def test_split_corpus():
""" Test corpus splitter """
n_exemplars = 10
corpus_dir = "tests/split-corpus"
orig_dir = pjoin(corpus_dir, "orig")
split_dir = pjoin(corpus_dir, "splits")
trn_dir = pjoin(split_dir, "train")
dev_dir = pjoin(split_dir, "dev")
setup_test_corpus(orig_dir, trn_dir, dev_dir, n_exemplars)
orig_corpus = corpus({"location": orig_dir})
split_corpus(
orig_dir,
split_dir=split_dir,
split_name="dev",
split_words=19,
min_split_segs=1,
leftover_data_split_name="train",
rand_seed=1337,
)
# Make sure we didn't destroy input data
final_corpus = corpus({"location": orig_dir})
assert orig_corpus.validate() == 1
assert final_corpus.validate() == 1
orig_hashes = [_.hash() for _ in orig_corpus.exemplars]
final_hashes = [_.hash() for _ in final_corpus.exemplars]
assert all(h in final_hashes for h in orig_hashes)
# Make sure correct number of words present in data split
dev_corpus = corpus({"location": dev_dir})
assert sum(e.count_words() for e in dev_corpus.exemplars) == 20
assert dev_corpus.validate()
def split_corpus(
in_dir,
split_dir,
split_name="split",
split_words=1000,
min_split_segs=10,
leftover_data_split_name="orig",
rand_seed=None,
):
"""
Splits an ASR corpus directory based on number of words outputting splits in split_dir.
At least 1000 words is recommended for dev or tests splits to make WER calculations significant ~0.1%
Invalid files, such as empty files, will not be included in data splits.
Set rand_seed for reproducible splits
"""
seed(rand_seed)
c = corpus({"location": in_dir})
LOGGER.debug("%d exemplars before validating them", len(c.exemplars))
valid_exemplars, total_words = c.count_exemplar_words()
c.exemplars = valid_exemplars
LOGGER.debug("%d exemplars after validating them", len(valid_exemplars))
if min_split_segs > c.calculate_number_of_segments():
LOGGER.error(
"Not enough valid segments in corpus, %d, to make a split with %d segments. Reduce min_split_segs or get more data",
c.calculate_number_of_segments(),
min_split_segs,
)
sys.exit(1)
perform_split(c, split_dir, split_name, split_words, min_split_segs,
leftover_data_split_name)
def main():
parser = argparse.ArgumentParser(
description=
'Copy and organize specified corpora into a target directory. Training, testing, and development sets will be created automatically if not already defined.'
)
parser.add_argument('--target-dir',
default='input-data',
required=False,
help="Path to target directory")
parser.add_argument(
'corpora',
nargs='+',
help="Name of one or more directories in directory this script is run")
args = parser.parse_args()
for data_dir in data_dirs:
os.makedirs(args.target_dir + "/" + data_dir, exist_ok=True)
corpora = {
"dev": corpus(),
"test": corpus(),
"train": corpus(),
"unsorted": corpus()
}
for corpus_dir in args.corpora:
for data_dir in data_dirs:
if os.path.exists(corpus_dir + "/" + data_dir):
corpora[data_dir] += corpus(
{"location": corpus_dir + "/" + data_dir})
corpora['unsorted'] += corpus({"location": corpus_dir})
all_ready = all(corpora[data_dir].validate() for data_dir in data_dirs)
# dump extra data into training data by default
corpora['train'] += corpora['unsorted']
if not all_ready:
print(
"Not all training corpora were prepared. Automatically shuffling into training, testing, development sets"
)
# first pass, populate train directory
if not corpora['train'].validate():
corpora['train'] += corpora['dev'] + corpora['test']
# pick a file from training set to be dev set
if not corpora['dev'].validate():
corpora['dev'] = corpora['train'][-1]
corpora['train'] = corpora['train'][:-1]
# pick 20% for testing
if not corpora['test'].validate():
split_index = int(corpora['train'].validate() * 4 // 5)
corpora['test'] = corpora['train'][split_index:]
corpora['train'] = corpora['train'][:split_index]
# prepare for training
for data_dir in data_dirs:
corpora[data_dir].prepare_for_training(args.target_dir + "/" +
data_dir)
def gather_all_corpora(corpora_dirs):
"""
Finds all existing corpora and gathers into a dictionary
"""
corpora = {
data_dir: get_corpus(corpus_dir + "/" + data_dir)
for corpus_dir in corpora_dirs for data_dir in data_dirs
}
corpora["unsorted"] = corpus()
for unsorted_corpus in list(map(get_corpus, corpora_dirs)):
corpora["unsorted"] += unsorted_corpus
return corpora
def get_corpus(loc):
""" returns corpus for input location """
return corpus({"location": loc})
