Commit 96c7ab7b authored by Bharath Ramsundar's avatar Bharath Ramsundar Committed by GitHub
Browse files

Merge pull request #278 from rbharath/muv_models 

Support Models for MUV dataset
parents 0c092b03 bf9b19a2

deepchem/splits/task_splitter.py

+11 −5
Original line number Diff line number Diff line
@@ -58,6 +58,8 @@ class TaskSplitter(Splitter): 
                             frac_test=.1):

    """Performs a train/valid/test split of the tasks for dataset.



    If split is uneven, spillover goes to test.



    Parameters

    ----------

    dataset: dc.data.Dataset
@@ -69,12 +71,11 @@ class TaskSplitter(Splitter): 
    frac_test: float, optional

      Proportion of tasks to be put into test. Rounded to nearest int.

    """

    np.testing.assert_almost_equal(frac_train + frac_valid + frac_test, 1)

    n_tasks = len(dataset.get_task_names())

    n_train = int(np.round(frac_train * n_tasks))

    n_valid = int(np.round(frac_valid * n_tasks))

    n_test = int(np.round(frac_test * n_tasks))

    if n_train + n_valid + n_test != n_tasks:

      raise ValueError("Train/Valid/Test fractions don't split tasks evenly.")

    n_test = n_tasks - n_train - n_valid 



    X, y, w, ids = dataset.X, dataset.y, dataset.w, dataset.ids
@@ -88,6 +89,8 @@ class TaskSplitter(Splitter): 
  def k_fold_split(self, dataset, K):

    """Performs a K-fold split of the tasks for dataset.



    If split is uneven, spillover goes to last fold.



    Parameters

    ----------

    dataset: dc.data.Dataset
@@ -98,13 +101,16 @@ class TaskSplitter(Splitter): 
    n_tasks = len(dataset.get_task_names())

    n_per_fold = int(np.round(n_tasks/float(K)))

    if K * n_per_fold != n_tasks:

      raise ValueError("Cannot perform a valid %d-way split" % K)

      print("Assigning extra tasks to last fold due to uneven split")

    

    X, y, w, ids = dataset.X, dataset.y, dataset.w, dataset.ids



    fold_datasets = []

    for fold in range(K):

      if fold != K-1:

        fold_tasks = range(fold*n_per_fold, (fold+1)*n_per_fold)

      else:

        fold_tasks = range(fold*n_per_fold, n_tasks)

      fold_datasets.append(

          NumpyDataset(X, y[:, fold_tasks], w[:, fold_tasks], ids))

    return fold_datasets

deepchem/splits/tests/test_task_splitter.py

+42 −0
Original line number Diff line number Diff line
@@ -58,6 +58,48 @@ class TestTaskSplitters(unittest.TestCase): 
    for fold_dataset in fold_datasets:

      assert len(fold_dataset.get_task_names()) == 2



  def test_uneven_k_fold_split(self):

    """

    Test k-fold-split works when K does not divide n_tasks.

    """

    n_samples = 100

    n_features = 10

    n_tasks = 17

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

    p = .05 # proportion actives

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

    dataset = dc.data.NumpyDataset(X, y)

    K = 4

    task_splitter = dc.splits.TaskSplitter()

    fold_datasets = task_splitter.k_fold_split(dataset, K)



    for fold in range(K-1):

      fold_dataset = fold_datasets[fold]

      assert len(fold_dataset.get_task_names()) == 4

    assert len(fold_datasets[-1].get_task_names()) == 5





  def test_uneven_train_valid_test_split(self):

    """

    Test train/valid/test split works when proportions don't divide n_tasks.

    """

    n_samples = 100

    n_features = 10

    n_tasks = 11

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

    p = .05 # proportion actives

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

    dataset = dc.data.NumpyDataset(X, y)



    task_splitter = dc.splits.TaskSplitter()

    train, valid, test = task_splitter.train_valid_test_split(

        dataset, frac_train=.4, frac_valid=.3, frac_test=.3)



    assert len(train.get_task_names()) == 4

    assert len(valid.get_task_names()) == 3

    # Note that the extra task goes to test

    assert len(test.get_task_names()) == 4



  def test_merge_fold_datasets(self):

    """

    Test that (K-1) folds can be merged into train dataset.

examples/low_data/datasets.py

+60 −7
Original line number Diff line number Diff line
@@ -12,13 +12,9 @@ import tempfile 
import numpy as np

import deepchem as dc



#sys.path.append("..")

