def main():
all_entries = csv_functions.read_csv_file(
'clotho_captions_development.csv', 'data')
all_words = []
for entry in all_entries:
entry_words = [
captions_functions.get_sentence_words(v, unique=True)
for k, v in entry.items() if not k.startswith('file')
]
all_words.extend(list(set(chain.from_iterable(entry_words))))
counter = Counter(all_words)
results = []
max_min = 0
files_to_use = []
max_files = 50
for entry in all_entries:
captions = [v for k, v in entry.items() if not k.startswith('file')]
min_freq = 1e6
for caption in captions:
min_freq = min(
min_freq, *[
counter.get(word) for word in
captions_functions.get_sentence_words(caption, unique=True)
])
max_min = max(max_min, min_freq)
results.append({'file': entry.get('file_name'), 'min_freq': min_freq})
if 10 < min_freq < 20:
files_to_use.append(entry.get('file_name'))
print(f'Max minimum freq is {max_min}')
print(f'Amount of files that I can use is {len(files_to_use)}')
# plt.hist([k['min_freq'] for k in results],
# bins=max_min,
# histtype='stepfilled')
# plt.grid()
# plt.show()
x = np.arange(len(files_to_use))
final_files = [
files_to_use[i] for i in np.random.permutation(x)[:max_files]
]
[print(f'File {i+1:02d} is {f}') for i, f in enumerate(final_files)]
p = Path('validation_file_names.pickle')
print('Saving list of validation files...', end=' ')
with p.open('wb') as f:
pickle.dump(final_files, f)
print('done.')
def get_hist_of_words(files: MutableSequence[Path],
data_dir: Path) \
-> Tuple[List[str], List[int]]:
all_captions = get_all_captions_from_files(files=files, data_dir=data_dir)
c = Counter(
chain.from_iterable([
captions_functions.get_sentence_words(v, unique=False)
for i in all_captions for v in i.values()
]))
return list(c.keys()), list(c.values())
def create_split_data(csv_split: MutableSequence[MutableMapping[str, str]],
dir_split: Path, dir_audio: Path, dir_root: Path,
words_list: MutableSequence[str],
chars_list: MutableSequence[str],
settings_ann: MutableMapping[str, Any],
settings_audio: MutableMapping[str, Any],
settings_output: MutableMapping[str, Any]) -> None:
"""Creates the data for the split.
:param csv_split: Annotations of the split.
:type csv_split: list[collections.OrderedDict]
:param dir_split: Directory for the split.
:type dir_split: pathlib.Path
:param dir_audio: Directory of the audio files for the split.
:type dir_audio: pathlib.Path
:param dir_root: Root directory of data.
:type dir_root: pathlib.Path
:param words_list: List of the words.
:type words_list: list[str]
:param chars_list: List of the characters.
:type chars_list: list[str]
:param settings_ann: Settings for the annotations.
:type settings_ann: dict
:param settings_audio: Settings for the audio.
:type settings_audio: dict
:param settings_output: Settings for the output files.
:type settings_output: dict
"""
# Make sure that the directory exists
dir_split.mkdir(parents=True, exist_ok=True)
captions_fields = [
settings_ann['captions_fields_prefix'].format(i)
for i in range(1,
int(settings_ann['nb_captions']) + 1)
]
# For each sound:
for csv_entry in csv_split:
file_name_audio = csv_entry[settings_ann['audio_file_column']]
audio = load_audio_file(audio_file=str(
dir_root.joinpath(dir_audio, file_name_audio)),
sr=int(settings_audio['sr']),
mono=settings_audio['to_mono'])
for caption_ind, caption_field in enumerate(captions_fields):
caption = csv_entry[caption_field]
words_caption = get_sentence_words(
caption,
unique=settings_ann['use_unique_words_per_caption'],
keep_case=settings_ann['keep_case'],
remove_punctuation=settings_ann['remove_punctuation_words'],
remove_specials=not settings_ann['use_special_tokens'])
chars_caption = list(
chain.from_iterable(
clean_sentence(caption,
keep_case=settings_ann['keep_case'],
remove_punctuation=settings_ann[
'remove_punctuation_chars'],
remove_specials=True)))
if settings_ann['use_special_tokens']:
chars_caption.insert(0, ' ')
chars_caption.insert(0, '<sos>')
chars_caption.append(' ')
chars_caption.append('<eos>')
indices_words = [words_list.index(word) for word in words_caption]
indices_chars = [chars_list.index(char) for char in chars_caption]
# create the numpy object with all elements
np_rec_array = np.rec.array(
np.array(
(file_name_audio, audio, caption, caption_ind,
np.array(indices_words), np.array(indices_chars)),
dtype=[('file_name', 'U{}'.format(len(file_name_audio))),
('audio_data', np.dtype(object)),
('caption', 'U{}'.format(len(caption))),
('caption_ind', 'i4'),
('words_ind', np.dtype(object)),
('chars_ind', np.dtype(object))]))
# save the numpy object to disk
dump_numpy_object(
np_obj=np_rec_array,
file_name=str(
dir_split.joinpath(
settings_output['file_name_template'].format(
audio_file_name=file_name_audio,
caption_index=caption_ind))))