def build_test_dataset(dataset_key,
returnn_python_exe,
returnn_root,
output_path,
bpe_size=2000):
ogg_zip_dict = get_ogg_zip_dict("corpora")
bliss_dict = get_bliss_corpus_dict()
test_ogg = ogg_zip_dict[dataset_key]
from i6_core.corpus.convert import CorpusToTextDictJob
test_reference_dict_file = CorpusToTextDictJob(
bliss_dict[dataset_key]).out_dictionary
train_bpe_datastream = get_bpe_datastream(bpe_size=bpe_size, is_recog=True)
audio_datastream = get_audio_datastream(returnn_python_exe, returnn_root,
output_path)
data_map = {
"audio_features": ("zip_dataset", "data"),
"bpe_labels": ("zip_dataset", "classes")
}
test_zip_dataset = returnn_standalone.data.datasets.OggZipDataset(
path=[test_ogg],
audio_opts=audio_datastream.as_returnn_audio_opts(),
target_opts=train_bpe_datastream.as_returnn_targets_opts(),
seq_ordering="sorted_reverse")
test_dataset = returnn_standalone.data.datasets.MetaDataset(
data_map=data_map,
datasets={"zip_dataset": test_zip_dataset},
seq_order_control_dataset="zip_dataset")
return test_dataset, test_reference_dict_file
def get_audio_datastream(returnn_python_exe, returnn_root, output_path):
ogg_zip_dict = get_ogg_zip_dict("corpora")
train_clean_100_ogg = ogg_zip_dict['train-clean-100']
audio_datastream = get_default_asr_audio_datastream(
statistics_ogg_zip=train_clean_100_ogg,
returnn_python_exe=returnn_python_exe,
returnn_root=returnn_root,
output_path=output_path,
)
return audio_datastream
def build_training_datasets(returnn_python_exe, returnn_root, output_path):
bpe_size = 2000
ogg_zip_dict = get_ogg_zip_dict("corpora")
train_clean_100_ogg = ogg_zip_dict['train-clean-100']
dev_clean_ogg = ogg_zip_dict['dev-clean']
dev_other_ogg = ogg_zip_dict['dev-other']
train_bpe_datastream = get_bpe_datastream(bpe_size=bpe_size,
is_recog=False)
audio_datastream = get_audio_datastream(
statistics_ogg_zip=train_clean_100_ogg,
returnn_python_exe=returnn_python_exe,
returnn_root=returnn_root,
output_path=output_path,
)
extern_data = {
'audio_features': audio_datastream.as_returnn_data_opts(),
'bpe_labels': train_bpe_datastream.as_returnn_data_opts()
}
data_map = {
"audio_features": ("zip_dataset", "data"),
"bpe_labels": ("zip_dataset", "classes")
}
train_zip_dataset = returnn_standalone.data.datasets.OggZipDataset(
path=train_clean_100_ogg,
audio_opts=audio_datastream.as_returnn_audio_opts(),
target_opts=train_bpe_datastream.as_returnn_targets_opts(),
partition_epoch=3,
seq_ordering="laplace:.1000",
other_opts={"epoch_wise_filter": {
(1, 5): {
"max_mean_len": 1000
}
}} # still hardcoded, future work
)
train_dataset = returnn_standalone.data.datasets.MetaDataset(
data_map=data_map,
datasets={"zip_dataset": train_zip_dataset},
seq_order_control_dataset="zip_dataset")
from i6_experiments.users.rossenbach.datasets.librispeech import get_mixed_cv_segments
cv_zip_dataset = returnn_standalone.data.datasets.OggZipDataset(
path=[dev_clean_ogg, dev_other_ogg],
audio_opts=audio_datastream.as_returnn_audio_opts(),
target_opts=train_bpe_datastream.as_returnn_targets_opts(),
segment_file=get_mixed_cv_segments(),
seq_ordering="sorted_reverse")
cv_dataset = returnn_standalone.data.datasets.MetaDataset(
data_map=data_map,
datasets={"zip_dataset": cv_zip_dataset},
seq_order_control_dataset="zip_dataset")
devtrain_zip_dataset = returnn_standalone.data.datasets.OggZipDataset(
path=train_clean_100_ogg,
audio_opts=audio_datastream.as_returnn_audio_opts(),
target_opts=train_bpe_datastream.as_returnn_targets_opts(),
segment_file=get_mixed_cv_segments(),
seq_ordering="sorted_reverse",
subset=3000,
)
devtrain_dataset = returnn_standalone.data.datasets.MetaDataset(
data_map=data_map,
datasets={"zip_dataset": devtrain_zip_dataset},
seq_order_control_dataset="zip_dataset")
return train_dataset, cv_dataset, devtrain_dataset, extern_data