#from muv.muv_datasets import load_muv



def load_tox21_ecfp(num_train=7200):

  """Load Tox21 datasets. Does not do train/test split"""

  # Set some global variables up top

  verbosity = "high"

  current_dir = os.path.dirname(os.path.realpath(__file__))

  dataset_file = os.path.join(

      current_dir, "../../datasets/tox21.csv.gz")
@@ -31,7 +27,7 @@ def load_tox21_ecfp(num_train=7200): 


  loader = dc.load.DataLoader(

      tasks=tox21_tasks, smiles_field="smiles", featurizer=featurizer,

      verbosity=verbosity)

      verbosity="high")

  dataset = loader.featurize(

      dataset_file, shard_size=8192)
@@ -48,7 +44,6 @@ def load_tox21_ecfp(num_train=7200): 
def load_tox21_convmol(base_dir=None, num_train=7200):

  """Load Tox21 datasets. Does not do train/test split"""

  # Set some global variables up top

  verbosity = "high"

  current_dir = os.path.dirname(os.path.realpath(__file__))

  dataset_file = os.path.join(

      current_dir, "../../datasets/tox21.csv.gz")
@@ -62,7 +57,7 @@ def load_tox21_convmol(base_dir=None, num_train=7200): 


  loader = dc.load.DataLoader(

      tasks=tox21_tasks, smiles_field="smiles",

      featurizer=featurizer, verbosity=verbosity)

      featurizer=featurizer, verbosity="high")

  dataset = loader.featurize(

      dataset_file, shard_size=8192)
@@ -75,3 +70,61 @@ def load_tox21_convmol(base_dir=None, num_train=7200): 
    dataset = transformer.transform(dataset)



  return tox21_tasks, dataset, transformers



def load_muv_ecfp():

  """Load MUV datasets. Does not do train/test split"""

  # Load MUV dataset

  print("About to load MUV dataset.")

  current_dir = os.path.dirname(os.path.realpath(__file__))

  dataset_file = os.path.join(

      current_dir, "../../datasets/muv.csv.gz")

  # Featurize MUV dataset

  print("About to featurize MUV dataset.")

  featurizer = dc.feat.CircularFingerprint(size=1024)

  MUV_tasks = sorted(['MUV-692', 'MUV-689', 'MUV-846', 'MUV-859', 'MUV-644',

                      'MUV-548', 'MUV-852', 'MUV-600', 'MUV-810', 'MUV-712',

                      'MUV-737', 'MUV-858', 'MUV-713', 'MUV-733', 'MUV-652',

                      'MUV-466', 'MUV-832'])



  loader = dc.load.DataLoader(

      tasks=MUV_tasks, smiles_field="smiles",

      featurizer=featurizer, verbosity="high")

  dataset = loader.featurize(dataset_file)



  # Initialize transformers 

  transformers = [

      dc.trans.BalancingTransformer(transform_w=True, dataset=dataset)]

  print("About to transform data")

  for transformer in transformers:

      dataset = transformer.transform(dataset)



  return MUV_tasks, dataset, transformers



def load_muv_convmol():

  """Load MUV datasets. Does not do train/test split"""

  # Load MUV dataset

  print("About to load MUV dataset.")

  current_dir = os.path.dirname(os.path.realpath(__file__))

  dataset_file = os.path.join(

      current_dir, "../../datasets/muv.csv.gz")

  # Featurize MUV dataset

  print("About to featurize MUV dataset.")

  featurizer = dc.feat.ConvMolFeaturizer()

  MUV_tasks = sorted(['MUV-692', 'MUV-689', 'MUV-846', 'MUV-859', 'MUV-644',

                      'MUV-548', 'MUV-852', 'MUV-600', 'MUV-810', 'MUV-712',

                      'MUV-737', 'MUV-858', 'MUV-713', 'MUV-733', 'MUV-652',

                      'MUV-466', 'MUV-832'])



  loader = dc.load.DataLoader(

      tasks=MUV_tasks, smiles_field="smiles",

      featurizer=featurizer, verbosity="high")

  dataset = loader.featurize(dataset_file)



  # Initialize transformers 

  transformers = [

      dc.trans.BalancingTransformer(transform_w=True, dataset=dataset)]

  print("About to transform data")

  for transformer in transformers:

      dataset = transformer.transform(dataset)



  return MUV_tasks, dataset, transformers

examples/low_data/muv_attn_one_fold.py

0 → 100644
+82 −0
Original line number Diff line number Diff line 
"""

Train low-data attn models on MUV. Test last fold only.

