Check failed: NDIMS == dims() (2 vs. 1)

[LucasMahieu]

I get the following error after building the graph of my NN. when I try to execute :

loss, _ = self._sess.run([self.loss_op, self.train_op], feed_dict=feed_dict)

I get this error :

F tensorflow/core/framework/tensor_shape.cc:36] Check failed: NDIMS == dims() (2 vs. 1) Asking for tensor of 2 dimensions from a tensor of 1 dimensions

I don't succeed to know to which part of the graph this error is related to ... Because, normaly the graph is build, so shapes should be correct ... are not ?

I am using Tensorflow version 1.0.0 with python API.

It could be a bug according to the first reply here : http://stackoverflow.com/questions/43413293/tensorflow-check-failed-ndims-dims-2-vs-1

[drpngx]

That seems like a bug. Could you share the simple possible repro you can find?

[LucasMahieu]

I succeed to find the source of the error.
It was caused by the bad shape of an input tensor of a custom op.
I have create a custom op, coded in C++, used in Python.
I think an improvement could be done in the error message. Indeed, with this line only, it is difficult to find the source of this error.