Update opts using values_eq_approx to the new interface.

theano/sandbox/cuda/extra_ops.py

@@ -460,5 +460,5 @@ def use_gpu_cumsum(node):
             axis = 0
 
         ret = host_from_gpu(GpuCumsum(axis)(x))
-        ret.values_eq_approx = values_eq_approx_high_tol
+        ret.tag.values_eq_approx = values_eq_approx_high_tol
         return [ret]

theano/sandbox/cuda/opt.py

@@ -1550,7 +1550,7 @@ def local_gpu_conv(node):
                            gpu_from_host(kern))
             out = tensor.patternbroadcast(out,
                                           node.outputs[0].broadcastable)
-            out.values_eq_approx = values_eq_approx_high_tol
+            out.tag.values_eq_approx = values_eq_approx_high_tol
             # in some case the ConvOp broadcast the last 2 dimensions
             # differently then the gpu ConvOp
             return [out]
@@ -1569,7 +1569,7 @@ def local_gpu_conv(node):
             out = tensor.patternbroadcast(
                 host_from_gpu(out),
                 node.outputs[0].broadcastable)
-            out.values_eq_approx = values_eq_approx_high_tol
+            out.tag.values_eq_approx = values_eq_approx_high_tol
             # in some case the ConvOp broadcast the last 2 dimensions
             # differently then the gpu ConvOp
             return [out]
@@ -2697,7 +2697,7 @@ def local_conv2d_gpu_conv(node):
             # out is on the GPU because both inputs are.
             out = theano.tensor.patternbroadcast(out,
                                                  node.outputs[0].broadcastable)
-            out.values_eq_approx = values_eq_approx_high_tol
+            out.tag.values_eq_approx = values_eq_approx_high_tol
             return [out]
 
     if isinstance(node.op, BaseAbstractConv2d):
@@ -2724,7 +2724,7 @@ def local_conv2d_gpu_conv(node):
             out = theano.tensor.patternbroadcast(
                 out,
                 node.outputs[0].broadcastable)
-            out.values_eq_approx = values_eq_approx_high_tol
+            out.tag.values_eq_approx = values_eq_approx_high_tol
             # If the original output was on CPU, we have to transfer it
             if isinstance(node.outputs[0].type, tensor.TensorType):
                 return [tensor.as_tensor_variable(out)]

theano/tensor/nnet/nnet.py

@@ -752,7 +752,7 @@ def local_logsoftmax(node):
         inVars = node.inputs[0].owner.inputs[0]
         new_op = LogSoftmax()
         ret = new_op(inVars)
-        ret.values_eq_approx = values_eq_approx_remove_inf
+        ret.tag.values_eq_approx = values_eq_approx_remove_inf
         return [ret]
 
 
@@ -787,7 +787,7 @@ def local_logsoftmax_grad(node):
             grads = tensor.alloc(grads, grads.shape[0], sm.shape[1])
 
         ret = grads - tensor.sum(grads, axis=1, keepdims=True) * sm
-        ret.values_eq_approx = values_eq_approx_remove_nan
+        ret.tag.values_eq_approx = values_eq_approx_remove_nan
         return [ret]
 
 

theano/tensor/nnet/sigm.py

@@ -269,7 +269,7 @@ def local_ultra_fast_sigmoid(node):
             # Other test could fail without good reason.
             return tensor.TensorType.values_eq_approx(a, b, atol=0.02)
         # Let DebugMode know that there this opt approx the values.
-        out.values_eq_approx = values_eq_approx_remove_low_prec
+        out.tag.values_eq_approx = values_eq_approx_remove_low_prec
         return [out]
 theano.compile.optdb['uncanonicalize'].register("local_ultra_fast_sigmoid",
                                                 local_ultra_fast_sigmoid)
@@ -308,7 +308,7 @@ def local_hard_sigmoid(node):
             # Other test could fail without good reason.
             return tensor.TensorType.values_eq_approx(a, b, atol=0.1)
         # Let DebugMode know that there this opt approx the values.
-        out.values_eq_approx = values_eq_approx_remove_low_prec
+        out.tag.values_eq_approx = values_eq_approx_remove_low_prec
         return [out]
 theano.compile.optdb['uncanonicalize'].register("local_hard_sigmoid",
                                                 local_hard_sigmoid)

theano/tensor/opt.py

@@ -3705,7 +3705,7 @@ def local_mul_switch_sink(node):
 
                     fct = [T.switch(switch.inputs[0], 0,
                                     fmul)]
-                    fct[0].values_eq_approx = values_eq_approx_remove_nan
+                    fct[0].tag.values_eq_approx = values_eq_approx_remove_nan
 
                     # Copy over stacktrace for switch op from both previous
                     #  elementwise multiplication op and previous switch op,
@@ -3729,7 +3729,7 @@ def local_mul_switch_sink(node):
 
                     fct = [T.switch(switch.inputs[0],
                                     fmul, 0)]
-                    fct[0].values_eq_approx = values_eq_approx_remove_nan
+                    fct[0].tag.values_eq_approx = values_eq_approx_remove_nan
 
                     # Copy over stacktrace for switch op from both previous
                     # elementwise multiplication op and previous switch op,
@@ -3773,7 +3773,7 @@ def local_div_switch_sink(node):
 
                 fct = [T.switch(switch.inputs[0], 0,
                                 fdiv)]
-                fct[0].values_eq_approx = values_eq_approx_remove_nan
+                fct[0].tag.values_eq_approx = values_eq_approx_remove_nan
 
                 # Copy over stacktrace for switch op from both previous
                 # elementwise division op and previous switch op,
@@ -3795,7 +3795,7 @@ def local_div_switch_sink(node):
 
                 fct = [T.switch(switch.inputs[0],
                                 fdiv, 0)]
-                fct[0].values_eq_approx = values_eq_approx_remove_nan
+                fct[0].tag.values_eq_approx = values_eq_approx_remove_nan
 
                 # Copy over stacktrace for switch op from both previous
                 # elementwise division op and previous switch op,
@@ -5599,7 +5599,7 @@ def local_log_add(node):
 
                 ret = max_pre + T.log1p(T.exp(T.add(*[p - max_pre
                                                       for p in pre_exp])))
-                ret.values_eq_approx = values_eq_approx_remove_inf
+                ret.tag.values_eq_approx = values_eq_approx_remove_inf
                 return [ret]
 
 
@@ -6023,7 +6023,7 @@ def local_log_erfc(node):
         threshold = 26.641747557
 
     ret = T.switch(x < threshold, node.outputs[0], stab_value)
-    ret.values_eq_approx = values_eq_approx_remove_inf
+    ret.tag.values_eq_approx = values_eq_approx_remove_inf
     return [ret]
 
 
@@ -6175,7 +6175,7 @@ def local_grad_log_erfc_neg(node):
     ret = T.switch(x < threshold, true_div_no_mul, stab_value)
     if y:
         ret = T.mul(ret, *y)
-    ret.values_eq_approx = values_eq_approx_remove_inf_nan
+    ret.tag.values_eq_approx = values_eq_approx_remove_inf_nan
     return [ret]
     """
 The libm used for the test is amdlibm