Commit e04c36a8 authored by Bharath Ramsundar's avatar Bharath Ramsundar Committed by GitHub
Browse files

Merge pull request #274 from rbharath/tox21_fixes 

Some minor fixes to Tox21 models
parents 1b94fc18 1a0a14ac

deepchem/models/tf_keras_models/multitask_classifier.py

+2 −18
Original line number Diff line number Diff line
@@ -177,37 +177,21 @@ class MultitaskGraphClassifier(Model): 
        softmax.append(tf.nn.softmax(logits, name='softmax_%d' % i))

    return softmax



  def fit(self, dataset, nb_epoch=10, pad_batches=False,

  def fit(self, dataset, nb_epoch=10, 

          max_checkpoints_to_keep=5, log_every_N_batches=50, **kwargs):

    # Perform the optimization

    log("Training for %d epochs" % nb_epoch, self.verbosity)

  

    # TODO(rbharath): Disabling saving for now to try to debug.

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

    # Save an initial checkpoint.

    #saver = tf.train.Saver(max_to_keep=max_checkpoints_to_keep)

    #saver.save(self.sess, self._save_path, global_step=0)

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

    for epoch in range(nb_epoch):

      # TODO(rbharath): This decay shouldn't be hard-coded.

      lr = self.learning_rate / (1 + float(epoch) / self.T)



      log("Starting epoch %d" % epoch, self.verbosity)

      # ToDo(hraut->rbharath) : what is the ids_b for? Is it the zero's? 

      for batch_num, (X_b, y_b, w_b, ids_b) in enumerate(dataset.iterbatches(

          self.batch_size, pad_batches=pad_batches)):

          self.batch_size, pad_batches=True)):

        if batch_num % log_every_N_batches == 0:

          log("On batch %d" % batch_num, self.verbosity)

        self.sess.run(

            self.train_op,

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

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

      #saver.save(self.sess, self._save_path, global_step=epoch)

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

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

    # Always save a final checkpoint when complete.

    #saver.save(self.sess, self._save_path, global_step=epoch+1)

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



  def save(self):

    """

deepchem/models/tf_keras_models/support_classifier.py

+24 −5
Original line number Diff line number Diff line
@@ -24,8 +24,8 @@ from deepchem.data import get_task_dataset_minus_support 
class SupportGraphClassifier(Model):

  def __init__(self, sess, model,

               test_batch_size=10, support_batch_size=10,

               learning_rate=.001, similarity="cosine",

               beta1=.9, beta2=.999, **kwargs):

               learning_rate=.001, decay_steps=20, decay_rate=1.,

               similarity="cosine", **kwargs):

    """Builds a support-based classifier.



    See https://arxiv.org/pdf/1606.04080v1.pdf for definition of support.
@@ -40,6 +40,10 @@ class SupportGraphClassifier(Model): 
      Number of positive examples in support.

    n_neg: int

      Number of negative examples in support.

    decay_steps: int, optional

      Corresponds to argument decay_steps in tf.train.exponential_decay

    decay_rate: float, optional

      Corresponds to argument decay_rate in tf.train.exponential_decay

    """

    self.sess = sess

    self.similarity = similarity
@@ -48,6 +52,8 @@ class SupportGraphClassifier(Model): 
    self.support_batch_size = support_batch_size



    self.learning_rate = learning_rate

    self.decay_steps = decay_steps

    self.decay_rate = decay_rate

    self.epsilon = K.epsilon()



    self.add_placeholders()
@@ -61,9 +67,17 @@ class SupportGraphClassifier(Model): 


  def get_training_op(self, loss):

    """Attaches an optimizer to the graph."""

    opt = tf.train.AdamOptimizer(self.learning_rate)

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

    global_step = tf.Variable(0, trainable=False)

    learning_rate = tf.train.exponential_decay(

        self.learning_rate, global_step,

        self.decay_steps, self.decay_rate, staircase=True)

    opt = tf.train.AdamOptimizer(learning_rate)

    # Get train function

    return opt.minimize(self.loss_op, name="train")

    return opt.minimize(self.loss_op, name="train", global_step=global_step)

    #opt = tf.train.AdamOptimizer(self.learning_rate)

    #return opt.minimize(self.loss_op, name="train")

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



  def add_placeholders(self):

