Load Videos from External Source

Describe the question.

I have a json file describing the whole video retrieval dataset. Part of the json is listed below:

[
    {
        "video": "/data/14111B1211bJYB4F5841E4-43JBb12C31842F2aFB58a5E38aB3F-a1EB7JCY1E5C7.mp4",
        "caption": "A woman is reading a book."
    },
    {
        "video": "/data/14111B1211bJYBBEJ8A-4CYb23Y---5C2b715-a-B7Fa5J7Ea-CEBaCF-F-8YYBbbF.mp4",
        "caption": "Students are discussing with each other."
    },
]

I don't know how to load the dataset using the json file in DALI pipeline.

I implement an external source input iterator:

class VideoTextInputIterator(object):
    def __init__(self, metadata_path, batch_size, shuffle=False):
        self.batch_size = batch_size
        self.shuffle = shuffle

        with open(metadata_path) as f:
            self.data_list = json.load(f)
        if shuffle:
            random.shuffle(self.data_list)
        self.length = len(self.data_list)

    def __iter__(self):
        self.i = 0
        return self

    def __len__(self):
        return len(self.data_list)

    def __next__(self):
        # stop iteration
        if self.i >= self.length:
            self.__iter__()
            raise StopIteration

        batch_videos = []
        batch_captions = []

        # avoid iterator overflow
        process_size = min(self.batch_size, self.length - self.i)
        for _ in range(process_size):
            data = self.data_list[self.i]
            video = data["video"]  # video path
            caption = data["caption"]  # video caption
            batch_videos.append(video)
            batch_captions.append(caption)
            self.i += 1

        return batch_videos, batch_captions

    next = __next__
    len = __len__

Then I use it in a DALI pipeline:

@pipeline_def
def create_video_text_pipeline(metadata_path, shuffle):
    pipe = Pipeline.current()
    videos, labels = fn.external_source(
        source=VideoTextInputIterator(
            metadata_path=metadata_path,
            batch_size=pipe.batch_size,
            shuffle=shuffle,
        ),
        num_outputs=2,
        dtype=[types.STRING, types.STRING]
    )
    videos, labels = fn.readers.video(
        sequence_length=4,
        device="gpu",
        filenames=videos
        normalized=False,
        random_shuffle=True,
        image_type=types.RGB,
        name="Reader",
    )
    return videos, labels

I run the pipe and get an error:

pipe = create_video_text_pipeline(
    metadata_path="/path/to/json/meta.json",
    batch_size=8,
    device_id=0,
    num_threads=1,
)
pipe.build()
out = pipe.run()

TypeError: The argument `filenames` for operator `Video` should not be a `DataNode` but a str or list of str

I read the documentation over and over again and can't find the best solution.
Maybe I need some help. Thank you guys !! :)

Check for duplicates

• I have searched the open bugs/issues and have found no duplicates for this bug report