Seperate training of reconstruction-task and property-prediction

[maxbernhard]

Dear All,

In the paper it is mentioned that 250 000 drug-like molecules were used to train the autoencoder-system And for training the Gaussian-process 2000 Molecules were used.

However the provided command line tool only provides one input.

Therefore the question: Is it possible to train the property-prediction and reconstruction-task separately, or how was this separated training achieved in the paper?

How should the code be executed, assuming we have the following two data sets?: • a limited number of Molecules with SMILES and properties and • a large number of Molecules with SMILES and without properties, (maybe overlapping with the smaller dataset)

Kind regards!

[wkl000]

Hello, do you figure out this question? Many thanks.

[jnwei-zz]

Hi All,

It sounds like you might like to do semi-supervised learning.

That is not currently supported by this repo, but you might be able to adapt it for your purposes. I would suggest looking at this paper for some ideas: https://arxiv.org/abs/1805.00108

Best, Jennifer

On Fri, Jun 28, 2019 at 9:35 AM wkl000 notifications@github.com wrote:

Hello, do you figure out this question? Many thanks.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_aspuru-2Dguzik-2Dgroup_chemical-5Fvae_issues_5-3Femail-5Fsource-3Dnotifications-26email-5Ftoken-3DADT3XUHH4DLN5OT3IFWU2XLP4YHSLA5CNFSM4E3BGN72YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODY2C5HA-23issuecomment-2D506736284&d=DwMCaQ&c=WO-RGvefibhHBZq3fL85hQ&r=UPzYrSHLXjnX3tYn90C8Ljjzb-yfrb1UtMOxOFh-tKk&m=qQfuQlGjhPcCvD22MpEG4HdqWm4ydziJI9XdkfoIEt4&s=gA8sbh69Qb1m6rNGN5AFpTIYFzWRID3tHKVAFtVc2e0&e=, or mute the thread https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_notifications_unsubscribe-2Dauth_ADT3XUGPYIARAV4WW3AAAPDP4YHSLANCNFSM4E3BGN7Q&d=DwMCaQ&c=WO-RGvefibhHBZq3fL85hQ&r=UPzYrSHLXjnX3tYn90C8Ljjzb-yfrb1UtMOxOFh-tKk&m=qQfuQlGjhPcCvD22MpEG4HdqWm4ydziJI9XdkfoIEt4&s=miVhcpk8csZyND4dXeg_YPXWVUwaQRTEYI-cpRWUlfI&e= .

[AustinApple]

Hi all, I also encounter the same problem, so I modified the code by myself. Please take it for your reference. After finishing training the VAE itself by a large number of Molecules with SMILES and without properties, I create a model which is composed of only the trained encoder and the property predictor and subsequently train the model by a limited number of Molecules with SMILES and properties.

the whole parameter setting https://github.com/AustinApple/modified_chemvae/blob/master/exp_property_training.json

In order to train the property predictor separately, I made some modifications in thetrain_vae.py https://github.com/AustinApple/modified_chemvae/blob/master/train_prop.py

However, in this way we need to freeze the weighting of encoder to train our property predictor or we will destroy the system of the trained auto-encoder. We can expect that the performance of property predictor would be worse than one training with encoder together. If there is any question. please let me know.

[UnixJunkie]

One problem I foresee is that if you don't train jointly the VAE with the property predictor, then the latent-space will not be organized according to the property. So, you are doing something quite different from what is reported in their paper.