added binary operators and any()/all() and added log1p and expm1

Author: jsalvatier

pymc/CommonDeterministics.py

@@ -676,9 +676,9 @@ def op_to_jacobians(op, module):
 for op in ['div', 'truediv', 'floordiv', 'mod', 'divmod', 'pow', 'lshift', 'rshift', 'and', 'xor', 'or']:
     create_rl_bin_method(op, Variable, jacobians = op_to_jacobians(op, locals()))
 
-# # Binary operators
-# for op in ['lt', 'le', 'eq', 'ne', 'gt', 'ge']:
-#     create_bin_method(op ,Variable)
+# Binary operators
+for op in ['lt', 'le', 'eq', 'ne', 'gt', 'ge']:
+     create_bin_method(op ,Variable)
 
 # Unary operators
 neg_jacobians = {'self' : lambda self: -ones(shape(self))}

pymc/Node.py

@@ -35,6 +35,8 @@ def logp_gradient_of_set(variable_set, calculation_set = None):
     
     Returns a dictionary of the gradients.
     """
+    calculation_set = set(calculation_set) # this is important because using the 'in' operator on lists will use the __eq__method which will create a new deterministic
+    
     logp_gradients = {}
     for variable in variable_set:   
         logp_gradients[variable] = logp_gradient(variable, calculation_set)

pymc/NumpyDeterministics.py

@@ -10,17 +10,18 @@
 import inspect
 
 #accumulations 
+_boolean_accumulation_deterministics = ['any' , 'all']
 _accumulation_deterministics = ['sum']#['sum', 'prod']
 
 
 #transformations (broadcasted)
-_generic = ['abs', 'exp', 'log', 'sqrt']
+_generic = ['abs', 'exp', 'log', 'sqrt','expm1', 'log1p]
 _trig = ['sin', 'cos', 'tan', 'arcsin', 'arccos', 'arctan']
 _hyp_trig = ['sinh', 'cosh', 'tanh', 'arcsinh', 'arccosh', 'arctanh']
 _transformation_deterministics = _generic + _trig + _hyp_trig
 _misc_funcs1 = ['arctan2', 'hypot']
 
-__all__ = _accumulation_deterministics + _transformation_deterministics + _misc_funcs1
+__all__ = _accumulation_deterministics  + _boolean_accumulation_deterministics+ _transformation_deterministics + _misc_funcs1
 
 def deterministic_from_funcs(name, eval, jacobians=None, jacobian_formats=None, dtype=np.float, mv=False):
     """
@@ -160,6 +161,8 @@ def sum_jacobian_a (a, axis):
 sqrt_jacobians = {'x': lambda x : .5    * x **-.5}
 hypot_jacobians = {'x1' : lambda x1, x2 : (x1**2 + x2**2)**-.5 * x1,
                    'x2' : lambda x1, x2 : (x1**2 + x2**2)**-.5 * x2}
+expm1_jacobians = exp_jacobians
+log1p_jacobians = {'x' : lambda x : 1.0/(1.0 + x)}
 
 sin_jacobians = {'x' : lambda x : np.cos(x)  }
 cos_jacobians = {'x' : lambda x : -np.sin(x) }
@@ -196,6 +199,13 @@ def wrapped_function(a, axis = None):
     locals()[function_name] = deterministic_from_funcs(function_name, wrapped_function, jacobians, jacobian_formats = {'a' : 'accumulation_operation'})
 
 
+for function_name in _boolean_accumulation_deterministics:
+    wrapped_function = wrap_function_accum(find_element(function_name, np, error_on_fail = True))
+
+    locals()[function_name] = deterministic_from_funcs(function_name, wrapped_function)
+
+
+
 def wrapped_function_trans(function):
     def wrapped_function(x):
         return function(x)

pymc/tests/test_gradients.py

@@ -311,6 +311,7 @@ def test_gradients(self):
 
     def check_model_gradients(self, model):
 
+        model = set(model)
         # find the markov blanket 
         children = set([])
         for s in model: