,12 @@ if __name__ == '__main__':
     sys.argv.extend([
         '--config-path', '../conf/conformer/',
         '--config-name', 'conformer_ctc_bpe',
-        'exp_manager.name=Name of our experiment',
+        'exp_manager.name=NeMo_Ja',
         'exp_manager.resume_if_exists=true',
         'exp_manager.resume_ignore_no_checkpoint=true',
         'exp_manager.exp_dir=results/',
+        'exp_manager.checkpoint_callback_params.save_top_k=1',
+        'exp_manager.checkpoint_callback_params.every_n_epochs=100',
         '++model.encoder.conv_norm_type=layer_norm',
         'model.tokenizer.dir=data/common_voice_11_0/ja/tokenizers/tokenizer_spe_bpe_v2491',
         'model.train_ds.tarred_audio_filepaths=data/common_voice_11_0/ja/train_tarred_1bk/audio__OP_0..1023_CL_.tar',
@@ -111,7 +113,7 @@ if __name__ == '__main__':
         'model.train_ds.manifest_filepath=data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json',
         'model.validation_ds.manifest_filepath=data/common_voice_11_0/ja/validation/validation_common_voice_11_0_manifest.json',
         'model.test_ds.manifest_filepath=data/common_voice_11_0/ja/test/test_common_voice_11_0_manifest.json',
-        'trainer.max_epochs=1',
+        'trainer.max_epochs=1000',
         'trainer.devices=1',
         'trainer.accelerator=gpu',
         'trainer.strategy=auto',