InvalidArgument error with tf.scan()

[danakianfar]

System information

• No custom code
• Error encountered on Windows 10 and CentOS 6.7
• binary
• TF 1.1.0 and 1.2.1 respectively
• 3.6
• CUDA 8.0, CUDNN 5.1
• GeForce 940MX 2GB (ran out of memory in execution), Titan X 12 GB

Describe the problem

I encountered this error when using tf.scan on an MultiRNNCell.

InvalidArgumentError (see above for traceback): Max scatter index must be <= array size (141 vs. 141)

Source code / logs

If you look here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/tensor_array_ops.cc#L1072

OP_REQUIRES(
          ctx, max_index < array_size,
          errors::InvalidArgument("Max scatter index must be <= array size (",
                                  max_index, " vs. ", array_size, ")"));
    }

It seems that the condition (<) is not consistent with the error message (<=).

[danakianfar]

I think that this is just a matter of fixing the error message. The error condition seems just to be checking for correct indexing.

[AswinBlue]

The code's hyperlink address was moved to here
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/tensor_array_ops.cc#L1072

can you upload the code that makes the error?

[yongtang]

Yes I think the error message should be fixed as index is 0-based. PR #12800.