Added flags DebugMode.allow_remove_inf and TensorType.value_eq_approx(…

theano/compile/debugmode.py

@@ -42,6 +42,9 @@ AddConfigVar('DebugMode.check_strides',
             "On difference: (0) - ignore, (1) warn, or (2) raise error"),
         IntParam(1, lambda i: i in (0,1,2)))
 
+AddConfigVar('DebugMode.allow_remove_inf',
+        "True -> do we allow an optimization to replace an inf with another value",
+        BoolParam(False))
 
 import logging
 _logger=logging.getLogger("theano.compile.debugmode")
@@ -599,7 +602,7 @@ def _lessbroken_deepcopy(a):
         assert rval.dtype == a.dtype
     return rval
 
-def _find_bad_optimizations0(order, reasons, r_vals):
+def _find_bad_optimizations0(order, reasons, r_vals, allow_remove_inf=False):
     """Use a simple algorithm to find broken optimizations.
     
     This algorithm is simple to understand, but sometimes when there's a problem it identifies
@@ -617,8 +620,8 @@ def _find_bad_optimizations0(order, reasons, r_vals):
                 new_r_val = r_vals[new_r]
                 r_val = r_vals[r]
                 assert r.type == new_r.type
-
-                if not r.type.values_eq_approx(r_val, new_r_val):
+                
+                if not r.type.values_eq_approx(r_val, new_r_val, allow_remove_inf=allow_remove_inf):
                     raise BadOptimization(old_r=r,
                             new_r=new_r, 
                             old_r_val=r_val, 
@@ -1045,7 +1048,8 @@ class _Linker(gof.link.LocalLinker):
                     if s[0] is not None:
                         print r, s
                     assert s[0] is None
-
+                
+                allow_remove_inf = self.maker.mode.allow_remove_inf
                 #try:
                 # compute the value of all variables
                 for i, (thunk_py, thunk_c, node) in enumerate(zip(thunks_py, thunks_c, order)):
@@ -1142,7 +1146,7 @@ class _Linker(gof.link.LocalLinker):
                                 #print >> sys.stderr, i, "DEBUGMODE clearing output", r
                                 # compares the version from thunk_py (in r_vals)
                                 # to the version produced by thunk_c (in storage_map)
-                                if not r.type.values_eq_approx(r_vals[r], storage_map[r][0]):
+                                if not r.type.values_eq_approx(r_vals[r], storage_map[r][0], allow_remove_inf=allow_remove_inf):
                                     raise BadCLinkerOutput(r, val_py=r_vals[r], val_c=storage_map[r][0])
                             else:
                                 #print >> sys.stderr, i, "DEBUGMODE storing reference output %x" % id(storage_map[r][0])
@@ -1162,7 +1166,7 @@ class _Linker(gof.link.LocalLinker):
                     if True: 
                         gc.collect()
 
-                _find_bad_optimizations(order, env.equivalence_tracker.reasons, r_vals)
+                _find_bad_optimizations(order, env.equivalence_tracker.reasons, r_vals, allow_remove_inf=allow_remove_inf)
 
                 #####
                 #  Postcondition: the input and output variables are in the storage map, nothing more
@@ -1523,6 +1527,12 @@ class DebugMode(Mode):
     different strides? (This can catch bugs, but is generally overly strict.) 0 no check, 1 warn, 2 err.
     """
 
+    allow_remove_inf = config.DebugMode.allow_remove_inf
+    """
+    Default False. Do we allow that an optimization remove inf value.
+    This is usefull to test stabilization optimization.
+    """
+
     # This function will be used to create a FunctionMaker in 
     # function_module.function
     def function_maker(self, i,o,m, *args, **kwargs):

theano/tensor/basic.py

@@ -488,7 +488,11 @@ class TensorType(Type):
         else:
             return False
     @staticmethod
-    def values_eq_approx(a, b):
+    def values_eq_approx(a, b, allow_remove_inf = False):
+        """
+        :param allow_remove_inf: If True, when their is an inf in a,
+                                 we allow any value in b in that position.
+        """
         if type(a) is numpy.ndarray and type(b) is numpy.ndarray:
             if a.shape != b.shape:
                 return False
@@ -516,7 +520,9 @@ class TensorType(Type):
                 # for now we use a home-made recipe, that should probably be
                 # revisited in the future.
                 a_missing = numpy.isnan(a)
-                if not a_missing.any():
+                a_inf = numpy.isinf(a)
+
+                if not (a_missing.any() or (allow_remove_inf and a_inf.any())):
                     # There are no missing values in a, thus this is not the
                     # reason why numpy.allclose(a, b) returned False.
                     _info('numpy allclose failed for abs_err %f and rel_err %f' %(
@@ -531,6 +537,9 @@ class TensorType(Type):
                         (atol + rtol * numpy.absolute(b)))
                 # Find places where both a and b have missing values.
                 both_missing = a_missing * numpy.isnan(b)
+                if allow_remove_inf:
+                    both_missing += a_inf
+                
                 # Combine all information.
                 return (cmp_elemwise + both_missing).all()