Commit 637858c8 authored by Bharath Ramsundar's avatar Bharath Ramsundar
Browse files

Many changes

parent 41477b40

deepchem/datasets/__init__.py

+4 −1
Original line number Diff line number Diff line
@@ -397,7 +397,10 @@ def _df_to_numpy(df, feature_types, tasks): 
  # perform common train/test split across all tasks

  n_samples = df.shape[0]

  n_tasks = len(tasks)

  y = df[tasks].values

  #y = df[tasks].values

  print("df.keys()")

  print(df.keys())

  y = np.array([df[task].values for task in tasks])

  y = np.reshape(y, (n_samples, n_tasks))

  w = np.ones((n_samples, n_tasks))

  missing = np.zeros_like(y).astype(int)

deepchem/models/__init__.py

+2 −2
Original line number Diff line number Diff line
@@ -135,7 +135,7 @@ class Model(object): 


    return X, y, w



  def predict(self, dataset, transformers):

  def predict(self, dataset, transformers=[]):

    """

    Uses self to make predictions on provided Dataset object.
@@ -157,7 +157,7 @@ class Model(object): 
    y_pred = np.reshape(y_pred, (n_samples, n_tasks))

    return y_pred



  def predict_proba(self, dataset, transformers, n_classes=2):

  def predict_proba(self, dataset, transformers=[], n_classes=2):

    """

    TODO: Do transformers even make sense here?

deepchem/models/tensorflow_models/__init__.py

+1 −4
Original line number Diff line number Diff line
@@ -186,13 +186,11 @@ class TensorflowGraph(object): 


  def fit(self,

          dataset,

          summaries=False,

          max_checkpoints_to_keep=5):

    """Fit the model.



    Args:

      dataset: Dataset object that represents data on disk.

      summaries: If True, add summaries for model parameters.

      max_checkpoints_to_keep: Integer. Maximum number of checkpoints to keep;

        older checkpoints will be deleted.
@@ -225,7 +223,7 @@ class TensorflowGraph(object): 
          y_bs, y_preds = [], []

          ########## DEBUG

          for (X_b, y_b, w_b, ids_b) in dataset.iterbatches(batch_size):

            # Run training op and compute summaries.

            # Run training op.

            feed_dict = self.construct_feed_dict(X_b, y_b, w_b, ids_b)

            #fetches = self.output + [

            #    train_op.values()[0], self.loss, self.updates]
@@ -314,7 +312,6 @@ class TensorflowGraph(object): 
          raise ValueError(

              'Unrecognized rank combination for output: %s' %

              (batch_output.shape,))

        batch_weights = batch_weights.transpose((1, 0))

        output.append(batch_output)



        outputs = np.array(from_one_hot(

deepchem/models/tensorflow_models/fcnet.py

+3 −3
Original line number Diff line number Diff line
@@ -191,9 +191,9 @@ class TensorflowMultiTaskClassifier(TensorflowClassifier): 
      self.restore()

    ######### DEBUG

    with self.graph.as_default():

      ########### DEBUG

      #assert not model_ops.is_training()

      ########### DEBUG

      ########## DEBUG

      assert not model_ops.is_training()

      ########## DEBUG

      self.require_attributes(['output'])



      # run eval data through the model

deepchem/models/tests/test_overfit.py

+111 −2
Original line number Diff line number Diff line
@@ -422,6 +422,7 @@ class TestOverfitAPI(TestAPI): 
    """Test tensorflow models can overfit 0/1 datasets with few actives."""

    tasks = ["task0"]

    task_types = {task: "classification" for task in tasks}

    #n_samples = 100

    n_samples = 100

    n_features = 3

    n_tasks = len(tasks)
@@ -440,13 +441,13 @@ class TestOverfitAPI(TestAPI): 
    model_params = {

      "layer_sizes": [1500],

      "dropouts": [.0],

      "learning_rate": 0.001,

      "learning_rate": 0.003,

      "momentum": .9,

      "batch_size": n_samples,

      "num_classification_tasks": 1,

      "num_classes": n_classes,

      "num_features": n_features,

      "weight_init_stddevs": [.3],

      "weight_init_stddevs": [1.],

      "bias_init_consts": [1.],

      "nb_epoch": 200,

      "penalty": 0.0,
@@ -474,3 +475,111 @@ class TestOverfitAPI(TestAPI): 
    scores = evaluator.compute_model_performance([classification_metric])



    assert scores[classification_metric.name] > .8



  def test_tf_skewed_missing_classification_overfit(self):

    """TF, skewed data, few actives



    Test tensorflow models overfit 0/1 datasets with missing data and few

    actives. This is intended to be as close to singletask MUV datasets as

    possible.

