CODEWITHSUNDEEP
pythontensorflowkeras
Feline
Feline

Reputation: 784

How to rescale output within tensorflow model?

I am trying to rescale the input in the model so that the output of my model is in the range [0,1]. But I don't want to do it with python code outside the model, instead, I want to do it inside the model. I have tried to use Keras Rescaling layer, but the results show values that are below 0, so I'm not sure what I'm doing wrong here..

class Generator(tf.Module):
    def __init__(self):
        super(Generator, self).__init__(name='Generator')
        self.rescale = layers.experimental.preprocessing.Rescaling(1.0  / 255.0)
        ....
        self.leakyrelu_3 = layers.LeakyReLU()
        self.conv2dT_3 = layers.Conv2DTranspose(3, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh')
        

    @tf.function
    def __call__(self):
        inputs = tf.random.normal([1, 100,])
        inputs = self.rescale(inputs)
        ....
        inputs = self.leakyrelu_3(inputs)
        output = self.conv2dT_3(inputs)
        return output
    
gen = Generator()
print(gen().numpy())

array([[[[-4.39650057e-06, -1.98987391e-06,  6.26171777e-06],
         [ 5.42304952e-06, -2.60572483e-06,  3.15579837e-06],
         [ 6.24479253e-06, -6.08476057e-06, -2.21408982e-07],
         ...,
         [ 1.29237583e-06,  1.32161574e-07,  7.94150219e-06],
         [ 1.25768611e-06, -6.39899235e-06,  4.49939853e-06],
         [-9.01997112e-07, -1.03935065e-06,  8.86256657e-07]],

        [[ 1.23434333e-07,  4.14229225e-06,  2.49262189e-06],
         [ 4.24094833e-06, -1.37856159e-05,  4.89129479e-06],
         [ 1.17599066e-05, -5.83954352e-06, -5.19962259e-06],
         ...,
         [-1.65802758e-05, -4.44178841e-06,  2.28360886e-06],
         [ 2.48740253e-06,  2.20288484e-06,  6.34336720e-06],
         [-6.80923858e-06, -7.12390374e-06, -1.00838906e-05]],

        ...,

Upvotes: 1

Views: 337

Answers (1)

Nicolas Gervais
Nicolas Gervais

Reputation: 36714

You can use, as last layer (or replace your leaky relu activation):

tf.keras.layers.ReLU(max_value=1)

See the docs.

You can also use a sigmoid activation.

Upvotes: 1

Related Questions