staging: Add tensorflow micro module.#9938
Conversation
|
@lyakh see the Readme for the extra steps needed to link archives to the llext module. We are linking a C and C++17 library to the llext module. |
|
@lgirdwood please try this version of your second patch. |
7d6b224 to
9283180
Compare
Thanks, this seems to have done the trick, I've had to add |
486217e to
8e7d05f
Compare
8e7d05f to
091c6ca
Compare
10042e9 to
c3cfacb
Compare
|
Ready for staging. |
jsarha
left a comment
There was a problem hiding this comment.
Tried to get this to compile, but it failed. The make file tried to look for sources under ${BASE_DIR}/tflm-micro , but the code is cloned under ${BASE_DIR}/tflite-micro.
marcinszkudlinski
left a comment
There was a problem hiding this comment.
Why special scripts/tensorflow-clone.sh script? We do use west, just add the repos there
c3cfacb to
afa8164
Compare
Thanks fixed, the built-in config was using the wrong directory but llext was building fine. |
That was my initial intention too, but I could only get this working with a west manifest in each child module repo (and this would be difficult/slow to upstream). I think the script is fine to unblock us today as either someone with better west know how or updates to west could fix this in the future. |
This is staging quality and adds a TFLM module based on micro_speech example to do audio classification. Signed-off-by: Liam Girdwood <liam.r.girdwood@linux.intel.com>
kv2019i
left a comment
There was a problem hiding this comment.
Some minor comments, but nothing worth blocking the PR.
| ``` | ||
| CONFIG_CPP=y | ||
| CONFIG_STD_CPP17=y | ||
| CONFIG_SOF_STAGING=y |
There was a problem hiding this comment.
This could be an overlay alike sof/app/shell_overlay.conf or sof/app/perf_overlay.conf.
| g_arena, kArenaSize); | ||
|
|
||
| // and allocate the tensors | ||
| if (interpreter->AllocateTensors() != kTfLiteOk) { |
There was a problem hiding this comment.
This is more of a tflite API question, but wondering can "interpreter" be NULL/nullptr here? Maybe this is ok and the new operator is just overloaded.
Integrates Tensorflow micro edition using xtensa optimized kernels into SOF as a module. WIP.