    """Adds placeholders to graph."""
@@ -139,6 +153,7 @@ class SupportGraphClassifier(Model): 
    # Create different support sets

    support_generator = SupportGenerator(dataset, range(n_tasks),

        n_pos, n_neg, n_trials, replace)

    recent_losses = []

    for ind, (task, support) in enumerate(support_generator):

      if ind % log_every_n_samples == 0:

        print("Sample %d from task %s" % (ind, str(task)))
@@ -150,7 +165,11 @@ class SupportGraphClassifier(Model): 
        ############################################################## DEBUG

        _, loss = self.sess.run([self.train_op, self.loss_op], feed_dict=feed_dict)

        if ind % log_every_n_samples == 0:

          print("\tloss is %s" % str(loss))

          mean_loss = np.mean(np.array(recent_losses))

          print("\tmean loss is %s" % str(mean_loss))

          recent_losses = []

        else:

          recent_losses.append(loss)

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



  def save(self):

examples/low_data/datasets.py

+6 −3
Original line number Diff line number Diff line
@@ -6,11 +6,15 @@ from __future__ import division 
from __future__ import unicode_literals



import os

import sys

import shutil

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
@@ -33,7 +37,7 @@ def load_tox21_ecfp(num_train=7200): 


  # Initialize transformers 

  transformers = [

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

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



  print("About to transform data")

  for transformer in transformers:
@@ -48,7 +52,6 @@ def load_tox21_convmol(base_dir=None, num_train=7200): 
  current_dir = os.path.dirname(os.path.realpath(__file__))

  dataset_file = os.path.join(

      current_dir, "../../datasets/tox21.csv.gz")

  #Make directories to store the raw and featurized datasets.



  # Featurize Tox21 dataset

  print("About to featurize Tox21 dataset.")
@@ -65,7 +68,7 @@ def load_tox21_convmol(base_dir=None, num_train=7200): 


  # Initialize transformers 

  transformers = [

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

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



  print("About to transform data")

  for transformer in transformers:

examples/low_data/tox_attn_one_fold.py

+10 −8
Original line number Diff line number Diff line
@@ -14,14 +14,14 @@ from datasets import load_tox21_convmol 
# Number of folds for split 

K = 4 

# Depth of attention module

max_depth = 4

max_depth = 8

# number positive/negative ligands

n_pos = 5 

n_pos = 1 

n_neg = 10

# Set batch sizes for network

test_batch_size = 100

test_batch_size = 128

support_batch_size = n_pos + n_neg

n_train_trials = 2000 

n_train_trials = 5000

n_eval_trials = 20

n_steps_per_trial = 1

# Sample supports without replacement (all pos/neg should be different)
@@ -51,18 +51,20 @@ support_model = dc.nn.SequentialSupportGraph(n_feat) 
# Adding 1st layer 

# output will be (n_atoms, 64)

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

# output will be (n_atoms, 64)

support_model.add_test(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

# output will be (n_atoms, 64)

support_model.add_support(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

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

# Addding 2nd layer

# output will be (n_atoms, 64)

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

support_model.add_test(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

support_model.add_support(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

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

# Adding 3rd layer

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

support_model.add_support(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

support_model.add_test(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

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



# Gather into molecules

support_model.add_test(dc.nn.GraphGather(test_batch_size))

examples/low_data/tox_res_one_fold.py

+14 −6
Original line number Diff line number Diff line
@@ -16,12 +16,12 @@ K = 4 
# Depth of attention module

max_depth = 4

# num positive/negative ligands

n_pos = 5 

n_pos = 1 

n_neg = 10

# Set batch sizes for network

test_batch_size = 100

test_batch_size = 128

support_batch_size = n_pos + n_neg

n_train_trials = 4000

n_train_trials = 11000

n_eval_trials = 20

n_steps_per_trial = 1

# Sample supports without replacement (all pos/neg should be different)
@@ -46,14 +46,22 @@ test_dataset = fold_datasets[-1] 
support_model = dc.nn.SequentialSupportGraph(n_feat)



# Add layers

# 1st conv layer + batchnorm

# output will be (n_atoms, 64)

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

# Need to add batch-norm separately to test/support due to differing

# shapes.

support_model.add_test(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

support_model.add_support(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

# 2nd conv layer + batchnorm

# output will be (n_atoms, 64)

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

support_model.add_test(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

support_model.add_support(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

# 3nd conv layer + batchnorm

# output will be (n_atoms, 64)

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

support_model.add_test(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))

support_model.add_support(dc.nn.BatchNormalization(epsilon=1e-5, mode=1))



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

support_model.add_test(dc.nn.GraphGather(test_batch_size))

support_model.add_support(dc.nn.GraphGather(support_batch_size))

CRA Git | Maintained and supported by SUSTech CRA and CCSE