changeset 117:3ef569b92fba

ported nnet_ops to new theano
author James Bergstra <bergstrj@iro.umontreal.ca>
date Wed, 07 May 2008 15:28:17 -0400
parents 9330d941fa1f
children 5fa46297191b
files _nnet_ops.py nnet_ops.py
diffstat 2 files changed, 57 insertions(+), 56 deletions(-) [+]
line wrap: on
line diff
--- a/_nnet_ops.py	Wed May 07 13:07:33 2008 -0400
+++ b/_nnet_ops.py	Wed May 07 15:28:17 2008 -0400
@@ -9,29 +9,31 @@
     def setUp(self):
         numpy.random.seed(9999)
     def test_elemwise(self):
-        TT.verify_grad(self, Sigmoid, [numpy.random.rand(3,4)])
+        TT.verify_grad(self, sigmoid, [numpy.random.rand(3,4)])
 
 class T_softplus(unittest.TestCase):
     def setUp(self):
         numpy.random.seed(9999)
     def test_elemwise(self):
-        TT.verify_grad(self, Softplus, [numpy.random.rand(3,4)])
+        TT.verify_grad(self, softplus, [numpy.random.rand(3,4)])
 
 class T_CrossentropySoftmax1Hot(unittest.TestCase):
     def setUp(self):
         numpy.random.seed(9999)
     def test0(self):
         y_idx = [0,1,3]
-        def output1(a,b):
-            return crossentropy_softmax_1hot_with_bias(a, b, y_idx)[0:1]
-        TT.verify_grad(self, output1, [numpy.random.rand(3,4),
+        class Dummy(object):
+            def make_node(self, a,b):
+                return crossentropy_softmax_1hot_with_bias(a, b, y_idx)[0:1]
+        TT.verify_grad(self, Dummy(), [numpy.random.rand(3,4),
             numpy.random.rand(4)])
 
     def test1(self):
         y_idx = [0,1,3]
-        def output1(a):
-            return crossentropy_softmax_1hot(a, y_idx)[0:1]
-        TT.verify_grad(self, output1, [numpy.random.rand(3,4)])
+        class Dummy(object):
+            def make_node(self, a):
+                return crossentropy_softmax_1hot(a, y_idx)[0:1]
+        TT.verify_grad(self, Dummy(), [numpy.random.rand(3,4)])
 
 
 
--- a/nnet_ops.py	Wed May 07 13:07:33 2008 -0400
+++ b/nnet_ops.py	Wed May 07 15:28:17 2008 -0400
@@ -1,5 +1,5 @@
 import theano
-from theano import tensor, gof, scalar
+from theano import tensor, scalar
 import numpy
 
 ############
@@ -7,7 +7,7 @@
 # SCALAR OPS
 #
 
-class ScalarSigmoid(scalar.FloatUnaryScalarOp):
+class ScalarSigmoid(scalar.UnaryScalarOp):
     @staticmethod
     def st_impl(x):
         if x < -30.0:
@@ -20,7 +20,7 @@
     def grad(self, (x,), (gz,)):
         y = scalar_sigmoid(x)
         return [gz * y * (1.0 - y)]
-    def c_foreach(self, (x,), (z,), sub):
+    def c_code(self, (x,), (z,), sub):
         if 'float' in self.inputs[0].dtype:
             return """%(z)s =
                 %(x)s < -30.0 
@@ -28,12 +28,11 @@
                 : %(x)s > 30.0 
                    ? 1.0
                    : 1.0 /(1.0+exp(-%(x)s));""" % locals()
-        raise NotImplementedError('only floatingpoint is implemented')
-scalar_sigmoid = gof.op.constructor(ScalarSigmoid)
-Sigmoid, sigmoid, SigmoidInplace, sigmoid_inplace =\
-        tensor.broadcast(ScalarSigmoid, 'Sigmoid')
+        return NotImplemented#Error('only floatingpoint is implemented')
+scalar_sigmoid = ScalarSigmoid(scalar.upgrade_to_float, name='scalar_sigmoid')
+sigmoid = tensor.Elemwise(scalar_sigmoid, name='sigmoid')
 
-class ScalarSoftplus(scalar.FloatUnaryScalarOp):
+class ScalarSoftplus(scalar.UnaryScalarOp):
     @staticmethod
     def static_impl(x):
         if x < -30.0:
@@ -45,7 +44,7 @@
         return ScalarSoftplus.static_impl(x)
     def grad(self, (x,), (gz,)):
         return [gz * scalar_sigmoid(x)]
-    def c_foreach(self, (x,), (z,), sub):
+    def c_code(self, (x,), (z,), sub):
         if 'float' in self.inputs[0].dtype:
             return """%(z)s =
                 %(x)s < -30.0 
@@ -53,10 +52,9 @@
                 : %(x)s > 30.0 
                    ? %(x)s
                    : log1p(exp(%(x)s));""" % locals()
-        raise NotImplementedError('only floating point x is implemented')
-scalar_softplus = gof.op.constructor(ScalarSoftplus)
-Softplus, softplus, SoftplusInplace, softplus_inplace =\
-        tensor.broadcast(ScalarSoftplus, 'Softplus')
+        return NotImplemented#Error('only floating point x is implemented')
+scalar_softplus = ScalarSoftplus(scalar.upgrade_to_float, name='scalar_softplus')
+softplus = tensor.Elemwise(scalar_softplus, name='softplus')
 
 
 ############
@@ -64,8 +62,7 @@
 # TENSOR OPS
 #
 
-
-class CrossentropySoftmax1HotWithBias(gof.op.Op):
+class CrossentropySoftmax1HotWithBias(theano.Op):
     """A special compound L{Op} for the output of neural-net classifiers.
 
