1

I have a dataframe with ~15k paths to audio files with on which I want to perform an operation (artificially add noise). Generally the whole thing works, but takes a long time even with less (16) records. The problem is not the execution time of the function but the time until all is initialized.

start = time.time()
data_augmented = data_augmented.swifter.progress_bar(True, desc="Merge Sounds") \
        .apply(merge_sounds(**settings), axis=1)
print(f"{time.time-(start)} - Map Timer")

Merge Sounds: 100%|█████████████████████████████| 16/16 [00:07<00:00,  2.09it/s]
26.973325729370117 - Map Timer

As you can see here, the initialization takes almost 4 times as long as the runtime of the Lambda Function (merge_sounds). With initialization-time im referring to elapsed_time_measured_by_myself - elapsed_time_measured_by_tqdm so in this case 26.97.. - 7 = 19.97

    start = time.time()
    lambda_fn = merge_sounds(**settings) #doesnt work if i put it in the line below.
    data_augmented = process_map(lambda_fn, data_augmented, max_workers=threads,
                                     desc=f"Merge_sounds [{threads} Threads]")
    print(f"{time.time-(start)} - Map Timer")

is stuck at:

Merge_sounds [16 Threads]:   0%|                         | 0/16 [00:00<?, ?it/s]

    with Pool(processes=16) as pool:
        data_augmented = pool.map(merge_sounds(**settings), tqdm(data_augmented, desc=f"Merge Sounds: {16} Threads"))

is stuck at:

Merge Sounds: 16 Threads:  38%|██████          | 6/16 [00:00<00:00, 4697.75it/s]

I know that parallelization doesn't make sense for smaller datasets, I'm just puzzled why I can easily parallelize everything everywhere in the code and I just can't get ahead here. I later run this code on alot of data so if parallelism would work, i would be very happy.

The Function used within Map is:

def merge_sounds(**settings):
    _range = settings.get("snr_range", (0.15, 0.65))
    assert len(_range), "snr_range -> e.g. (0.15, 0.75)"
    target_sample_rate = settings.get("target_sample_rate", "16000")

    if "target_path" not in settings.keys():
        raise Exception("please Specify target_path in Settings-Dict")
    target_path = Path(settings["target_path"])
    target_path.mkdir(parents=True, exist_ok=True)

    def __call__(item):
        _target_path = item["path_augmented"]

        snr = round(uniform(_range[0], _range[1]), 4)
        pad_idx = item.name

        yp, _ = IO.load(item["path"], sample_rate=target_sample_rate)
        yn, _ = IO.load(item["path_noise"], sample_rate=target_sample_rate)
        item["snr"] = snr

        y_augmented = Effect.add_noise(yp, yn, snr=snr, pad_idx=pad_idx)
        IO.save_wav(y_augmented, _target_path, target_sample_rate)
        return item

    return __call__

Is there anything im Forghetting to Parallelise the Map Function (seems to work with one of this variations everywhere else in my code just as excpected)

Ty in advanced.

Ichi Ban
  • 33
  • 5

1 Answers1

0

If fixed this Issue by Bypassing TQDM. I zipped the Columns i needed something like this

    _paths_in = _df["path_input"]
    _paths_out = _df["path_output"]
    _path_noise = _df["path_noise"]
    job = zip(_paths_in, _path_noise, _paths_out, _filter_jobs)

and then i just passed this to MultiProcessor Function.

    jobs = list(enumerate((zip_jobs(df))))

    with Pool(processes=_threads) as pool:
        data_augmented = pool.map(execute_job, tqdm(jobs, desc=f"Audio-Augmentation: {_threads} Threads"))

merge_sounds and the new execute_jobare similar, it just changed, what parameter the function needs.

Ichi Ban
  • 33
  • 5