Yakov Pechersky

I've just pushed a change which allows you to pass `test_split` as an optional command-line argument. Test-epoch size will be equal to `epoch_size / int((1 - test_split) / test_split)`. The...

#43 got merged 3 days ago, it seems. After 50 epochs of 600,000 strings (batch size 300) using the gen-method, I got 97% acc. After 150 epochs, I've plateau'd out...

@dribnet You might like to take a look at my PR #43, which hard codes a charset, and provides a helper object for decoding and encoding strings given that charset....

If I understand correctly, you might have an issue with how programs write to your HDD/SSD. What Python and TF/Theano are you using? On Fri, May 19, 2017 at 10:18...

You're right, the latter should be "data_test". In general, "train_gen.py" should be used instead, it should be less demanding on your machine. I wouldn't call an autoencoder or a VAE...

I suggest checking out the paper and repo I cite in #62. It also has pretrained models if you need that.

Yeah, go ahead and make a PR. On Thu, Mar 30, 2017 at 8:48 AM, Eli wrote: > @pechersky do you accept pull requests? > I've made some improvements to...

Could you also change the generator-based approach? Specifically, there is a `random.shuffle` call here that can be seeded: https://github.com/maxhodak/keras-molecules/blob/master/molecules/vectorizer.py. Additionally, perhaps seed should be a flag that can be passed...

I meant something like that SmilesDataGenerator.__init__ could take a seed kwarg, which would be passed in at train_gen.py, using some sort of flag. In that case, it's ok to pass...

I've committed a couple changes to train and train_gen to take the seed as a cli parameter. Could you test that they work as you expect?