How to freeze Auto Encoder Layers in tensorflow

Question

This is an autoencoder network, the first part is encoder and second part is decoder. I want to freeze the first three convolution layers and save the encoder parts. can you help me how can i do it? Thank you

def encoder(input_img):
    #encoder
    #input = 28 x 28 x 1 (wide and thin)
    conv1 = Conv2D(64, (2,2), activation='relu', padding='same')(input_img) #28 x 28 x 32
    conv2 = BatchNormalization()(conv1)
    conv3 = Conv2D(32, (2,2), activation='relu', padding='same')(conv2)
    conv4 = BatchNormalization()(conv3)
    pool5 = MaxPooling2D(pool_size=(2,2))(conv4) #14 x 14 x 32
    conv6 = Conv2D(16, (2,2), activation='relu', padding='same')(pool5) #14 x 14 x 64
    conv7 = BatchNormalization()(conv6)
    conv8 = Conv2D(8, (2,2), activation='relu', padding='same')(conv7)
    conv9 = BatchNormalization()(conv8)
    conv10 = Conv2D(4, (2,2), activation='relu', padding='same')(conv9)
    return conv10



def decoder(conv11):    
    #decoder
    conv12 = Conv2D(4, (2,2), activation='relu', padding='same')(conv11)
    conv13 = Conv2D(8, (2,2), activation='relu', padding='same')(conv12) #7 x 7 x 128
    conv14 = BatchNormalization()(conv13)
    conv15 = Conv2D(16, (2,2), activation='relu', padding='same')(conv14)
    conv16 = BatchNormalization()(conv15)
    conv17 = Conv2D(32, (2,2), activation='relu', padding='same')(conv16) #7 x 7 x 64
    conv18 = BatchNormalization()(conv17)
    conv19 = Conv2D(64, (2,2), activation='relu', padding='same')(conv18)
    conv20 = BatchNormalization()(conv19)
    up21 = UpSampling2D((2,2))(conv20) #14 x 14 x 64
    decoded = Conv2D(3, (2,2), activation='sigmoid', padding='same')(up21) # 28 x 28 x
    return decoded

autoencoder = Model(input_img, decoder(encoder(input_img)))
autoencoder.compile(loss='mae', optimizer = 'SGD')
autoencoder.summary()


train = np.concatenate((normal[0:1900,:,:,:],un_informative[0:1900,:,:,:]),axis=0) 
valid = np.concatenate((normal[1900:,:,:,:],un_informative[1900:,:,:,:]),axis=0)
history = autoencoder.fit(train ,train , batch_size=batch_size,epochs=200,verbose=1, validation_data=(valid, valid))

score 0 · Accepted Answer · answered Jun 17 '22 at 10:09

0

After you compile your model you can choose which layers to freeze by specifying:

layer_x.trainable = False

So I would suggest doing the following:

layers_to_freeze = [name of the layers]

for layer in model.layers:
    if layer in layers_to_freeze:
       layer.trainable=False

answered Jun 17 '22 at 10:09

DPM

845
7
33

Thank you. you mean i must compile the model at the first, then specify the non-trainable layers and at the last train the model? – aliashraf afrah Jun 17 '22 at 10:14
Yes, exactly. First compile the model, then freeze the layers and finally you are ready to train your model! – DPM Jun 17 '22 at 10:25
excuse me, can you help me how can i save the encoder part(with trainable and non-trainable layers) and then use it? – aliashraf afrah Jun 17 '22 at 10:50
Yeah, so there are two ways you can do this. You can save the whole model and then specify which portion to use or you can do what is suggested on the second answer of this post: https://stackoverflow.com/questions/58318195/how-to-bypass-portion-of-neural-network-in-tensorflow-for-some-but-not-all-fea. – DPM Jun 17 '22 at 12:07

How to freeze Auto Encoder Layers in tensorflow

1 Answers1