Merge pull request #3768 from nouiz/nanguardmode

doc/extending/extending_theano.txt

@@ -913,6 +913,23 @@ documentation:
 .. automodule:: theano.misc.doubleop
     :members:
 
+NanGuardMode and AllocEmpty
+---------------------------
+
+NanGuardMode help users find where in the graph NaN appear. But
+sometimes, we want some variables to not be checked. For example, in
+the old GPU back-end, we use a float32 CudaNdarray to store the MRG
+random number generator state (they are integers). So if NanGuardMode
+check it, it will generate false positive. Another case is related to
+[Gpu]AllocEmpty or some computation on it (like done by Scan).
+
+You can tell NanGuardMode to do not check a variable with:
+``variable.tag.nan_guard_mode_check``. Also, this tag automatically
+follow that variable during optimization. This mean if you tag a
+variable that get replaced by an inplace version, it will keep that
+tag.
+
+
 Final Note
 ----------
 

theano/compile/function_module.py

@@ -199,7 +199,7 @@ def std_fgraph(input_specs, output_specs, accept_inplace=False):
     return fgraph, list(map(SymbolicOutput, updates))
 
 
-std_fgraph.features = [gof.toolbox.PreserveNames]
+std_fgraph.features = [gof.toolbox.PreserveVariableAttributes]
 
 
 class AliasedMemoryError(Exception):

theano/compile/mode.py

@@ -416,7 +416,7 @@ def get_mode(orig_string):
         elif string == 'NanGuardMode':
             # need to import later to break circular dependency.
             from .nanguardmode import NanGuardMode
-            # DebugMode use its own linker.
+            # NanGuardMode use its own linker.
             ret = NanGuardMode(True, True, True, optimizer=config.optimizer)
         else:
             # This might be required if the string is 'ProfileMode'

theano/compile/nanguardmode.py

@@ -297,11 +297,13 @@ class NanGuardMode(Mode):
                 # If the input is the result of computation, then we
                 # don't need to check it. It is already done after the
                 # computation.
-                if var.owner is not None:
+                if (var.owner is None and
+                        getattr(var.tag, 'nan_guard_mode_check', True)):
                     do_check_on(x[0], node, fn, True)
             fn()
             outputs = fn.outputs
-            for x in outputs:
+            for x, var in zip(outputs, node.outputs):
+                if getattr(var.tag, 'nan_guard_mode_check', True):
                     do_check_on(x[0], node, fn, False)
 
         wrap_linker = theano.gof.WrapLinker([theano.gof.OpWiseCLinker()],

theano/gof/toolbox.py

@@ -455,10 +455,28 @@ class PrintListener(Feature):
 
 
 class PreserveNames(Feature):
+    """
+    This preserve some variables names during optimization.
+
+    Deprecated. We need to keep it to allow unpickling.
+    """
+
+    def on_change_input(self, fgraph, node, i, r, new_r, reason=None):
+        if r.name is not None and new_r.name is None:
+            new_r.name = r.name
+
+
+class PreserveVariableAttributes(Feature):
+    """
+    This preserve some variables attributes and tag during optimization.
+    """
 
     def on_change_input(self, fgraph, node, i, r, new_r, reason=None):
         if r.name is not None and new_r.name is None:
             new_r.name = r.name
+        if getattr(r.tag, 'nan_guard_mode_check', False) and getattr(
+                new_r.tag, 'nan_guard_mode_check', False) is False:
+            new_r.tag.nan_guard_mode_check = r.tag.nan_guard_mode_check
 
 
 class NoOutputFromInplace(Feature):

theano/misc/do_nightly_build_send

@@ -51,7 +51,7 @@ def mysend(subject, file):
     
     # Open the files in binary mode.  Let the MIMEImage class automatically
     # guess the specific image type.
-    fp = open(file, 'rb')
+    with open(file, 'rb') as fp:
         s=fp.read()
     failures=0
     errors=0
@@ -115,7 +115,6 @@ def mysend(subject, file):
     s = ("Summary of the output:\n\n" + filter_output(open(file)) +
          "\n\nFull output:\n\n" + s)
     img = MIMEText(s)
-    fp.close()
     msg.attach(img)
     
