ONE
ONE copied to clipboard
Samsung
[onert] Support TFlite 2.8 for gpu_cl
- Update the tensorflow version used in gpu_cl to 2.8.
- I added config file to download source code to build tensorflow lite 2.8.
- Applied changed api with update to 2.8.
- Delete unnecessary code.
ONE-DCO-1.0-Signed-off-by: HyunJin Yi [email protected]
@seanshpark There's no reason to use 2.7 instead of 2.8. I updated to 2.7 because my existing working repo used 2.7. Upgrading to 2.8 requires some modifications.
There's no reason to use 2.7 instead of 2.8. I updated to 2.7 because my existing working repo used 2.7. Upgrading to 2.8 requires some modifications.
Can you please share background or purpose of this change? If there is an issue for this, please add a link.
There's no reason to use 2.7 instead of 2.8. I updated to 2.7 because my existing working repo used 2.7. Upgrading to 2.8 requires some modifications.
Can you please share background or purpose of this change? If there is an issue for this, please add a link.
https://github.sec.samsung.net/tizen-sys-con/ONE/pull/185 https://github.sec.samsung.net/tizen-sys-con/ONE/pull/197
The OP not implemented in tf2.4 implemented ex separately. However, it is implemented in tf2.7, so I updated it to tf2.7. I don't want to use ex anymore.
@nnfw-bot test cross-runtime-test-debug @nnfw-bot test cross-runtime-test-release
https://github.sec.samsung.net/tizen-sys-con/ONE/pull/185 https://github.sec.samsung.net/tizen-sys-con/ONE/pull/197
Thesw two are PRs.
I asked Can you please share background or purpose of this change? If there is an issue for this, please add a link.
I don't understand how I could get the background or purpose of this PR.
https://github.sec.samsung.net/tizen-sys-con/ONE/pull/185 https://github.sec.samsung.net/tizen-sys-con/ONE/pull/197
Thesw two are PRs.
I asked
Can you please share background or purpose of this change? If there is an issue for this, please add a link.I don't understand how I could get the background or purpose of this PR.
The OP not implemented in tf2.4 implemented ex separately.
However, it is implemented in tf2.7, so I updated it to tf2.7.
I don't want to use ex anymore.
Ex folder was removed in these PRs. So I linked these PRs.
We need Reduce Op, Split Op.... But Those OPs are not in TF2.4.1, but are in TF2.7.0. So I update TF2.4.1 to TF2.7.0.
Ex folder was removed in these PRs. So I linked these PRs.
We need Reduce Op, Split Op.... But Those OPs are not in TF2.4.1, but are in TF2.7.0. So I update TF2.4.1 to TF2.7.0.
Are these only for Tizen? I know you work on Tizen so I can understand the changes for Tizen.
I see changes for Android and also for Linux? But why these too?
There are too many changes that just So I update TF2.4.1 to TF2.7.0. explanation is not enough.
I don't review codes related for OneRT so actully don't care,
but there are changes related with common part in infra (so I requested to separate them) and I also sometimes test with OneRT as a user and options_armv7l-linux.cmake changes affect this.
Ex folder was removed in these PRs. So I linked these PRs. We need Reduce Op, Split Op.... But Those OPs are not in TF2.4.1, but are in TF2.7.0. So I update TF2.4.1 to TF2.7.0.
Are these only for Tizen? I know you work on Tizen so I can understand the changes for Tizen. I see changes for Android and also for Linux? But why these too? There are too many changes that just
So I update TF2.4.1 to TF2.7.0.explanation is not enough. I don't review codes related for OneRT so actully don't care, but there are changes related with common part ininfra(so I requested to separate them) and I also sometimes test with OneRT as a user andoptions_armv7l-linux.cmakechanges affect this.
I am currently developing gpu_cl. I am updating the tensorflow version used in gpu_cl to 2.7 now.
For this reason,
We need Reduce Op, Split Op.... But Those OPs are not in TF2.4.1, but are in TF2.7.0.
however, the gpu_cl test is running on Tizen, Andorid, and Linux. So I fixed them all.
Changing the Tensorflow version will change the api we use. So it is difficult to change only the tizen.
however, the gpu_cl test is running on Tizen, Andorid, and Linux. So I fixed them all.
Does this change work also in RPi4 Ubuntu 18.04/20.04?
Does this change work also in RPi4 Ubuntu 18.04/20.04?
I didn't check it directly, but I checked it through "cross-runtime-test-debug" and "cross-runtime-test-release".
@YiHyunJin , how about rebase this PR to version 2.8 of TensorFlow and send it back if it doesn't require a lot of effort?
ONE's TensorFlow support history is a bit complicated. It has been proceeding in the order of 2.2 -> 2.3 -> 2.4 -> 2.6 -> 2.8. Currently, version 2.8 is officially supported, but some modules are still stuck in the old version. For version 2.7, we never supported it. It doesn't seem appropriate to introduce new kernels to a version that has never been supported in the past.
Sooner or later, delayed upgrades of several modules will all have to match 2.8. In addition, it is necessary to upgrade again to version 2.9 afterward.
Considering these various situations, it is recommended to upgrade your work to support version 2.8 immediately.
@YiHyunJin , how about rebase this PR to version 2.8 of TensorFlow and send it back if it doesn't require a lot of effort?
ONE's TensorFlow support history is a bit complicated. It has been proceeding in the order of 2.2 -> 2.3 -> 2.4 -> 2.6 -> 2.8. Currently, version 2.8 is officially supported, but some modules are still stuck in the old version. For version 2.7, we never supported it. It doesn't seem appropriate to introduce new kernels to a version that has never been supported in the past.
Sooner or later, delayed upgrades of several modules will all have to match 2.8. In addition, it is necessary to upgrade again to version 2.9 afterward.
Considering these various situations, it is recommended to upgrade your work to support version 2.8 immediately.
Okay. I will rebase this PR to version 2.8. :)
I removed @glistening from reviewer and merged because there is no reaction for a long time.
