15 KiB
parakeet.data
This short guide shows the design of parakeet.data and how we use it in an experiment.
The most important concepts of parakeet.data are DatasetMixin, DataCargo, Sampler, batch function and DataIterator.
Dataset
Dataset, as we assume here, is a list of examples. You can get its length by len(dataset)(which means it length is known, and we have to implement __len__ method for it). And you can access its items randomly by dataset[i](which means we have to implement __getitem__ method for it). Furthermore, you can iterate over it by iter(dataset) or for example in dataset, which means we have to implement __iter__ method for it.
DatasetMixin
We provide an DatasetMixin object which provides the above methods. You can inherit DatasetMixin and implement get_example method for it to define your own dataset class. The get_example method is called by __getitem__ method automatically.
We also define several high-order Dataset classes, the obejcts of which can be built from some given Dataset objects.
TupleDataset
Dataset that is a combination of sevral datasets of the same length. An example of a Tupledataset is a tuple of examples of its constituent datasets.
DictDataset
Dataset that is a combination of sevral datasets of the same length. An example of the Dictdataset is a dict of examples of its constituent datasets.
SliceDataset
SliceDataset is a slice of the base dataset.
SubsetDataset
SubsetDataset is a subset of the base dataset.
ChainDataset
ChainDataset is the concatenation of several datastes with the same fields.
TransformDataset
A TransformeDataset is created by applying a transform to the base dataset. The transform is a callable object which takes an example of the base dataset as parameter and returns an example of the TransformDataset. The transformation is lazy, which means it is applied to an example only when requested.
FilterDataset
A FilterDataset is created by applying a filter to the base dataset. A filter is a predicate that takes an example of the base dataset as parameter and returns a boolean. Only those examples that pass the filter are included in the FilterDataset.
Note that the filter is applied to all the examples in the base dataset when initializing a FilterDataset.
CacheDataset
By default, we preprocess dataset lazily in DatasetMixin.get_example. An example is preprocessed whenever requested. But CacheDataset caches the base dataset lazily, so each example is processed only once when it is first requested. When preprocessing the dataset is slow, you can use Cachedataset to speed it up, but caching may consume a lot of RAM if the dataset is large.
Finally, if preprocessing the dataset is slow and the processed dataset is too large to cache, you can write your own code to save them into files or databases, and then define a Dataset to load them. Dataset is flexible, so you can create your own dataset painlessly.
DataCargo
DataCargo, like Dataset, is an iterable, but it is an iterable of batches. We need Datacargo because in deep learning, batching examples into batches exploits the computational resources of modern hardwares. You can iterate it by iter(datacargo) or for batch in datacargo. DataCargo is an iterable but not an iterator, in that in can be iterated more than once.
batch function
The concept of batch is something transformed from a list of examples. Assume that an example is a structure(tuple in python, or struct in C and C++) consists of several fields, then a list of examples is an array of structures(AOS, e.g. a dataset is an AOS). Then a batch here is a structure of arrays (SOA). Here is an example:
The table below represents 2 examples, each of which contains 5 fields.
| weight | height | width | depth | density |
|---|---|---|---|---|
| 1.2 | 1.1 | 1.3 | 1.4 | 0.8 |
| 1.6 | 1.4 | 1.2 | 0.6 | 1.4 |
The AOS representation and SOA representation of the table are shown below.
AOS:
[(1.2, 1,1, 1,3, 1,4, 0.8),
(1.6, 1.4, 1.2, 0.6, 1.4)]
SOA:
([1,2, 1.6],
[1.1, 1.4],
[1.3, 1.2],
[1.4, 0.6],
[0.8, 1.4])
For the example above, converting an AOS to an SOA is trivial, just stacking every field for all the examples. But it is not always the case. When a field contains a sequence, you may have to pad all the sequences to the largest length then stack them together. In some other cases, we may want to add a field for the batch, for example, valid_length for each example. So in general, a function to transform an AOS to SOA is needed to build a Datacargo from a dataset. We call this the batch function (batch_fn), but you can use any callable object if you need to.
Usually we need to define the batch function as an callable object which stores all the options and configurations as its members. Its __call__ method transforms a list of examples into a batch.
Sampler
Equipped with a batch function(we have known how to batch), here comes the next question. What to batch? We need to decide which examples to pick when creating a batch. Since a dataset is a list of examples, we only need to pick indices for the corresponding examples. A sampler object is what we use to do this.
