"""Training data pipeline shared by model.py and load_balanced_model.py. Simulates a low-resolution input for a random target sample rate: lowpass filter the 48kHz ground truth, downsample, then cubic-spline upsample it back to 48kHz so the generator sees fixed-size input/target pairs. """ import os import random import librosa import torch import torch.nn.functional as F import torchaudio import torchaudio.transforms as T from ssr_eval.lowpass import lowpass_filter from layers import cubic_upsampler FILTER_TYPES = ["butter", "cheby1", "ellip", "bessel"] INPUT_SAMPLE_RATES = [8000, 16000, 24000] LOWPASS_FILTER_ORDER = 8 def _get_flist(folder): wav_files = [] for root, dirs, files in os.walk(folder): for filename in files: if filename.endswith(('.wav', '.flac', '.mp4')): wav_files.append(os.path.join(root, filename)) return wav_files class VoiceDataset(torch.utils.data.Dataset): def __init__(self, root: str, segment_length: int, target_sample_rate: int) -> None: self._flist = _get_flist(root) self.segment_length = segment_length self.target_sample_rate = target_sample_rate def __getitem__(self, n: int): x, sr = torchaudio.load(self._flist[n]) if sr > self.target_sample_rate: resample_transform_x = T.Resample(orig_freq=sr, new_freq=self.target_sample_rate) x = resample_transform_x(x) if x.size()[0] == 2: x = x[random.randint(0, 1)].unsqueeze(dim=0) assert x.shape[0] == 1 x = torch.squeeze(x) assert x.dim() == 1 if x.shape[0] < self.segment_length: x = F.pad(x, [0, self.segment_length - x.shape[0]], "constant") pos = random.randint(0, x.shape[0] - self.segment_length) x = x[pos:pos + self.segment_length] freq = random.choice(INPUT_SAMPLE_RATES) cutoff = freq * 0.5 z = lowpass_filter(x, cutoff, self.target_sample_rate, LOWPASS_FILTER_ORDER, ftype=random.choice(FILTER_TYPES)) z = librosa.resample(z, orig_sr=self.target_sample_rate, target_sr=freq) v = torch.tensor(z.copy(), dtype=torch.float32) v_ = cubic_upsampler(v, freq, self.target_sample_rate) y = torch.tensor(v_, dtype=torch.float32) return [x, y] def __len__(self) -> int: return len(self._flist) def _collate(examples): x_stack = torch.stack([example[0] for example in examples]) y_stack = torch.stack([example[1] for example in examples]) return [x_stack, y_stack] def make_dataloader(root: str, segment_length: int, target_sample_rate: int, batch_size: int) -> torch.utils.data.DataLoader: ds = VoiceDataset(root, segment_length, target_sample_rate) return torch.utils.data.DataLoader( ds, batch_size=batch_size, shuffle=True, collate_fn=_collate)