Sanggyu Lee

Results 17 issues of Sanggyu Lee

It adds two tflite models of nnpkg under nnpackage/examples. MANIFEST has pkg-inputs, pkg-outputs and model-connect fields. ONE-DCO-1.0-Signed-off-by: Sanggyu Lee Related: #9116 - Please don't merge this PR. I would like...

PR/NO TEST
PR/NO MERGE

I would like to add tests for multiple model nnpkg. Currently nnfw_api tests are located in `tests/nnfw_api`, and it can only handle single model nnpkg. We need to define how...

area/onert
type/discussion

I can see **~_26_~ 13** occurrences of `#if defined(ONERT_HAVE_HDF5) && ONERT_HAVE_HDF5 == 1` I think it would be better to eliminate them by introducing `H5FormatterDummy`. I found it become ugly...

type/suggestion

I see several if-conditions around our code for pipelined execution. - It is not documented much - seems not be used - make it hard to introduce new feature and...

area/onert
type/discussion
type/suggestion

It uses quantize type internally. ONE-DCO-1.0-Signed-off-by: Sanggyu Lee It includes #11494. This PR may be landed as whole. Then, I will close #11494.

I rewrite the internal issue for on-device quantizer (hereafter `ODQ`) that I wrote last year. Current implementation of on-device quantizer uses `hdf5`. It is good for prototyping and offline tools....

As basic training is working, I would like to clean up and reduce code changes which was introduced for training in tight schedule. - [ ] #12111 - [ ]...

area/onert
type/discussion

It removes `ir::train::operation` by adding clone() in IOperation. It removes op to `train::op` convertor and visitor. ONE-DCO-1.0-Signed-off-by: Sanggyu Lee Related to #12110, https://github.com/Samsung/ONE/pull/12073#issuecomment-1823807053 ``` $ Product/x86_64-linux.debug/out/bin/onert_train \ mnist.circle \ --load_input:raw...

PR/NO MERGE
type/discussion

I've found `Permute`s are eliminated during running `mnist.circle` with `onert_run`. ``` $ BACKENDS=cpu GRAPH_DOT_DUMP=1 ONERT_LOG_ENABLE=1 \ Product/x86_64-linux.debug/out/bin/onert_run \ mnist.circle ``` you can download `mnist.circle` in https://github.com/Samsung/ONE/pull/12031#issue-1991835752. ``` [ PassRunner ]...

future

It will use "wo_" to distinguish weight-only and full quantization. ONE-DCO-1.0-Signed-off-by: Sanggyu Lee #11497 must be landed before this PR.