 # Send the email via our own SMTP server.

theano/sandbox/cuda/basic_ops.py

@@ -2436,7 +2436,7 @@ class GpuReshape(tensor.Reshape, GpuOp):
 
     """
 
-    # __hash__, __eq__, __str__ come from tensor.Subtensor
+    # __hash__, __eq__, __str__ come from tensor.Reshape
     def make_node(self, x, shp):
         x = as_cuda_ndarray_variable(x)
         shp = tensor.as_tensor_variable(shp)
@@ -3680,6 +3680,7 @@ class GpuAllocEmpty(GpuOp):
         # The outut can contain nan/inf.  output.type is a new
         # instance, so we can do this only for that variable.
         output.type.filter_checks_isfinite = False
+        output.tag.nan_guard_mode_check = False
         return Apply(self, shape, [output])
 
     def debug_perform(self, node, inputs, out_):

theano/sandbox/cuda/opt.py

@@ -1190,14 +1190,16 @@ def local_gpu_incsubtensor(node):
                 # The IncSubtensor upcast to float32 y, so we do it
                 # explicitly to move it to the GPU.
                 y = y.astype('float32')
-
-            return [GpuIncSubtensor(
+            ret = GpuIncSubtensor(
                 incsubt.idx_list,
                 inplace=incsubt.inplace,
                 set_instead_of_inc=incsubt.set_instead_of_inc)(
                     as_cuda_ndarray_variable(x),
                     as_cuda_ndarray_variable(y),
-                    *coords)]
+                    *coords)
+            ret.tag.nan_guard_mode_check = getattr(
+                host_output.tag, 'nan_guard_mode_check', True)
+            return [ret]
     # Incrementing a float32 x results in a float32
     # output even if y is float64, so we can downcast
     # y to put it on GPU
@@ -1221,10 +1223,16 @@ def local_gpu_incsubtensor(node):
                 y = tensor.cast(y, 'float32')
             gpu_y = as_cuda_ndarray_variable(y)
         if go_gpu:
-            return [host_from_gpu(GpuIncSubtensor(
+            ret = GpuIncSubtensor(
                 node.op.idx_list, inplace=node.op.inplace,
                 set_instead_of_inc=node.op.set_instead_of_inc)(
-                    gpu_x, gpu_y, *coords))]
+                    gpu_x, gpu_y, *coords)
+
+            val = getattr(node.outputs[0].tag, 'nan_guard_mode_check', True)
+            ret.tag.nan_guard_mode_check = val
+            ret = host_from_gpu(ret)
+            ret.tag.nan_guard_mode_check = val
+            return [ret]
     return False
 
 
@@ -2532,6 +2540,20 @@ def local_gpu_allocempty(node):
     return False
 
 
+# Don't register by default.
+@gof.local_optimizer([GpuAllocEmpty])
+def local_gpu_alloc_empty_to_zeros(node):
+    # We need the exact match as GpuAlloc inherit from GpuAllocEmpty.
+    if type(node.op) is GpuAllocEmpty:
+        return [gpu_alloc(theano.tensor.constant(0, dtype='float32'),
+                          *node.inputs)]
+optdb.register('local_gpu_alloc_empty_to_zeros',
+               theano.tensor.opt.in2out(local_gpu_alloc_empty_to_zeros),
+               # After move to gpu and merge2, before inplace.
+               49.3,
+               'alloc_empty_to_zeros',)
+
+
 def typeInfer(node):
     return typeConstructor
 

theano/sandbox/gpuarray/basic_ops.py

@@ -721,6 +721,7 @@ class GpuAllocEmpty(HideC, Alloc):
         output.tag.values_eq_approx = tensor.type.values_eq_approx_always_true
         # The outut can contain nan/inf.
         output.type.filter_checks_isfinite = False
+        output.tag.nan_guard_mode_check = False
         return Apply(self, sh, [output])
 
     def debug_perform(self, node, inputs, out_, ctx):

theano/sandbox/gpuarray/opt.py

@@ -300,6 +300,21 @@ def local_gpualloc_memset_0(node):
             return [new_op(*node.inputs)]
 
