Decord reader does not work with tf.Data.Dataset.from_generator

[douglas125]

Hello;

I would like to use tf.Data to process some video frames but I can't get Decord to work with from_generator. Versions: Decord: 0.3.9 tf: 2.2.0

Here's some minimal code to reproduce the issue. Any help is appreciated.

from decord import VideoReader
from decord import cpu, gpu
import tensorflow as tf

# use whatever video here e.g. http://file.all-free-download.com/downloadfiles/footage/traffic_jam_new_work_street_583.zip
vr = VideoReader('traffic_jam_new_work_street_583.mp4', ctx=cpu(0))
def frame_gen_decord():
    for k in range(len(vr)):
        sample = vr.next().asnumpy()
        yield sample

g = frame_gen_decord()
x = next(g)
h, w, c = x.shape

dataset = tf.data.Dataset.from_generator(
    frame_gen_decord,
    output_types=(tf.int32),
    output_shapes=(tf.TensorShape([h, w, c]))
)

# this throws an error
for frame in dataset.take(1):
    x = frame

[zhreshold]

@douglas125 Have you tried setting the output format to be tensorflow?

For example

import decord
decord.bridge.set_bridge('tensorflow')

[douglas125]

I have not. I'll do that and post the results.