Add support for model saving/loading

    self.build_graph()

    self.grad = None

  def save(self):

    self.grad = None # recompute grad on restore

    super(ANIRegression, self).save()

  def build_grad(self):

    self.grad = tf.gradients(self.outputs, self.atom_feats)

  def compute_grad(self, dataset, batch_size=1):

    with self._get_tf("Graph").as_default():

      if not self.built:

        self.build()

      if not self.grad:

        self.grad = tf.gradients(self.outputs, self.atom_feats)

        self.build_grad()

      feed_dict = dict()

      X = dataset.X

      flags = np.sign(np.array(X[:batch_size, :, 0]))

import pyanitools as pya

import app

import dill

def convert_species_to_atomic_nums(s):

  PERIODIC_TABLE = {"H": 1, "C": 6, "N": 7, "O": 8}

  res = []

      dc.metrics.Metric(dc.metrics.pearson_r2_score, mode="regression")

  ]

  model_dir = "/tmp/ani.pkl"

  if os.path.exists(model_dir):

    print("Restoring existing model...")

    model = dc.models.ANIRegression.load_from_dir(model_dir=model_dir)

  else:

    print("Fitting new model...")

    model = dc.models.ANIRegression(

        1,

        max_atoms,

        batch_size=batch_size,

        learning_rate=0.001,

        use_queue=False,

      model_dir="/tmp/ani.pkl",

        model_dir=model_dir,

        mode="regression")

    # For production, set nb_epoch to 100+

    model.fit(train_dataset, nb_epoch=1, checkpoint_interval=100)

    print("Saving model...")

    model.save()

  print("Evaluating model")

  train_scores = model.evaluate(train_dataset, metric, transformers)

  valid_scores = model.evaluate(valid_dataset, metric, transformers)

  print("Gradient of a single test set element:")

  print(model.grad_one(coords, atomic_nums))

  # currently broken

  # model.save()

  app.webapp.model = model

  app.webapp.run(host='0.0.0.0', debug=False)