 
+# Don't register by default.
+@gof.local_optimizer([GpuAllocEmpty])
+def local_gpua_alloc_empty_to_zeros(node):
+    if isinstance(node.op, GpuAllocEmpty):
+        context_name = infer_context_name(*node.inputs)
+        z = numpy.asarray(0, dtype=node.outputs[0].dtype)
+        return [GpuAlloc()(as_gpuarray_variable(z, context_name),
+                           *node.inputs)]
+optdb.register('local_gpua_alloc_empty_to_zeros',
+               theano.tensor.opt.in2out(local_gpua_alloc_empty_to_zeros),
+               # After move to gpu and merge2, before inplace.
+               49.3,
+               'alloc_empty_to_zeros',)
+
+
 @register_opt()
 @local_optimizer([GpuContiguous])
 def local_gpu_contiguous_gpu_contiguous(node):
@@ -569,9 +584,13 @@ def local_gpua_subtensor(node, context_name):
 @register_opt('fast_compile')
 @op_lifter([tensor.IncSubtensor])
 def local_gpua_incsubtensor(node, context_name):
-    return GpuIncSubtensor(node.op.idx_list, node.op.inplace,
+    op = GpuIncSubtensor(node.op.idx_list, node.op.inplace,
                          node.op.set_instead_of_inc,
                          node.op.destroyhandler_tolerate_aliased)
+    ret = op(*node.inputs)
+    val = getattr(node.outputs[0].tag, 'nan_guard_mode_check', True)
+    ret.tag.nan_guard_mode_check = val
+    return ret
 
 
 @register_opt('fast_compile')

theano/scan_module/scan_utils.py

@@ -620,7 +620,9 @@ def expand_empty(tensor_var, size):
     new_shape = [size + shapes[0]] + shapes[1:]
     empty = tensor.AllocEmpty(tensor_var.dtype)(*new_shape)
 
-    return tensor.set_subtensor(empty[:shapes[0]], tensor_var)
+    ret = tensor.set_subtensor(empty[:shapes[0]], tensor_var)
+    ret.tag.nan_guard_mode_check = False
+    return ret
 
 
 def equal_computations(xs, ys, in_xs=None, in_ys=None):

theano/tensor/basic.py

@@ -6241,6 +6241,13 @@ class AllocEmpty(gof.Op):
         # The outut can contain nan/inf.  output.type is a new
         # instance, so we can do this only for that variable.
         output.type.filter_checks_isfinite = False
+
+        # We can't reuse filter_checks_isfinite as by default it is
+        # False and it is set to true only in DebugMode.
+        # We can't set it in the type as other make_node can reuse the type.
+        # We can't set it in the variable as it isn't copied when we copy
+        # the variale. So we set it in the tag.
+        output.tag.nan_guard_mode_check = False
         return Apply(self, shape, [output])
 
     def debug_perform(self, node, inputs, out_):

theano/tensor/opt.py

@@ -1733,6 +1733,26 @@ def local_useless_alloc(node):
             return [node.inputs[0]]
 
 
+# Don't register by default.
+@gof.local_optimizer([T.AllocEmpty])
+def local_alloc_empty_to_zeros(node):
+    """This convert AllocEmpty to Alloc of 0.
+
+    This help investigate NaN with NanGuardMode.  Not registered by
+    default. To activate it, use the Theano flag
+    optimizer_including=alloc_empty_to_zeros. This also enable
+    the GPU version of this optimizations.
+
+    """
+    if isinstance(node.op, T.AllocEmpty):
+        return [T.zeros(node.inputs, dtype=node.outputs[0].dtype)]
+compile.optdb.register('local_alloc_empty_to_zeros',
+                       in2out(local_alloc_empty_to_zeros),
+                       # After move to gpu and merge2, before inplace.
+                       49.3,
+                       'alloc_empty_to_zeros',)
+
+
 @register_specialize
 @register_canonicalize
 @gof.local_optimizer([T.shape])
@@ -3043,6 +3063,9 @@ def local_inplace_setsubtensor(node):
             set_instead_of_inc=node.op.set_instead_of_inc,
             destroyhandler_tolerate_aliased=dta)
         new_node = new_op(*node.inputs)
+        val = getattr(node.outputs[0].tag, 'nan_guard_mode_check', True)
+        new_node.tag.nan_guard_mode_check = val
+
         # Copy stacktrace from original outputs to new outputs.
         # This is sensible, because the new operation is the
         # same as the old one, but now with different attributes.