A Sampler is represented as an iterable of integers. Assume the dataset has N examples, then an iterable of intergers in the range[0, N) is an appropriate sampler for this dataset to build a DataCargo.
We provide several samplers that is ready to use. The SequentialSampler, RandomSampler and so on.
DataIterator
DataIterator is what returned by iter(data_cargo). It can only be iterated over once.
Here's the analogy.
Dataset --> Iterable[Example] | iter(Dataset) -> Iterator[Example]
DataCargo --> Iterable[Batch] | iter(DataCargo) -> Iterator[Batch]
In order to construct an iterator of batches from an iterator of examples, we construct a DataCargo from a Dataset.
Code Example
Here's an example of how we use parakeet.data to process the LJSpeech dataset with a wavenet model.
First, we would like to define a class which represents the LJSpeech dataset and loads it as-is. We try not to apply any preprocessings here.
import csv
import numpy as np
import librosa
from pathlib import Path
import pandas as pd
from parakeet.data import DatasetMixin
from parakeet.data import batch_spec, batch_wav
class LJSpeechMetaData(DatasetMixin):
def __init__(self, root):
self.root = Path(root)
self._wav_dir = self.root.joinpath("wavs")
csv_path = self.root.joinpath("metadata.csv")
self._table = pd.read_csv(
csv_path,
sep="|",
header=None,
quoting=csv.QUOTE_NONE,
names=["fname", "raw_text", "normalized_text"])
def get_example(self, i):
fname, raw_text, normalized_text = self._table.iloc[i]
fname = str(self._wav_dir.joinpath(fname + ".wav"))
return fname, raw_text, normalized_text
def __len__(self):
return len(self._table)
We make this dataset simple in purpose. It requires only the path of the dataset, nothing more. It only loads the metadata.csv in the dataset when it is initialized, which includes file names of the audio files, and the transcriptions. We do not even load the audio files at get_example.
Then we define a Transform object to transform an example of LJSpeechMetaData into an example we want for the model.
class Transform(object):
def __init__(self, sample_rate, n_fft, win_length, hop_length, n_mels):
self.sample_rate = sample_rate
self.n_fft = n_fft
self.win_length = win_length
self.hop_length = hop_length
self.n_mels = n_mels
def __call__(self, example):
wav_path, _, _ = example
sr = self.sample_rate
n_fft = self.n_fft
win_length = self.win_length
hop_length = self.hop_length
n_mels = self.n_mels
wav, loaded_sr = librosa.load(wav_path, sr=None)
assert loaded_sr == sr, "sample rate does not match, resampling applied"
# Pad audio to the right size.
frames = int(np.ceil(float(wav.size) / hop_length))
fft_padding = (n_fft - hop_length) // 2 # sound
desired_length = frames * hop_length + fft_padding * 2
pad_amount = (desired_length - wav.size) // 2
if wav.size % 2 == 0:
wav = np.pad(wav, (pad_amount, pad_amount), mode='reflect')
else:
wav = np.pad(wav, (pad_amount, pad_amount + 1), mode='reflect')
# Normalize audio.
wav = wav / np.abs(wav).max() * 0.999
# Compute mel-spectrogram.
# Turn center to False to prevent internal padding.
spectrogram = librosa.core.stft(
wav,
hop_length=hop_length,
win_length=win_length,
n_fft=n_fft,
center=False)
spectrogram_magnitude = np.abs(spectrogram)
# Compute mel-spectrograms.
mel_filter_bank = librosa.filters.mel(sr=sr,
n_fft=n_fft,
n_mels=n_mels)
mel_spectrogram = np.dot(mel_filter_bank, spectrogram_magnitude)
mel_spectrogram = mel_spectrogram
# Rescale mel_spectrogram.
min_level, ref_level = 1e-5, 20 # hard code it
mel_spectrogram = 20 * np.log10(np.maximum(min_level, mel_spectrogram))
mel_spectrogram = mel_spectrogram - ref_level
mel_spectrogram = np.clip((mel_spectrogram + 100) / 100, 0, 1)
# Extract the center of audio that corresponds to mel spectrograms.
audio = wav[fft_padding:-fft_padding]
assert mel_spectrogram.shape[1] * hop_length == audio.size
# there is no clipping here
return audio, mel_spectrogram
Transform loads the audio files, and extracts mel_spectrogram from them. This transformation actually needs a lot of options to specify, namely, the sample rate of the audio files, the n_fft, win_length, hop_length of stft transformation, and n_mels for transforming spectrogram into mel_spectrogram. So we define it as a callable class. You can also use a closure, or a partial if you want to.
