Is there any other methods to use tensorflow and tensorlayer mutually?

[SunnerLi]

Recently I want to use tensorlayer and tensorflow mutually, but I encounter a trouble. For my design, I want to use batch normalization between dense layer and activation layer. Moreover, it's recommend to use this mechanism to reduce the variance. It's not problem that add batchNormLayer after DenseLayer, but I should to use native tensorflow function to reach the goal which returns tensor object. If I want to concatenate the other part of the network, the redundant InputLayer should be added.

For example:

import tensorlayer as tl
import tensorflow as tf

class Net(object):
    def __init__(self, _placeholder):
        # Assign input
        self.network = tl.layers.InputLayer(_placeholder, name='input1')

        # 1st fc + bn + elu
        self.network = tl.layers.DenseLayer(self.network, name ='fc1')
        self.network = tl.layers.BatchNormLayer(self.network, name='bn1')
        self.network = tf.nn.elu(self.network.outputs)

        # Add redundant layer with redundant name
        self.network = tl.layers.InputLayer(self.network, name='input2')

        # 2nd fc + bn + elu
        self.network = tl.layers.DenseLayer(self.network, name ='fc2')
        self.network = tl.layers.BatchNormLayer(self.network, name='bn2')
        self.network = tf.nn.elu(self.network.outputs)

ph = tf.placeholder(tf.float32, [None, 1000])
net = Net(ph)

The idea is from here. Is there any other method to use tensorlayer more directly? Does tensorlayer provide any other way to become more flexible, or adding such this Inputlayer is the only solution?

[zsdonghao]

Hi, you can put the activation function into BatchNormLayer, this is how I do :

• https://github.com/zsdonghao/SRGAN/blob/master/model.py

But I am not sure what do you mean by "concatenate the other part of the network", here is some tricks summary by others:

• https://github.com/wagamamaz/tensorlayer-tricks

Hope it can solve your problem.

[SunnerLi]

@zsdonghao I'm sorry that I miss the activation arg in batchNormLayer...... It's my fault..... Thanks for your model script! And the trick link is a really good demonstration (or tutorial).