pseudolabeled_labels = pseudolabeled_df.labels.tolist()
else:
pseudolabeled_audio_files = []
pseudolabeled_labels = []
train_loader = torch.utils.data.DataLoader(
AntispoofDataset(
audio_files=[
os.path.join(args.train_data_dir, fname)
for fname in train_df.fname.values[train]
] + pseudolabeled_audio_files,
labels=train_df.labels.values[train].tolist() +
pseudolabeled_labels,
transform=Compose([
LoadAudio(),
SampleLongAudio(max_length=args.max_audio_length),
MixUp(p=args.p_mixup),
AudioAugmentation(p=args.p_aug),
audio_transform,
DropFields(("audio", "sr")),
]),
clean_transform=Compose([
LoadAudio(),
SampleLongAudio(max_length=args.max_audio_length)
])),
shuffle=True,
drop_last=True,
batch_size=config.train.batch_size,
collate_fn=make_collate_fn(
{"signal": audio_transform.padding_value}),
**loader_kwargs)
"pin_memory": True
} if torch.cuda.is_available() else {})
valid_loader = torch.utils.data.DataLoader(
SoundDataset(
audio_files=[
os.path.join(args.train_data_dir, fname)
for fname in train_df.fname.values[valid]
],
labels=[
item.split(",") for item in train_df.labels.values[valid]
],
transform=Compose([
LoadAudio(),
MapLabels(class_map=class_map),
SampleLongAudio(args.max_audio_length),
ShuffleAudio(chunks_range=(12, 20), p=1.0),
audio_transform,
DropFields(("audio", "filename", "sr")),
]),
clean_transform=Compose([
LoadAudio(),
MapLabels(class_map=class_map),
]),
),
shuffle=False,
batch_size=args.batch_size,
collate_fn=make_collate_fn(
{"signal": audio_transform.padding_value}),
**loader_kwargs)