Merge pull request #84 from CStephenson970/decorator_fix

Change dnnseparate models to run on CPU by default

Author: karllab41

src/dnnseparate/DANmodel.py

@@ -6,7 +6,8 @@
 class DANModel:
     def __init__(self, F=257, num_speakers=251,
                  layer_size=600, embedding_size=40,
-                 nonlinearity='logistic',normalize=False):
+                 nonlinearity='logistic',normalize=False,
+                 device='/cpu:0'):
         """
         Initializes a Deep Attractor Network[i].  Default architecture is the
         same as for the Lab41 model and the deep clustering model.
@@ -23,6 +24,7 @@ def __init__(self, F=257, num_speakers=251,
             nonlinearity: Nonlinearity to use in BLSTM layers (default logistic)
             normalize: Do you normalize vectors coming into the final layer?
                        (default False)
+            device: Which device to run the model on
         """
 
         self.F = F
@@ -34,28 +36,28 @@ def __init__(self, F=257, num_speakers=251,
 
         self.graph = tf.Graph()
         with self.graph.as_default():
+            with tf.device(device):
+                # Placeholder tensor for the magnitude spectrogram
+                self.S = tf.placeholder("float", [None, None, self.F])
 
-            # Placeholder tensor for the magnitude spectrogram
-            self.S = tf.placeholder("float", [None, None, self.F])
+                # Placeholder tensor for the input data
+                self.X = tf.placeholder("float", [None, None, self.F])
 
-            # Placeholder tensor for the input data
-            self.X = tf.placeholder("float", [None, None, self.F])
+                # Placeholder tensor for the labels/targets
+                self.y = tf.placeholder("float", [None, None, self.F, None])
 
-            # Placeholder tensor for the labels/targets
-            self.y = tf.placeholder("float", [None, None, self.F, None])
+                # Placeholder for the speaker indicies
+                self.I = tf.placeholder(tf.int32, [None,None])
 
-            # Placeholder for the speaker indicies
-            self.I = tf.placeholder(tf.int32, [None,None])
-
-            # Define the speaker vectors to use during training
-            self.speaker_vectors = tf_utils.weight_variable(
+                # Define the speaker vectors to use during training
+                self.speaker_vectors = tf_utils.weight_variable(
                                        [self.num_speakers,self.embedding_size],
                                        tf.sqrt(2/self.embedding_size))
 
-            # Model methods
-            self.network
-            self.cost
-            self.optimizer
+                # Model methods
+                self.network
+                self.cost
+                self.optimizer
 
             # Saver
             self.saver = tf.train.Saver()

src/dnnseparate/L41model.py

@@ -6,7 +6,8 @@
 class L41Model:
     def __init__(self, F=257, num_speakers=251,
                  layer_size=600, embedding_size=40,
-                 nonlinearity='logistic',normalize=False):
+                 nonlinearity='logistic',normalize=False,
+                 device='/cpu:0'):
         """
         Initializes Lab41's clustering model.  Default architecture comes from
         the parameters used by the best performing model in the paper[1].
@@ -21,8 +22,9 @@ def __init__(self, F=257, num_speakers=251,
             layer_size: Size of BLSTM layers
             embedding_size: Dimension of embedding vector
             nonlinearity: Nonlinearity to use in BLSTM layers (default logistic)
-            normalize: Do you normalize vectors coming into the final layer? 
+            normalize: Do you normalize vectors coming into the final layer?
                        (default False)
+            device: Which device to run the model on
         """
 
         self.F = F
@@ -33,26 +35,27 @@ def __init__(self, F=257, num_speakers=251,
         self.normalize = normalize
 
         self.graph = tf.Graph()
-        with self.graph.as_default():
 
-            # Placeholder tensor for the input data
-            self.X = tf.placeholder("float", [None, None, self.F])
+        with self.graph.as_default():
+            with tf.device(device):
+                # Placeholder tensor for the input data
+                self.X = tf.placeholder("float", [None, None, self.F])
 
-            # Placeholder tensor for the labels/targets
-            self.y = tf.placeholder("float", [None, None, self.F, None])
+                # Placeholder tensor for the labels/targets
+                self.y = tf.placeholder("float", [None, None, self.F, None])
 
-            # Placeholder for the speaker indicies
-            self.I = tf.placeholder(tf.int32, [None,None])
+                # Placeholder for the speaker indicies
+                self.I = tf.placeholder(tf.int32, [None,None])
 
-            # Define the speaker vectors to use during training
-            self.speaker_vectors = tf_utils.weight_variable(
+                # Define the speaker vectors to use during training
+                self.speaker_vectors = tf_utils.weight_variable(
                                        [self.num_speakers,self.embedding_size],
                                        tf.sqrt(2/self.embedding_size))
 
-            # Model methods
-            self.network
-            self.cost
-            self.optimizer
+                # Model methods
+                self.network
+                self.cost
+                self.optimizer
 
             # Saver
             self.saver = tf.train.Saver()

src/dnnseparate/deep_clustering_model.py

@@ -6,7 +6,8 @@
 class DeepClusteringModel:
     def __init__(self, F=257,
                  layer_size=600, embedding_size=40,
-                 nonlinearity='logistic'):
+                 nonlinearity='logistic',
+                 device='/cpu:0'):
         """
         Initializes the deep clustering model from [1].  Defaults correspond to
         the parameters used by the best performing model in the paper.
@@ -21,6 +22,7 @@ def __init__(self, F=257,
             layer_size: Size of BLSTM layers
             embedding_size: Dimension of embedding vector
             nonlinearity: Nonlinearity to use in BLSTM layers
+            device: Which device to run the model on
         """
 
         self.F = F
@@ -30,17 +32,17 @@ def __init__(self, F=257,
 
         self.graph = tf.Graph()
         with self.graph.as_default():
+            with tf.device(device):
+                # Placeholder tensor for the input data
+                self.X = tf.placeholder("float", [None, None, self.F])
 
-            # Placeholder tensor for the input data
-            self.X = tf.placeholder("float", [None, None, self.F])
+                # Placeholder tensor for the labels/targets
+                self.y = tf.placeholder("float", [None, None, self.F, None])
 
-            # Placeholder tensor for the labels/targets
-            self.y = tf.placeholder("float", [None, None, self.F, None])
-
-            # Model methods
-            self.network
-            self.cost
-            self.optimizer
+                # Model methods
+                self.network
+                self.cost
+                self.optimizer
 
             # Saver
             self.saver = tf.train.Saver()

src/utils/tf_utils.py

@@ -17,9 +17,8 @@ def scope_decorator(function):
     @functools.wraps(function)
     def decorator(self):
         if not hasattr(self,attribute):
-            with tf.device("/cpu:0"):
-                with tf.variable_scope(name):
-                    setattr(self,attribute,function(self))
+            with tf.variable_scope(name):
+                setattr(self,attribute,function(self))
         return getattr(self,attribute)
 
     return decorator