連結

from keras.applications.vgg16 import VGG16
from keras.models import Sequential, Model
from keras.layers import Input, TimeDistributed, Flatten, GRU, Dense, Dropout
from keras import optimizers

def build_model():
    pretrained_cnn = VGG16(weights='imagenet', include_top=False)
    # pretrained_cnn.trainable = False
    for layer in pretrained_cnn.layers[:-5]:
        layer.trainable = False
    # input shape required by pretrained_cnn
    input = Input(shape = (224, 224, 3)) 
    x = pretrained_cnn(input)
    x = Flatten()(x)
    x = Dense(2048)(x)
    x = Dropout(0.5)(x)
    pretrained_cnn = Model(inputs = input, output = x)

    input_shape = (None, 224, 224, 3) # (seq_len, width, height, channel)
    model = Sequential()
    model.add(TimeDistributed(pretrained_cnn, input_shape=input_shape))
    model.add(GRU(1024, kernel_initializer='orthogonal', bias_initializer='ones', dropout=0.5, recurrent_dropout=0.5))
    model.add(Dense(categories, activation = 'softmax'))

    model.compile(loss='categorical_crossentropy',
                optimizer = optimizers.SGD(lr=0.01, momentum=0.9, clipnorm=1., clipvalue=0.5),
                metrics=['accuracy'])
    return model
 

keras 官方給出的例子

from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.applications import InceptionV3

video=keras.Input(shape=(None,150,150,3))

cnn=InceptionV3(weights='imagenet',include_top=False,pooling='avg')

cnn.trainable=False

frame_features=layers.TimeDistributed(cnn)(video)

video_vector=layers.LSTM(256)(frame_features)