Dali: How to resize image after read batch of video using video reader?

[liuchanglab]

Hi, When we reader video using nvidia-dali, we get tensorlist with size of (N, F, H, W, C) . How to resize these images using resize op? We get such error: The layout "FHWC" does not match any of the allowed layouts for input 0. Valid layouts are:
HWC

[JanuszL]

Hi,
Resize operator doesn't support Video kind of input yet. We plan to enable this in the near feature.
In special case, when you extract a sequence of length 1 you can use Reshape or ElementExtract operators and treat the data as an image:

class VideoPipe(Pipeline):
    def __init__(self, batch_size, num_threads, device_id, data, shuffle):
        super(VideoPipe, self).__init__(batch_size, num_threads, device_id, seed=16)
        self.input = ops.VideoReader(device="gpu", filenames=data, sequence_length=1,
                                     shard_id=0, num_shards=1,
                                     random_shuffle=shuffle, initial_fill=initial_prefetch_size)
        self.extract = ops.ElementExtract(device="gpu", element_map=0)
        self.reshape = ops.Reshape(device="gpu", shape=[720, 1280, 3], layout="HWC") # assumes one frame, this is more efficient, no-copy Op
        self.resize = ops.Resize(device="gpu", resize_x=360, resize_y=720)


    def define_graph(self):
        frames = self.input(name="Reader")
        extracted = self.extract(frames)
        reshaped = self.reshape(frames) 
        out0 = self.resize(extracted)
        out1 = self.resize(reshaped)
        return out0, out1

or

class VideoPipe(Pipeline):
    def __init__(self, batch_size, num_threads, device_id, data, shuffle):
        super(VideoPipe, self).__init__(batch_size, num_threads, device_id, seed=16)
        self.input = ops.VideoReader(device="gpu", filenames=data, sequence_length=4,
                                     shard_id=0, num_shards=1,
                                     random_shuffle=shuffle, initial_fill=initial_prefetch_size)
        # Can be used to extract subset of elements from sequence, here taking 0th and 1st element from 4 element sequence
        self.extract = ops.ElementExtract(device="gpu", element_map=[0, 1]) 
        self.resize = ops.Resize(device="gpu", resize_x=360, resize_y=720)


    def define_graph(self):
        frames = self.input(name="Reader")
        extracted_0, extracted_1 = self.extract(frames)
        out0 = self.resize(extracted_0)
        out1 = self.resize(extracted_1)
        return out0, out1