def train_dataloader(self, batch_size):
return DataLoader(AudioDataset(HPText.dataset,
start_idx=0,
end_idx=HPText.num_train,
durations=False),
batch_size=batch_size,
collate_fn=self.collate,
shuffle=True)
def val_dataloader(self, batch_size):
dataset = AudioDataset(HPText.dataset,
start_idx=HPText.num_train,
end_idx=HPText.num_valid,
durations=False)
return DataLoader(dataset,
batch_size=batch_size,
collate_fn=self.collate,
shuffle=False,
sampler=SequentialSampler(dataset))
def _parse_function(example_proto, y_label=None):
features = {
LABEL: tf.FixedLenFeature([], dtype=tf.string),
NAME: tf.FixedLenFeature([], dtype=tf.string),
AUDIO: tf.VarLenFeature(dtype=tf.float32), #max 396900
}
parsed_features = tf.parse_single_example(example_proto, features)
# Expand dims for channels
audio = tf.expand_dims(tf.concat(parsed_features[AUDIO].values, axis=0), axis=-1)
label = AudioDataset.str_label_to_int(parsed_features[y_label], LABEL_TO_INT_DICT)
if y_label:
return (audio, label,)
else:
return (audio,)
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", default=64, type=int, help="Batch size")
parser.add_argument("--num_samples",
default=4096,
type=int,
help="On how many samples to compute statistics")
parser.add_argument("--dataset",
default=HPText.dataset,
type=str,
help="Path to dataset")
args = parser.parse_args()
collate = Collate('cuda' if torch.cuda.is_available() else 'cpu',
standardize=False)
dl = DataLoader(AudioDataset(HPText.dataset,
alignments=True,
end_idx=args.num_samples),
collate_fn=collate,
batch_size=args.batch_size,
shuffle=False)
maxi = float('-inf')
mini = float('inf')
mean = 0
std = 0
w = 0
for i, b in enumerate(Bar(dl), 1):
s, slen, _, plen, _, _ = b
m = mask(s, slen, dim=1)
dx = x[1:] - x[:-1]
return torch.all(dx >= 0, dim=0)
if __name__ == '__main__':
import argparse
import sys
sys.path.append('code')
from torch.utils.data import DataLoader
from torch.utils.data.sampler import SequentialSampler
from datasets.AudioDataset import AudioDataset
from duration_extractor_custom import DurationExtractor, Collate
parser = argparse.ArgumentParser()
parser.add_argument("checkpoint", type=str, help="Path to checkpoint of convolutional_cacotron model")
parser.add_argument("data_folder", type=str, help="Where the data live and where to save durations.")
parser.add_argument("--durations_filename", default='durations.txt', type=str, help="Name of the final durations file.")
parser.add_argument("--batch_size", default=64, type=int, help="Batch size")
args = parser.parse_args()
# Load pretrained checkpoint and extract alignments to data_folder
m = DurationExtractor().load(args.checkpoint)
dataset = AudioDataset(root=args.data_folder, durations=False)
dataloader = DataLoader(dataset, batch_size=args.batch_size, collate_fn=Collate(m.device),
shuffle=False, sampler=SequentialSampler(dataset))
save_alignments_as_fertilities(m, dataloader, args.data_folder, args.durations_filename)
#python code/extract_durations.py checkpoints/aligner_ind_10500.pth code/datasets/data/TITML-IDN-LJ --batch_size 16 --durations_filename durations_aligner_ind_10500.txt