Then we defines a functor to batch examples into a batch. Because the two fields ( audio and mel_spectrogram) are both sequences, batching them is not trivial. Also, because the wavenet model trains in audio clips of a fixed length(0.5 seconds, for example), we have to truncate the audio when creating batches. We want to crop audio randomly when creating batches, instead of truncating them when preprocessing each example, because it allows for an audio to be truncated at different positions.
class DataCollector(object):
def __init__(self,
context_size,
sample_rate,
hop_length,
train_clip_seconds,
valid=False):
frames_per_second = sample_rate // hop_length
train_clip_frames = int(
np.ceil(train_clip_seconds * frames_per_second))
context_frames = context_size // hop_length
self.num_frames = train_clip_frames + context_frames
self.sample_rate = sample_rate
self.hop_length = hop_length
self.valid = valid
def random_crop(self, sample):
audio, mel_spectrogram = sample
audio_frames = int(audio.size) // self.hop_length
max_start_frame = audio_frames - self.num_frames
assert max_start_frame >= 0, "audio is too short to be cropped"
frame_start = np.random.randint(0, max_start_frame)
# frame_start = 0 # norandom
frame_end = frame_start + self.num_frames
audio_start = frame_start * self.hop_length
audio_end = frame_end * self.hop_length
audio = audio[audio_start:audio_end]
return audio, mel_spectrogram, audio_start
def __call__(self, samples):
# transform them first
if self.valid:
samples = [(audio, mel_spectrogram, 0)
for audio, mel_spectrogram in samples]
else:
samples = [self.random_crop(sample) for sample in samples]
# batch them
audios = [sample[0] for sample in samples]
audio_starts = [sample[2] for sample in samples]
mels = [sample[1] for sample in samples]
mels = batch_spec(mels)
if self.valid:
audios = batch_wav(audios, dtype=np.float32)
else:
audios = np.array(audios, dtype=np.float32)
audio_starts = np.array(audio_starts, dtype=np.int64)
return audios, mels, audio_starts
When these 3 components are defined, we can start building our dataset with them.
# building the ljspeech dataset
ljspeech_meta = LJSpeechMetaData(root)
transform = Transform(sample_rate, n_fft, win_length, hop_length, n_mels)
ljspeech = TransformDataset(ljspeech_meta, transform)
# split them into train and valid dataset
ljspeech_valid = SliceDataset(ljspeech, 0, valid_size)
ljspeech_train = SliceDataset(ljspeech, valid_size, len(ljspeech))
# building batch functions (they can be differnt for training and validation if you need it)
train_batch_fn = DataCollector(context_size, sample_rate, hop_length,
train_clip_seconds)
valid_batch_fn = DataCollector(
context_size, sample_rate, hop_length, train_clip_seconds, valid=True)
# building the data cargo
train_cargo = DataCargo(
ljspeech_train,
train_batch_fn,
batch_size,
sampler=RandomSampler(ljspeech_train))
valid_cargo = DataCargo(
ljspeech_valid,
valid_batch_fn,
batch_size=1, # only batch=1 for validation is enabled
sampler=SequentialSampler(ljspeech_valid))
Here comes the next question, how to bring batches into Paddle's computation. Do we need some adaptor to transform numpy.ndarray into Paddle's native Variable type? Yes.
First we can use var = dg.to_variable(array) to transform ndarray into Variable.
for batch in train_cargo:
audios, mels, audio_starts = batch
audios = dg.to_variable(audios)
mels = dg.to_variable(mels)
audio_starts = dg.to_variable(audio_starts)
# your training code here
In the code above, processing of the data and training of the model run in the same process. So the next batch starts to load after the training of the current batch has finished. There is actually better solutions for this. Data processing and model training can be run asynchronously. To accomplish this, we would use DataLoader from Paddle. This serves as an adapter to transform an Iterable of batches into another iterable of batches, which runs asynchronously and transform each ndarray into Variable.
# connects our data cargos with corresponding DataLoader
# now the data cargo is connected with paddle
with dg.guard(place):
train_loader = fluid.io.DataLoader.from_generator(
capacity=10,return_list=True).set_batch_generator(train_cargo, place)
valid_loader = fluid.io.DataLoader.from_generator(
capacity=10, return_list=True).set_batch_generator(valid_cargo, place)
# iterate over the dataloader
for batch in train_loader:
audios, mels, audio_starts = batch
# your trains cript here