     @type x: is a matrix of floats (32 or 64)
@@ -90,28 +87,31 @@
     """
     nin=3
     nout=2
-    def __init__(self, x, b, y_idx, **kwargs):
-        x = tensor._as_tensor(x)
-        b = tensor._as_tensor(b)
-        y_idx = tensor._as_tensor(y_idx)
-        if len(x.broadcastable) != 2 \
-                or x.dtype not in ['float32', 'float64']:
+    def __init__(self, **kwargs):
+        theano.Op.__init__(self, **kwargs)
+
+    def make_node(self, x, b, y_idx):
+        x = tensor.as_tensor(x)
+        b = tensor.as_tensor(b)
+        y_idx = tensor.as_tensor(y_idx)
+        if x.type.ndim != 2 \
+                or x.type.dtype not in ['float32', 'float64']:
             raise ValueError('x must be 2-d tensor of floats')
-        if len(b.broadcastable) != 1 \
-                or x.dtype not in ['float32', 'float64']:
+        if b.type.ndim != 1 \
+                or x.type.dtype not in ['float32', 'float64']:
             raise ValueError('x must be 1-d tensor of floats')
-        if len(y_idx.broadcastable) != 1 \
-                or y_idx.dtype not in ['int32', 'int64']:
+        if y_idx.type.ndim != 1 \
+                or y_idx.type.dtype not in ['int32', 'int64']:
             raise ValueError('x must be 1-d tensor of ints')
 
 #       TODO: Is this correct? It used to be y, not y_idx
-        nll = tensor.Tensor(x.dtype, y_idx.broadcastable)
+        nll = tensor.Tensor(x.type.dtype, 
+                y_idx.type.broadcastable).make_result()
 #        nll = Tensor(x.dtype, y.broadcastable)
-        sm = tensor.Tensor(x.dtype, x.broadcastable)
-        self.inputs = [x, b, y_idx]
-        self.outputs = [nll, sm]
-    def perform(self):
-        x, b, y_idx = [i.data for i in self.inputs]
+        sm = x.type.make_result()
+        return theano.Apply(self, [x, b, y_idx],[nll, sm])
+    def perform(self, node, input_storage, output_storage):
+        x, b, y_idx = input_storage
         if b.shape[0] != x.shape[1]:
             raise ValueError('b must have same number of columns as x')
         if y_idx.shape[0] != x.shape[0]:
@@ -124,14 +124,14 @@
             sm[i] = numpy.exp(row - numpy.max(row)) #softmax
             sm[i] *= 1.0 / numpy.sum(sm[i]) #vector scale
             nll[i] = -numpy.log( sm[i, y_idx[i]]) #cross-entropy
-        self.outputs[0].data = nll
-        self.outputs[1].data = sm
+        output_storage[0][0] = nll
+        output_storage[1][0] = sm
     def grad(self, (x, b, y_idx), (g_nll, g_sm)):
         if g_sm is not None:
             raise NotImplementedError()
         nll, sm = crossentropy_softmax_1hot_with_bias(x, b, y_idx)
-        dx = CrossentropySoftmax1HotWithBiasDx(g_nll, sm, y_idx).outputs[0]
-        db = tensor.Sum(dx, axis = [0]).outputs[0]
+        dx = CrossentropySoftmax1HotWithBiasDx()(g_nll, sm, y_idx)
+        db = tensor.sum(dx, axis = [0])
         return dx, db, None
 
     def c_headers(self): return ['<iostream>']
@@ -280,27 +280,26 @@
               //mat_at(y,i,0) =  - mat_at(x,i,t[i]) - mat_at(b,0,t[i]) + (discount_max ? maxi : 0.0) + log(sum);
         }
         """ % dict(locals(), **sub)
+crossentropy_softmax_1hot_with_bias = CrossentropySoftmax1HotWithBias()
 
-crossentropy_softmax_1hot_with_bias = \
-        gof.op.constructor(CrossentropySoftmax1HotWithBias)
-
-class CrossentropySoftmax1HotWithBiasDx (gof.op.Op):
+class CrossentropySoftmax1HotWithBiasDx (theano.Op):
     nin=3
     nout=1
     """Gradient wrt x of the CrossentropySoftmax1Hot Op"""
-    def __init__(self, dy, sm, y_idx,**kwargs):
-        dy = tensor._as_tensor(dy)
-        sm = tensor._as_tensor(sm)
-        y_idx = tensor._as_tensor(y_idx)
-        self.inputs = [dy, sm, y_idx]
-        self.outputs = [tensor.Tensor(sm.dtype, sm.broadcastable)]
-    def perform(self):
-        dy,sm,y_idx = [i.data for i in self.inputs]
+    def __init__(self, **kwargs):
+        theano.Op.__init__(self,**kwargs)
+    def make_node(self, dy, sm, y_idx,**kwargs):
+        dy = tensor.as_tensor(dy)
+        sm = tensor.as_tensor(sm)
+        y_idx = tensor.as_tensor(y_idx)
+        return theano.Apply(self, [dy, sm, y_idx],[sm.type.make_result()])
+    def perform(self, node, input_storage, output_storage):
+        dy,sm,y_idx = input_storage
         dx = numpy.zeros_like(sm)
         for i in xrange(sm.shape[0]):
             dx[i] = dy[i] * sm[i] #vector scale
             dx[i, y_idx[i]] -= dy[i] #scalar decrement
-        self.outputs[0].data = dx
+        output_storage[0][0] = dx
     def grad(self, *args):
         raise NotImplementedError()
     def c_code(self,  (dnll, sm, y_idx), (dx,), sub):