def _iter_subset_batch_data(self, subset_df, batch_size, repeat, shuffle):
# 전체 데이터를 순회할 때까지 반복
while True:
# 주어진 데이터에서 현재 iterator의 위치
i = 0
if shuffle:
subset_df = subset_df.sample(frac=1)
while True:
# 배치 데이터 slicing
batch_df = subset_df.iloc[i:i + batch_size]
# 데이터가 없으면 iteration 종료
if len(batch_df) == 0:
break
# 모든 배치 데이터에 대해 segment 데이터 읽어와서 리스트 생성
batch_data = [
cu.load(segment['path'])
for _, segment in batch_df.iterrows()
]
# x, y 데이터 분리
batch_x_video, batch_x_audio, batch_y = zip(
*[(segment['video'], segment['audio'], segment['label'])
for segment in batch_data])
batch_x_video = np.array(batch_x_video)
batch_x_audio = np.array(batch_x_audio)
batch_y = np.array(batch_y).reshape(-1, 1)
# 데이터를 iterator로 반환
batch_x = []
if 'video' in self.x_includes:
batch_x.append(batch_x_video)
if 'audio' in self.x_includes:
batch_x.append(batch_x_audio)
yield batch_x, batch_y
i += batch_size
if not repeat:
break
def convert_data(input_video_dir,
segment_length,
video_sample_rate,
video_width,
video_height,
audio_sample_rate,
apply_mfcc,
output_dataset_dir=None):
# output_dataset_dir이 주어지지 않은 경우 기본값으로 정의
if not output_dataset_dir:
output_dataset_dir = f'dataset_sl{segment_length}_vsr{video_sample_rate}_vw{video_width}_vh{video_height}_asr{audio_sample_rate}{"_mfcc" if apply_mfcc else ""}'
config = {
'segment_length': segment_length,
'video_sample_rate': video_sample_rate,
'video_width': video_width,
'video_height': video_height,
'audio_sample_rate': audio_sample_rate,
'apply_mfcc': apply_mfcc
}
print(f'input_video_dir: {input_video_dir}')
print(f'output_dataset_dir: {output_dataset_dir}')
print('config:')
pp(config)
print()
# output_dataset_dir이 이미 존재하는 경우 종료
if os.path.exists(output_dataset_dir):
print(
f'WARNING: output_dataset_dir \"{output_dataset_dir}\" already exists.'
)
if input('Remove all contents and continue? (y/n): ').lower().strip(
) == 'y':
import shutil
shutil.rmtree(output_dataset_dir)
else:
quit(1)
os.makedirs(output_dataset_dir, exist_ok=True)
# highlight 구간 파일 경로를 기준으로 영상 파일 경로와 제목을 생성
hl_section_path_list = glob.glob(os.path.join(input_video_dir, '*.txt'))
video_path_list = [
os.path.splitext(path)[0] + '.mp4' for path in hl_section_path_list
]
title_list = [
os.path.splitext(os.path.split(path)[1])[0]
for path in hl_section_path_list
]
# video_path_list 중 존재하지 않는 파일이 있으면 종료
video_path_exists = list(map(os.path.exists, video_path_list))
if not all(video_path_exists):
for video_path in [
video_path for exists, video_path in zip(
video_path_exists, video_path_list) if not exists
]:
print(f'ERROR: {video_path} not exists')
quit(1)
start = dt.now()
# 각 원본영상과 highlight 구간 파일을 읽고 segment 단위로 나누어 저장
with mp.Pool() as pool:
params = zip(repeat(config), title_list, video_path_list,
hl_section_path_list, repeat(output_dataset_dir))
output_path_list_list = pool.starmap(generate_segment_data, params)
total_segment_count = sum(map(len, output_path_list_list))
# get video/audio data shape
segment_data = cu.load(output_path_list_list[0][0])
video_data_shape = segment_data['video'].shape
audio_data_shape = segment_data['audio'].shape
# metadata 기록
metadata = {
'created': str(dt.now()),
'config': config,
'data_shape': {
'video': video_data_shape,
'audio': audio_data_shape
},
'total_segment_count': total_segment_count,
'segment_counts': {
title: count
for title, count in zip(title_list, map(len,
output_path_list_list))
}
}
metadata_path = os.path.join(output_dataset_dir, 'metadata.json')
with open(metadata_path, 'w') as f:
json.dump(metadata, f, indent=4)
end = dt.now()
print()
print(
f'Total {total_segment_count} segments saved, elapsed time: {end - start}s'
)
def _iter_subset_batch_data(self, subset_df, batch_size, repeat, shuffle):
# 전체 데이터를 순회할 때까지 반복
while True:
# 주어진 데이터에서 현재 iterator의 위치
i = 0
if shuffle:
subset_df = subset_df.sample(frac=1)
while True:
# 배치 데이터 slicing
batch_df = subset_df.iloc[i:i + batch_size]
# 데이터가 없으면 iteration 종료
if len(batch_df) == 0:
break
# 모든 배치 데이터에 대해 segment 데이터 읽어와서 리스트 생성
batch_data = []
for _, segment in batch_df.iterrows():
target_segment_data = cu.load(segment['path'])
# 시간값 추가
target_segment_data['time'] = target_segment_data[
'start_sec'] / target_segment_data['total_duration']
if self.x_expand > 0:
# 앞/뒤 segment 결합
segment_data_list = []
title = segment['title']
target_index = segment['index']
title_segment_list = self.all_segment_dict[title]
segment_data_zero = {
'video':
np.zeros_like(target_segment_data['video']),
'audio':
np.zeros_like(target_segment_data['audio'])
}
for index in range(target_index - self.x_expand,
target_index + self.x_expand + 1):
if index == target_index:
segment_data_list.append(target_segment_data)
elif 0 <= index < len(title_segment_list):
segment_data_list.append(
cu.load(title_segment_list[index]['path']))
else:
segment_data_list.append(segment_data_zero)
video_list, audio_list = zip(
*[(s['video'], s['audio'])
for s in segment_data_list])
# 원본 데이터 교체
target_segment_data['video'] = np.array(video_list)
target_segment_data['audio'] = np.array(audio_list)
batch_data.append(target_segment_data)
# x, y 데이터 분리
batch_x_video, batch_x_audio, batch_x_time, = zip(
*[(s['video'], s['audio'], s['time']) for s in batch_data])
batch_x_video = np.array(batch_x_video, dtype=np.float16)
batch_x_audio = np.array(batch_x_audio, dtype=np.float16)
batch_x_time = np.array(batch_x_time, dtype=np.float16)
# 데이터를 iterator로 반환
batch_x = []
if 'video' in self.x_includes:
batch_x.append(batch_x_video)
if 'audio' in self.x_includes:
batch_x.append(batch_x_audio)
if 'time' in self.x_includes:
batch_x.append(batch_x_time)
yield batch_x
i += batch_size
if not repeat:
break