How to use a pre-prepared model as non-trained subnets in a tensor flow?

Question

I would like to train a network that contains a subnet network, which I will need to correct during training. The basic idea is to add and add some layers of a pre-prepared network (inceptionV3)

new_layers -> pre-trained and fixed sub-net (inceptionv3) -> new_layers

and start the training process for the task that I have without changing the previously prepared task. I also need to connect directly to some level of the pre-prepared network. For example, with inceptionV3 I like to use it from conv 299x299 to the last pool level or from conv 79x79 to the last pool level.

Answer 1

, "", , , . Optimizer , , , minimize, ..

opt.minimize(loss, <subset of variables you want to train>)

tf.gradients , , , tf.gradients.

, " " , , . tf.Conv2D 299x299, , , , 79x79, .