Fergus Henderson

I'm on the "TF Lite in Play services" team at Google, and I just came across this issue. (For faster support on similar issues in future, file an issue in...

While I am firmly in favour of addressing this issue, I realize that the changes to take proto field numbers into account may actually break binary backwards compatibility for existing...

Hi @ThomasRichtsfeld , Firstly, thank you for reporting this issue! Secondly, if you're having trouble creating a PR, feel free to just attach a patch file here.

I have made a fix for this.

The PR is this one: https://github.com/tensorflow/tflite-support/commit/e4bc12500c1ae3110c11345c8d93599515013b55 It was merged on November 21st.

Just a brief comment on this issue: TF Lite has not promised binary backwards compatibility. In particular for TFLiteRegistration, the structure was only documented as being "semi-ABI stable". See the...

Please add some release notes in `RELEASE.md` to the `tf.lite` section to give open source contributors details about this change when it gets included into a new release. This is...

It looks like Alan Kelly already wrote an update for the RELEASE.md: * Add int16x8 support for the built-in ops `space_to_batch_nd` and `batch_to_space_nd` However, it looks like some tests are...

I was able to reproduce this on Linux. Building with `-c dbg` results in SIGILL rather than SIGSEGV. ``` Program received signal SIGILL, Illegal instruction. 0x0000555555b725ff in flatbuffers::ts::TsGenerator::AddImport (this=0x7fffffffac90, imports=...,...