Dev gzf deepspeed (#1750)

zhifu gao

2024-05-23 f47d43c020bc2724423b9d9ee2fa61499bf8cb99

 funasr/bin/compute_audio_cmvn.py

@@ -52,7 +52,7 @@
        frontend=frontend,
        tokenizer=None,
        is_training=False,
        **kwargs.get("dataset_conf")
        **kwargs.get("dataset_conf"),
    )

    # dataloader
@@ -68,11 +68,14 @@
        dataset_train, collate_fn=dataset_train.collator, **batch_sampler_train
    )

    iter_stop = int(kwargs.get("scale", 1.0) * len(dataloader_train))

    total_frames = 0
    for batch_idx, batch in enumerate(dataloader_train):
        if batch_idx >= iter_stop:
        iter_stop = int(kwargs.get("scale", -1.0) * len(dataloader_train))
        log_step = iter_stop // 100
        if batch_idx % log_step == 0:
            logging.info(f"prcessed: {batch_idx}/{iter_stop}")
        if batch_idx >= iter_stop and iter_stop > 0.0:
            logging.info(f"prcessed: {iter_stop}/{iter_stop}")
            break

        fbank = batch["speech"].numpy()[0, :, :]