File size: 3,960 Bytes
5403e87 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 | import os
import random
import torch
from torch.utils.data import Dataset, DataLoader
from torch.nn.utils.rnn import pad_sequence
import pandas as pd
import sentencepiece as spm
import torchaudio
from torchaudio.transforms import Resample
# -------------------------
# Tokenizer
# -------------------------
sp = spm.SentencePieceProcessor()
sp.Load("./ressources/tokenizer/128_v7.model")
# -------------------------
# Load CSVs
# -------------------------
train_data = pd.read_csv("./ressources/train.csv", low_memory=False)
validation_data = pd.read_csv("./ressources/dev.csv", low_memory=False)
test_data = pd.read_csv("./ressources/test.csv", low_memory=False)
X_train, y_train = train_data["path"], train_data["sentence"]
X_val, y_val = validation_data["path"], validation_data["sentence"]
X_test, y_test = test_data["path"], test_data["sentence"]
del train_data, validation_data, test_data
audio_location = os.environ.get("AUDIO_LOCATION")
# -------------------------
# Collate Function
# -------------------------
def collate_fn(batch):
batch = [b for b in batch if b is not None]
if len(batch) == 0:
return None
transcriptions, waveforms, audio_lengths = zip(*batch)
transcriptions = [torch.tensor(t, dtype=torch.long) for t in transcriptions]
waveforms = [torch.tensor(w, dtype=torch.float32) for w in waveforms]
transcription_lengths = torch.tensor(
[t.size(0) for t in transcriptions], dtype=torch.int32
)
audio_lengths = torch.tensor(audio_lengths, dtype=torch.int32)
padded_waveforms = pad_sequence(waveforms, batch_first=True, padding_value=0.0)
padded_transcriptions = pad_sequence(
transcriptions, batch_first=True, padding_value=0
)
return padded_waveforms, padded_transcriptions, audio_lengths, transcription_lengths
# -------------------------
# Dataset
# -------------------------
class AudioDataset(Dataset):
def __init__(self, X, y, audio_location=audio_location, train=False):
self.audio_dirs = X.reset_index(drop=True)
self.transcriptions = y.reset_index(drop=True)
self.train = train
self.audio_location = audio_location
self.target_sr = 16000
self.resampler = None
def __len__(self):
return len(self.transcriptions)
def __getitem__(self, idx):
paths = str(self.audio_dirs[idx]).split(",")
if self.train:
chosen = random.randint(0, len(paths) - 1)
else:
chosen = 0
audio_location = f"{self.audio_location}/{paths[chosen]}.mp3"
# ---- Text ----
transcription = sp.Encode(self.transcriptions[idx], out_type=int)
# ---- Audio ----
waveform, sr = torchaudio.load(audio_location)
# Convert to mono
if waveform.size(0) > 1:
waveform = waveform.mean(dim=0, keepdim=True)
if sr != self.target_sr:
if self.resampler is None or self.resampler.orig_freq != sr:
self.resampler = Resample(orig_freq=sr, new_freq=self.target_sr)
waveform = self.resampler(waveform)
waveform = waveform.squeeze(0) # [T]
return transcription, waveform, waveform.size(0)
# -------------------------
# Datasets
# -------------------------
train_data = AudioDataset(X_train, y_train, train=True)
validation_data = AudioDataset(X_val, y_val)
test_data = AudioDataset(X_test, y_test)
# -------------------------
# DataLoaders
# -------------------------
train_dataloader = DataLoader(
train_data,
shuffle=True,
drop_last=True,
batch_size=64,
num_workers=8,
collate_fn=collate_fn,
pin_memory=True,
persistent_workers=True,
)
validation_dataloader = DataLoader(
validation_data,
batch_size=64,
num_workers=4,
collate_fn=collate_fn,
persistent_workers=True,
)
test_dataloader = DataLoader(
test_data,
batch_size=4,
num_workers=4,
collate_fn=collate_fn,
)
|