"""

from __future__ import print_function

from __future__ import division

from __future__ import unicode_literals



import tempfile

import numpy as np

import deepchem as dc

import tensorflow as tf

from datasets import load_muv_convmol



# Number of folds for split 

K = 4 

# Depth of attention module

max_depth = 3

# number positive/negative ligands

n_pos = 1

n_neg = 5

# Set batch sizes for network

test_batch_size = 128

support_batch_size = n_pos + n_neg

nb_epochs = 1

n_train_trials = 2000

n_eval_trials = 20

learning_rate = 1e-4

log_every_n_samples = 50

# Number of features on conv-mols

n_feat = 71



muv_tasks, dataset, transformers = load_muv_convmol()



# Define metric

metric = dc.metrics.Metric(

    dc.metrics.roc_auc_score, verbosity="high", mode="classification")



task_splitter = dc.splits.TaskSplitter()

fold_datasets = task_splitter.k_fold_split(dataset, K)



train_folds = fold_datasets[:-1] 

train_dataset = dc.splits.merge_fold_datasets(train_folds)

test_dataset = fold_datasets[-1]



# Train support model on train

support_model = dc.nn.SequentialSupportGraph(n_feat)



# Add layers

support_model.add(dc.nn.GraphConv(64, activation='relu'))

support_model.add(dc.nn.GraphPool())

support_model.add(dc.nn.GraphConv(128, activation='relu'))

support_model.add(dc.nn.GraphPool())

support_model.add(dc.nn.GraphConv(64, activation='relu'))

support_model.add(dc.nn.GraphPool())

support_model.add(dc.nn.Dense(128, activation='tanh'))



support_model.add_test(dc.nn.GraphGather(test_batch_size, activation='tanh'))

support_model.add_support(dc.nn.GraphGather(support_batch_size, activation='tanh'))



# Apply an attention lstm layer

support_model.join(dc.nn.AttnLSTMEmbedding(

    test_batch_size, support_batch_size, max_depth))



with tf.Session() as sess:

  model = dc.models.SupportGraphClassifier(

    sess, support_model, test_batch_size=test_batch_size,

    support_batch_size=support_batch_size, learning_rate=learning_rate,

    verbosity="high")



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

  print("FIT")

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

  model.fit(train_dataset, nb_epochs=nb_epochs, 

            n_episodes_per_epoch=n_train_trials,

            n_pos=n_pos, n_neg=n_neg, log_every_n_samples=log_every_n_samples)

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

  print("EVAL")

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

  scores = model.evaluate(

      test_dataset, metric, n_pos, n_neg, n_trials=n_eval_trials)

  print("Scores on evaluation dataset")

  print(scores)

examples/low_data/muv_res_one_fold.py

0 → 100644
+82 −0
Original line number Diff line number Diff line 
"""

Train low-data res models on MUV. Test last fold only.

"""

from __future__ import print_function

from __future__ import division

from __future__ import unicode_literals



import tempfile

import numpy as np

import deepchem as dc

import tensorflow as tf

from datasets import load_muv_convmol



# Number of folds for split 

K = 4 

# Depth of attention module

max_depth = 3

# num positive/negative ligands

n_pos = 1

n_neg = 5

# Set batch sizes for network

test_batch_size = 128

support_batch_size = n_pos + n_neg

nb_epochs = 1

n_train_trials = 2000

n_eval_trials = 20

learning_rate = 1e-4

log_every_n_samples = 50

# Number of features on conv-mols

n_feat = 71



muv_tasks, dataset, transformers = load_muv_convmol()



# Define metric

metric = dc.metrics.Metric(

    dc.metrics.roc_auc_score, verbosity="high", mode="classification")



task_splitter = dc.splits.TaskSplitter()

fold_datasets = task_splitter.k_fold_split(dataset, K)



train_folds = fold_datasets[:-1] 

train_dataset = dc.splits.merge_fold_datasets(train_folds)

test_dataset = fold_datasets[-1]



# Train support model on train

support_model = dc.nn.SequentialSupportGraph(n_feat)



# Add layers

support_model.add(dc.nn.GraphConv(64, activation='relu'))

support_model.add(dc.nn.GraphPool())

support_model.add(dc.nn.GraphConv(128, activation='relu'))

support_model.add(dc.nn.GraphPool())

support_model.add(dc.nn.GraphConv(64, activation='relu'))

support_model.add(dc.nn.GraphPool())

support_model.add(dc.nn.Dense(128, activation='tanh'))



support_model.add_test(dc.nn.GraphGather(test_batch_size, activation='tanh'))

support_model.add_support(dc.nn.GraphGather(support_batch_size, activation='tanh'))



# Apply a residual lstm layer

support_model.join(dc.nn.ResiLSTMEmbedding(

    test_batch_size, support_batch_size, max_depth))



with tf.Session() as sess:

  model = dc.models.SupportGraphClassifier(

    sess, support_model, test_batch_size=test_batch_size,

    support_batch_size=support_batch_size, learning_rate=learning_rate,

    verbosity="high")



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

  print("FIT")

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

  model.fit(train_dataset, nb_epochs=nb_epochs,

            n_episodes_per_epoch=n_train_trials,

            n_pos=n_pos, n_neg=n_neg, log_every_n_samples=log_every_n_samples)

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

  print("EVAL")

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

  scores = model.evaluate(

      test_dataset, metric, n_pos, n_neg, n_trials=n_eval_trials)

  print("Scores on evaluation dataset")

  print(scores)

CRA Git | Maintained and supported by SUSTech CRA and CCSE