transformations now are their own special thing

Author: jsalvatier

pymc/distributions/dist_math.py

@@ -6,7 +6,13 @@
 from __future__ import division
 from ..quickclass import *
 import theano.tensor as t
-from theano.tensor import sum, switch, log,exp, eq, neq, lt, gt, le, ge, zeros_like, cast,arange, round, max, min
+from theano.tensor import (
+        sum, switch, log,exp, 
+        eq, neq, lt, gt, le, ge, all, any, 
+        cast,arange, round, max, min, 
+        zeros_like, ones, ones_like,
+        concatenate, constant)
+
 
 from numpy import pi, inf, nan
 from special import gammaln

pymc/distributions/transforms.py

@@ -12,10 +12,12 @@ def TransformedDistribtuion():
             support = dist.support 
 
             def logp(x): 
-                return dist.logp(x) + jacobian_det(x)
+                return dist.logp(backward(x)) + jacobian_det(x)
 
             if hasattr(dist, "mode"): 
-                mode = backward(dist.mode)
+                mode = forward(dist.mode)
+            if hasattr(dist, "median"): 
+                mode = forward(dist.median)
 
             return locals()
 
@@ -30,5 +32,5 @@ def __str__():
 
 simplex = transform("simplex",
         lambda p: p[:-1],
-        lambda p: concatenate([p, 1- sum(p)]),
-        0)
+        lambda p: concatenate([p, 1- sum(p, keepdims = True)]),
+        lambda p: constant([0]))

pymc/model.py

@@ -76,18 +76,10 @@ def Var(model, name, dist, shape = 1, dtype = None, testval = try_defaults):
         model.factors.append(dist.logp(var))
         return var
 
-    def TransformedVar(model, name, dist, transform, logjacobian, shape = 1, dtype = None, testval = try_defaults): 
-        if not dtype:
-            dtype = default_type[dist.support]
-        
-        var = Variable('transformed_' + name, shape, dtype, get_test_val(dist, testval))
-
-        model.vars.append(var)
-
-        tvar = transform(var)
-        model.factors.append(dist.logp(tvar) + logjacobian(var))
+    def TransformedVar(model, name, dist, trans, shape = 1, dtype = None, testval = try_defaults): 
+        var = model.Var(trans.name + '_' + name, trans.apply(dist), shape, dtype, testval) 
 
-        return tvar, var
+        return trans.backward(var), var
 
     @property
     def logp(model):