Final2

import skvideo.io
import skvideo.datasets
import tensorflow as tf
from tensorflow.keras.applications.resnet50 import ResNet50
from tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions

# enabling eager execution for easier explanation
tf.enable_eager_execution()

model = ResNet50(weights='imagenet')

reader = skvideo.io.FFmpegReader(skvideo.datasets.bigbuckbunny(),
                             inputdict={},
                             outputdict={})


def gen_frames():
    for frame in reader.nextFrame():
        yield frame


dataset = tf.data.Dataset.from_generator(gen_frames, tf.int64)

def preprocess(frame):
    x = tf.image.resize_bilinear(frame, [224, 224])
    x = preprocess_input(x)
    return x

dataset = dataset.batch(64).map(preprocess, 10).prefetch(1)

def predict():
    with tf.device("/gpu:0"):
        for frames in dataset:
            yield model.predict(frames.numpy())

dataset2 = tf.data.Dataset.from_generator(predict, tf.float64)

def postprocess(output):
    # do some post processing
    return tf.argsort(output)[:3]

dataset2 = dataset2.map(postprocess, 10)

# unbatch the output if needed
# dataset2 = dataset2.apply(tf.data.experimental.unbatch())

for value in dataset2:
    print(decode_predictions(value.numpy()))