    """

    

    tasks = ["task0"]

    task_types = {task: "classification" for task in tasks}

    #n_samples = 250

    #n_samples = 500

    #n_samples = 1000

    #n_samples = 2000

    #n_samples = 5000

    n_samples = 5120

    #n_features = 3

    n_features = 6

    n_tasks = len(tasks)

    n_classes = 2

    

    # Generate dummy dataset

    np.random.seed(123)

    #p = .002

    #p = .2

    #p = .1

    #p = .05

    #p = .01

    #p = .005

    p = .002

    ids = np.arange(n_samples)

    X = np.random.rand(n_samples, n_features)

    y = np.random.binomial(1, p, size=(n_samples, n_tasks))

    w = np.ones((n_samples, n_tasks))

    print("np.count_nonzero(y)")

    print(np.count_nonzero(y))

    #w = np.random.binomial(1, p, size=(n_samples, n_tasks))

    #print("np.amin(w), np.amax(w)")

    #print(np.amin(w), np.amax(w))



    #print("y_nonzero.shape")

    #print(y_nonzero.shape)

    #print("np.count_nonzero(y_nonzero)")

    #print(np.count_nonzero(y_nonzero))

    ##### DEBUG

    y_flat, w_flat = np.squeeze(y), np.squeeze(w)

    y_nonzero = y_flat[w_flat != 0]

    num_nonzero = np.count_nonzero(y_nonzero)

    weight_nonzero = len(y_nonzero)/num_nonzero

    print("weight_nonzero")

    print(weight_nonzero)

    w_flat[y_flat != 0] = weight_nonzero

    w = np.reshape(w_flat, (n_samples, n_tasks))

    print("np.amin(w), np.amax(w)")

    print(np.amin(w), np.amax(w))

    ##### DEBUG

  

    dataset = Dataset.from_numpy(self.train_dir, tasks, X, y, w, ids)



    model_params = {

      "layer_sizes": [1200],

      "dropouts": [.0],

      "learning_rate": 0.003,

      "momentum": .9,

      #"batch_size": n_samples,

      #"batch_size": n_samples/2,

      #"batch_size": n_samples/4,

      #"batch_size": n_samples/8,

      #"batch_size": n_samples/16,

      #"batch_size": n_samples/32,

      "batch_size": n_samples/64,

      # TODO(rbharath): Is there a bug in the padding code? Why does it fail to

      # learn for non-multiples?

      #"batch_size": 600,

      "num_classification_tasks": 1,

      "num_classes": n_classes,

      "num_features": n_features,

      "weight_init_stddevs": [1.],

      "bias_init_consts": [1.],

      "nb_epoch": 250,

      "penalty": 0.0,

      "optimizer": "adam",

      "data_shape": dataset.get_data_shape()

    }

    print("model_params['batch_size']")

    print(model_params['batch_size'])



    verbosity = "high"

    classification_metric = Metric(metrics.roc_auc_score, verbosity=verbosity)

    model = TensorflowModel(

        tasks, task_types, model_params, self.model_dir,

        tf_class=TensorflowMultiTaskClassifier,

        verbosity=verbosity)



    # Fit trained model

    model.fit(dataset)

    model.save()



    y_pred_model = model.predict(dataset, transformers=[])

    y_pred_proba_model = model.predict_proba(dataset, transformers=[])



    # Eval model on train

    transformers = []

    evaluator = Evaluator(model, dataset, transformers, verbosity=verbosity)

    scores = evaluator.compute_model_performance([classification_metric])



    assert scores[classification_metric.name] > .8

CRA Git | Maintained and supported by SUSTech CRA and CCSE