How to initialize the parameters in build_cnn18 if I want to train the model?

[kalyo-zjl]

Hi,

I found in your training code that you initialized the model with the trained checkpoint. I may wonder where is the initialization weights for build_cnn18, it is not a VGG-net neither.

Best.

[zaiweizhang]

I would like to know the answer as well. Thanks!

[walsvid]

Hi, everyone. In our experiments, the encoder network can be trained from scratch. And another option is to replace it with VGG16 and use the weight of ImageNet pretrained.

In our current code, the initialization part is in inits.py, all layers are assigned some initialization method when created, generally Xavier initialization. And then, before training all variables are initialized by tensorflow code:

sess.run(tf.global_variables_initializer())