In constructing a neural network in Tensorflow, user can choose to initialise the weight value as a tf.varialble or tf.const. The latter means that the weight will not be updated during training. I find this property useful in some use cases, such as neural style transfer. I think it would be great to have such "freeze" mechanism in Owl's NN module.
In constructing a neural network in Tensorflow, user can choose to initialise the weight value as a
tf.varialble
ortf.const
. The latter means that the weight will not be updated during training. I find this property useful in some use cases, such as neural style transfer. I think it would be great to have such "freeze" mechanism in Owl's NN module.