train_model for fingerprint data about conv_qsar_fast HOT 2 OPEN

Hi @chepyle ,

When using a fingerprint representation, the use_fp keyword in architecture specification should be making its way to main.core.build_model() and main.data.get_data_full(), which will use a fixed FP input instead of the learned FP. I see that the problem is coming from train_model(), because this feeds the three input arrays in (regardless of use_fp - this was a late change to the code after running the baseline models).

I've created a branch chepyle and modified main.core.train_model() to add two dummy inputs to the model. That way, when training or testing, these inputs will be accepted but ignored. Because these will be disconnected from the computational graph, you will have to pass on_unused_input='warn' to Theano when running the script.

An alternate solution would be to change all of the calls to .fit or .predict to be inside conditionals that look at use_fp and pass only the first element of the inputs if using a fixed FP representation.

Let me know if this solves the issue; if so, I will merge it into the master branch.

-Connor

from conv_qsar_fast.

Thanks Connor, the model does run now for with the fingerprint data sets. However, the disconnected graph causes problems with save_model similar to this:
keras-team/keras#2790

Jake

from conv_qsar_fast.