Albert Zeyer

@critias @pavelgolik @manish-kumar-garg maybe you can try with TF 2?

Note: I was able to reproduce the problem again, with a slightly adopted test case ([current version](https://github.com/albertz/playground/blob/21d081ef01a4b8a97b99ce12c99d7e72e4d5685b/tf-test-tensorarray-bug.py)). This test case now removes the `check_numerics` calls (which get false positives for...

For reference, with LSTMBlockCell, I also got this inf detection once (but not sure whether it really originates from there): ``` (1) Invalid argument: gradients/s/LSTMBlockCell_1_grad/LSTMBlockCellGrad:1 : Tensor had Inf values...

@valerie94 Can you try TF 1.14 and TF 2.2 and report whether you see the same problem?

Thanks, good to know. So we can just recommend to use TF 1.14 and >=TF 2.2.

`derived_from_tag`/`derived_matches` is not the right option here. `derived_from_tag` means that it is derived from this tag (via some op, like +- or so). We use it as a heuristic to...

Note that I stumbled upon this (or related) specifically for the `DotLayer` in #1154, and fixed via #1155. In #1155, I introduced a new behavior version to trigger the more...

I'm now thinking whether my proposed solution from from #666 is maybe really the better way which would have avoided these issues: > As another solution, maybe we can disallow...

We have the `auto_generated` flag.

This change should also directly enforce not to use the tuple-style declaration (or maybe that's already the case of `extern_data`, I forgot).