tf_ucf101_dataset.py

# # First create and activate conda python3 environment:
# conda create -n video python=3.6
# conda activate video
# # Then install the requirements:
# conda install ffmpeg
# conda install tensorflow-gpu==1.13.1
# pip install tensorflow_datasets
# # The bellow code would still produce an error because of the missing file ("ucf101_labels.txt")
# # So manually download the "ucf101_labels.txt" and put it in place:
# cd "/home/$USER/anaconda3/envs/video/lib/python3.6/site-packages/tensorflow_datasets/video/"
# wget https://raw.githubusercontent.com/tensorflow/datasets/master/tensorflow_datasets/video/ucf101_labels.txt

import tensorflow as tf
import tensorflow_datasets as tfds
tf.enable_eager_execution()                                                                                                                              
builder = tfds.builder(name="ucf101")
builder.download_and_prepare()  # this needs to be called only once
dtrain = builder.as_dataset(split=tfds.Split.TRAIN)
dtrain = dtrain.repeat().padded_batch(1, padded_shapes={"video":[tf.Dimension(None), 256, 256, 3], "label":[]})
example, = dtrain.take(1)
print(example.keys())  # dict_keys(['video', 'label'])
print(example['video'].shape)  # TensorShape([Dimension(1), Dimension(74), Dimension(256), Dimension(256), Dimension(3)])