Merge pull request #1313 from rbharath/tfbinding

Multilayer SequenceDNN
deepchem · Jul 15, 2018 · da74fda · da74fda
2 parents 52c2350 + ad19082
commit da74fda
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 10 deletions.
diff --git a/deepchem/models/tensorgraph/models/sequence_dnn.py b/deepchem/models/tensorgraph/models/sequence_dnn.py
@@ -5,6 +5,7 @@
 tasks working with genomic data.
 """
 import tensorflow as tf
+import collections
 from deepchem.models import Sequential
 from deepchem.models.tensorgraph import layers
 from deepchem.data import NumpyDataset
@@ -14,9 +15,6 @@ class SequenceDNN(Sequential):
   """
   Sequence DNN models.
 
-  # TODO(rbharath): This model only supports one-conv layer. Extend
-  # so that conv layers of greater depth can be implemented.
-
   Parameters
   ----------
   seq_length : int
@@ -27,7 +25,7 @@ class SequenceDNN(Sequential):
       number of tasks. Default: 1.
   num_filters : list[int] | tuple[int]
       number of convolutional filters in each layer. Default: (15,).
-  conv_width : list[int] | tuple[int]
+  kernel_size: list[int] | tuple[int]
       width of each layer's convolutional filters. Default: (15,).
   pool_width : int
       width of max pooling after the last layer. Default: 35.
@@ -44,17 +42,25 @@ def __init__(self,
                loss,
                use_RNN=False,
                num_tasks=1,
-               num_filters=15,
-               kernel_size=15,
+               num_filters=(15,),
+               kernel_size=(15,),
                pool_width=35,
                L1=0,
                dropout=0.0,
                verbose=True,
                **kwargs):
     super(SequenceDNN, self).__init__(loss, **kwargs)
+    if not isinstance(num_filters, collections.Sequence) or not isinstance(
+        kernel_size, collections.Sequence):
+      raise ValueError("num_filters and kernel_size must be lists")
+    if not len(num_filters) == len(kernel_size):
+      raise ValueError("num_filters and kernel_size must be of same length")
+    n_layers = len(num_filters)
     self.num_tasks = num_tasks
     self.verbose = verbose
-    self.add(layers.Conv2D(num_filters, kernel_size=kernel_size))
-    self.add(layers.Dropout(dropout))
+    for layer in range(n_layers):
+      self.add(
+          layers.Conv2D(num_filters[layer], kernel_size=kernel_size[layer]))
+      self.add(layers.Dropout(dropout))
     self.add(layers.Flatten())
     self.add(layers.Dense(self.num_tasks, activation_fn=tf.nn.relu))
diff --git a/deepchem/models/tensorgraph/tests/test_sequencednn.py b/deepchem/models/tensorgraph/tests/test_sequencednn.py
@@ -19,7 +19,7 @@ def test_seq_dnn_singlefilter_train(self):
     #  #    False: num_sequences / num_negatives
     #  #} if not multitask else None,
     dataset = dc.data.NumpyDataset(X, y)
-    model = dc.models.SequenceDNN(50, "binary_crossentropy", num_filters=1)
+    model = dc.models.SequenceDNN(50, "binary_crossentropy", num_filters=[1])
     model.fit(dataset, nb_epoch=1)
 
   def test_seq_dnn_multifilter_train(self):
@@ -32,5 +32,14 @@ def test_seq_dnn_multifilter_train(self):
     #  #    False: num_sequences / num_negatives
     #  #} if not multitask else None,
     dataset = dc.data.NumpyDataset(X, y)
-    model = dc.models.SequenceDNN(50, "binary_crossentropy", num_filters=15)
+    model = dc.models.SequenceDNN(50, "binary_crossentropy", num_filters=[15])
+    model.fit(dataset, nb_epoch=1)
+
+  def test_seq_dnn_multilayer_train(self):
+    """Test SequenceDNN with multiple layers works."""
+    X = np.random.rand(10, 1, 4, 50)
+    y = np.random.randint(0, 2, size=(10, 1))
+    dataset = dc.data.NumpyDataset(X, y)
+    model = dc.models.SequenceDNN(
+        50, "binary_crossentropy", num_filters=[1, 1], kernel_size=[15, 15])
     model.fit(dataset, nb_epoch=1)