silviaamam / qml Goto Github PK

The "unpadding" assumes that the data set is composed of only one molecule.

When the repository is installed using pip, it doesn't install tensorflow-gpu when there are gpus present

When using feed_dict to pass data to the tensorflow session, a copy of the data is made. Upgrade to using TFRecords so that it doesn't break with large amounts of data

This will be a problem for the TFRecord decoding f one inputs zs as int.

The descriptor parameters come out as None when the get_params function is used.

Even if you are using an already trained model to predict the energies of some structures, it still requires you to set the properties.

When training ARMP from a loaded model, two different cost functions seem to appear.

When the model is reloaded the predict function seem to output all zeros

def _fit(self, x, y, classes, dy)

needs to be

def _fit(self, x, y, classes, dy, dgdr=None)

since the fit method will pass dgdr as well.

tf.Dataset.shuffle() is meant for shuffling data that is in memory. So for now, we could just shuffle the data between 2 batches. Eventually implement this:
tensorflow/tensorflow#14857

The other problem with this is that if the shuffling is added, also the predictions will be shuffled, which is wrong.

When a model is reloaded, the hyper parameters such as the learning rate, the regularisation... cannot be changed, even though it may look from the instantiation of the object that they have been changed

Add a way of killing gracefully the training.

https://stackoverflow.com/questions/18499497/how-to-process-sigterm-signal-gracefully

Ideally the model should be saved if it gets killed, so that it can be restarted.

This problem was solved in ARMP but still not working in ARMP_G

1. No need to download data set to run example
2. Add how long an example script takes to run (ideally should be short)
3. Explain how you should format your own data to work with the script