--- a/sandbox/rbm/README.txt	Tue Jul 08 21:42:21 2008 -0400
+++ b/sandbox/rbm/README.txt	Tue Jul 08 23:59:57 2008 -0400
@@ -1,1 +1,4 @@
+An RBM with binomial units trained with CD-1.
+by Joseph Turian
+    
 This seems to work fine.

--- a/sandbox/rbm/globals.py	Tue Jul 08 21:42:21 2008 -0400
+++ b/sandbox/rbm/globals.py	Tue Jul 08 23:59:57 2008 -0400
@@ -4,7 +4,7 @@
 
 INPUT_DIMENSION = 1000
 #INPUT_DIMENSION = 100
-HIDDEN_DIMENSION = 100
+HIDDEN_DIMENSION = 10
 #HIDDEN_DIMENSION = 20
 #HIDDEN_DIMENSION = 6
 LEARNING_RATE = 0.1

--- a/sandbox/rbm/main.py	Tue Jul 08 21:42:21 2008 -0400
+++ b/sandbox/rbm/main.py	Tue Jul 08 23:59:57 2008 -0400
@@ -1,6 +1,7 @@
 #!/usr/bin/python
 """
-    An RBM with binomial units trained with CD-1.
+Simple SGD RBM training.
+(An example of how to use the model.)
 """
 
 
@@ -22,4 +23,4 @@
     instance = nonzero_instances[i % len(nonzero_instances)]
 
     # SGD update over instance
-    model.update(instance)
+    model.update([instance])

--- a/sandbox/rbm/model.py	Tue Jul 08 21:42:21 2008 -0400
+++ b/sandbox/rbm/model.py	Tue Jul 08 23:59:57 2008 -0400
@@ -14,13 +14,21 @@
 random.seed(globals.SEED)
 
 import pylearn.nnet_ops
+import pylearn.sparse_instance
 
 def sigmoid(v):
+    """
+    @todo: Move to pylearn.more_numpy
+    @todo: Fix to avoid floating point overflow.
+    """
 #    if x < -30.0: return 0.0
 #    if x > 30.0: return 1.0 
     return 1.0 / (1.0 + numpy.exp(-v))
 
 def sample(v):
+    """
+    @todo: Move to pylearn.more_numpy
+    """
     assert len(v.shape) == 2
     x = numpy.zeros(v.shape)
     for j in range(v.shape[0]):
@@ -35,14 +43,28 @@
     Compute the crossentropy of binary output wrt binary target.
     @note: We do not sum, crossentropy is computed by component.
     @todo: Rewrite as a scalar, and then broadcast to tensor.
+    @todo: Move to pylearn.more_numpy
+    @todo: Fix to avoid floating point overflow.
     """
     return -(target * numpy.log(output) + (1 - target) * numpy.log(1 - output))
 
 
 class Model:
+    """
+    @todo: input dimensions should be stored here! not as a global.
+    """
     def __init__(self):
         self.parameters = parameters.Parameters(randomly_initialize=True)
 
+    def sample(self, instances, iterations=1):
+        v = pylearn.sparse_instance.to_vector(instances, globals.INPUT_DIMENSION)
+        for i in range(iterations):
+            q = sigmoid(self.parameters.b + dot(v, self.parameters.w))
+            h = sample(q)
+            p = sigmoid(self.parameters.c + dot(h, self.parameters.w.T))
+            v = sample(p)
+        return v
+
     def update(self, instances):
         """
         Update the L{Model} using one training instance.
@@ -50,12 +72,8 @@
         @todo: Should assert that nonzero_indices and zero_indices
         are correct (i.e. are truly nonzero/zero).
         """
-        v0 = numpy.zeros((len(instances), globals.INPUT_DIMENSION))
         minibatch = len(instances)
-        for i in range(minibatch):
-            for idx in instances[i].keys():
-                v0[i][idx] = instances[i][idx]
-
+        v0 = pylearn.sparse_instance.to_vector(instances, globals.INPUT_DIMENSION)
         q0 = sigmoid(self.parameters.b + dot(v0, self.parameters.w))
         h0 = sample(q0)
         p0 = sigmoid(self.parameters.c + dot(h0, self.parameters.w.T))

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/sparse_instance.py	Tue Jul 08 23:59:57 2008 -0400
@@ -0,0 +1,22 @@
+"""
+Sparse instances.
+Each instance is represented as dict with key dimension.
+Dimensions not present in the dict have value 0.
+"""
+
+from numpy import zeros
+
+def to_vector(instances, dimensions):
+    """
+    Convert sparse instances to vectors.
+    @type instances: list of sparse instances
+    @param dimensions: The number of dimensions in each instance.
+    @rtype: numpy matrix (instances x dimensions)
+    """
+    assert isinstance(instances, list)
+    v = zeros((len(instances), dimensions))
+    l = len(instances)
+    for i in range(l):
+        for idx in instances[i].keys():
+            v[i][idx] = instances[i][idx]
+    return v