train

游雁

2023-05-16 3d9f094e9652d4b84894c6fd4eae39a4a753b0f0

 funasr/utils/prepare_data.py

@@ -27,7 +27,7 @@
        parts = line.strip().split()
        if len(parts) < 2:
            continue
        text_dict[parts[0]] = " ".join(parts[1:]).lower()
        text_dict[parts[0]] = " ".join(parts[1:])
    filter_count = 0
    with open(wav_file, "w") as f_wav, open(text_file, "w") as f_text:
        for sample_name, wav_path in wav_dict.items():
@@ -36,10 +36,8 @@
                f_text.write(sample_name + " " + text_dict[sample_name] + "\n")
            else:
                filter_count += 1
    logging.info(
        "{}/{} samples in {} are filtered because of the mismatch between wav.scp and text".format(len(wav_lines),
                                                                                                   filter_count,
                                                                                                   dataset))
    logging.info("{}/{} samples in {} are filtered because of the mismatch between wav.scp and text".
                 format(filter_count, len(wav_lines), dataset))


def wav2num_frame(wav_path, frontend_conf):
@@ -72,7 +70,7 @@
                f.flush()


def calc_shape(args, dataset, nj=32):
def calc_shape(args, dataset, nj=64):
    shape_path = os.path.join(args.data_dir, dataset, "speech_shape")
    if os.path.exists(shape_path):
        logging.info('Shape file for small dataset already exists.')
@@ -92,7 +90,7 @@
    start = 0
    for i in range(nj):
        end = start + num_job_lines
        file = os.path.join(shape_path, "wav.scp.{}".format(str(i + 1)))
        file = os.path.join(split_shape_path, "wav.scp.{}".format(str(i + 1)))
        with open(file, "w") as f:
            if i == nj - 1:
                f.writelines(lines[start:])
@@ -117,7 +115,7 @@
    logging.info('Generating shape files done.')


def generate_data_list(data_dir, dataset, nj=100):
def generate_data_list(data_dir, dataset, nj=64):
    list_file = os.path.join(data_dir, dataset, "data.list")
    if os.path.exists(list_file):
        logging.info('Data list for large dataset already exists.')
@@ -160,19 +158,52 @@
    distributed = distributed_option.distributed
    if not distributed or distributed_option.dist_rank == 0:
        filter_wav_text(args.data_dir, args.train_set)
        filter_wav_text(args.data_dir, args.dev_set)
        filter_wav_text(args.data_dir, args.valid_set)

        if args.dataset_type == "small" and args.train_shape_file is None:
        if args.dataset_type == "small":
            calc_shape(args, args.train_set)
            calc_shape(args, args.dev_set)
            calc_shape(args, args.valid_set)

        if args.dataset_type == "large" and args.train_data_file is None:
        if args.dataset_type == "large":
            generate_data_list(args.data_dir, args.train_set)
            generate_data_list(args.data_dir, args.dev_set)
            generate_data_list(args.data_dir, args.valid_set)

    args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "speech_shape")]
    args.valid_shape_file = [os.path.join(args.data_dir, args.dev_set, "speech_shape")]
    args.train_data_file = os.path.join(args.data_dir, args.train_set, "data.list")
    args.valid_data_file = os.path.join(args.data_dir, args.dev_set, "data.list")
    if args.dataset_type == "small":
        args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "speech_shape")]
        args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "speech_shape")]
        data_names = args.dataset_conf.get("data_names", "speech,text").split(",")
        data_types = args.dataset_conf.get("data_types", "sound,text").split(",")
        args.train_data_path_and_name_and_type = [
            ["{}/{}/wav.scp".format(args.data_dir, args.train_set), data_names[0], data_types[0]],
            ["{}/{}/text".format(args.data_dir, args.train_set), data_names[1], data_types[1]]
        ]
        args.valid_data_path_and_name_and_type = [
            ["{}/{}/wav.scp".format(args.data_dir, args.valid_set), data_names[0], data_types[0]],
            ["{}/{}/text".format(args.data_dir, args.valid_set), data_names[1], data_types[1]]
        ]
        if args.embed_path is not None:
            args.train_data_path_and_name_and_type.append(
                [os.path.join(args.embed_path, "embeds", args.train_set, "embeds.scp"), "embed", "kaldi_ark"])
            args.valid_data_path_and_name_and_type.append(
                [os.path.join(args.embed_path, "embeds", args.valid_set, "embeds.scp"), "embed", "kaldi_ark"])
    else:
        args.train_data_file = os.path.join(args.data_dir, args.train_set, "data.list")
        args.valid_data_file = os.path.join(args.data_dir, args.valid_set, "data.list")
        if args.embed_path is not None:
            if not distributed or distributed_option.dist_rank == 0:
                for d in [args.train_set, args.valid_set]:
                    file = os.path.join(args.data_dir, d, "data.list")
                    with open(file) as f:
                        lines = f.readlines()
                    out_file = os.path.join(args.data_dir, d, "data_with_embed.list")
                    with open(out_file, "w") as out_f:
                        for line in lines:
                            parts = line.strip().split()
                            idx = parts[0].split("/")[-2]
                            embed_file = os.path.join(args.embed_path, "embeds", args.valid_set, "ark",
                                                      "embeds.{}.ark".format(idx))
                            out_f.write(parts[0] + " " + parts[1] + " " + embed_file + "\n")
            args.train_data_file = os.path.join(args.data_dir, args.train_set, "data_with_embed.list")
            args.valid_data_file = os.path.join(args.data_dir, args.valid_set, "data_with_embed.list")
    if distributed:
        dist.barrier()