tf.raw_ops.Unbatch aborts with "Check failed: d < dims()"

[Assoap]

tf.raw_ops.Unbatch aborts with "Check failed: d < dims()"

[sushreebarsa]

@Assoap Generally this issue occurs when a program encounters a critical error and crashes. The core memory, which holds essential data for the program's execution, is dumped to a file for potential debugging. TensorFlow might not have explicit error handling for every possible scenario within tf.raw_ops.Unbatch. If the internal checks fail due to unexpected input, the program might not have a proper way to recover and raise a user-friendly exception.

Thank you!

[mihaimaruseac]

Would be recommended to fix/eliminate the check fails, but they should be treated as bugs, not security issues.

[Assoap]

gist

[sushreebarsa]

@Assoap tf.raw_ops.Unbatch is designed to split a batched tensor along the first dimension (batch dimension) into separate elements. However, a scalar tensor has no dimension to split along. This mismatch in dimensions (d < dims()) leads to the error. Could you please remove the tf.raw_ops.Unbatch call altogether. The scalar tensor already represents a single value. Please let us know if it helps?
Thank you!

[Assoap]

@sushreebarsa Thank you so much for your prompt reply. I was just wondering, why didn't the program raise any exceptions, such as ValueError, but instead crashed (core dumped) directly?

[svkgn4DL]

@Assoap Generally this issue occurs when a program encounters a critical error and crashes. The core memory, which holds essential data for the program's execution, is dumped to a file for potential debugging. TensorFlow might not have explicit error handling for every possible scenario within tf.raw_ops.Unbatch. If the internal checks fail due to unexpected input, the program might not have a proper way to recover and raise a user-friendly exception.

Thank you!

That means users don't need to report check fails and they will stay forever then? In that case it will be better to highlight the same in README.md so that users can save their time in reporting such unnecessary (?) issues no ?

CC: @mihaimaruseac