Hi! I’m implementing a dataset that loads data from videos. I am following the example from X3D pytorch, which use EncodedVideo to load a video and get_clip to segment it to a clip.
class MyDataset(Dataset):
def __init__(self, data_root):
csv_list = get_csv_list(data_root)
action_list = create_action_list(csv_list)
self.data = self._prepare_data(action_list) # This function returns a list of [
# [path, label, start, end],
# [path, label, start, end], ...]
def _prepare_data(self, action_list):
data = []
for action in action_list:
filename = action['filename']
label = action['label']
start, end = time2sec([action['start'], action['end']])
data.append([filename, label, start, end])
return data
def __getitem__(self, idx):
filename, label, start, end = self.data[idx]
start_sec = 0
end_sec = clip_duration + start_sec
video_data = EncodedVideo.from_path(filename)
video_data = video_data.get_clip(start_sec=start_sec, end_sec=end_sec)
input = video_data["video"]
return label, input
def __len__(self):
return len(self.data)
clip_duration = 5
dataset = MyDataset(data_root)
for data in dataset:
label, input = data
The get_clip call is very expensive (1~1.5s when the clip_duration is 5). The memory usage is also high, and for some reason the RAM usage doesn’t drop when the code stops (I’m using Colab and the session crashes after a few iterations).
Is there a better way to implement a Dataset (and DataLoader) for datasets contaning videos?