CODEWITHSUNDEEP
tensorflow
Gilfoyle
Gilfoyle

Reputation: 3636

Tensorflow: Train two networks in one network

I face the following problem in Tensorflow: I constructed a network with the following layer sizes:

input_0 = 100
input_1 = 1000
output = 10

The entire network looks like:

[input_0, 300, 500, input_1, 800, 400, output]

Now I feed data to input_0 and run an optimizing step afterwards. Here I want to use the whole network which is straight forward. But after doing that I also want to be able to feed some data into input_1 and run an optimizing step which runs backwards but only until the input_1 layer. Is that even possible? I mean there should be a way to do that.

In short: How to train networks [input_0,...,output] and [input_1,...,output] independently of each other when they are both part of the same graph?

I tried to implement it in Tensorflow which resulted in many errors. I also tried to split the network into two networks. But than I don't know how to connect them properly.

Any suggestion?

Upvotes: 1

Views: 255

Answers (1)

Davis Yoshida
Davis Yoshida

Reputation: 1785

The optimizers accept a var_list parameter which will let you only update some weights.

See the documentation for GradientDescentOptimizer here.

Upvotes: 1

Related Questions