[microTVM] Add tutorial on how to generate MLPerfTiny submissions #13783
Conversation
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.
Generated by tvm-bot |
mehrdadh
force-pushed
the
micro/mlperftiny_tutorial
branch
from
b344fed to
2bcb15e
Compare
| # If you like to build the submission with CMSIS-NN, modify USE_CMSIS variable. | ||
| # | ||
|
|
||
| MODEL_URL = "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite" |
There was a problem hiding this comment.
should we include other models, since we reference them elsewhere?
There was a problem hiding this comment.
I think it would redundant here. We can make a micro/test api call that returns the address to the model.
There was a problem hiding this comment.
I'm with @mehrdadh - I worry having multiple URLS might be confusing. Adding a micro API call would be a good way to clean this up in a future PR.
There was a problem hiding this comment.
Can this line (101) be pleased changed as follows?
MODEL_URL = os.environ.get("MODEL_URL", "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite")
This can give flexibility to the user to change the url without having to modify the code.
There was a problem hiding this comment.
@arjunsuresh Unfortunately that's not the only change that is required to use a different model in this tutorial. We could change this tutorial in a follow up PR to get all the model URL, name and index from test APIs in TVM
| f"-DWORKSPACE_SIZE={workspace_size + 512}", | ||
| f"-DTARGET_MODEL={MODEL_INDEX}", | ||
| f"-DTH_MODEL_VERSION=EE_MODEL_VERSION_{MODEL_SHORT_NAME}01", | ||
| f"-DMAX_DB_INPUT_SIZE={input_total_size}", |
There was a problem hiding this comment.
Would love for you to briefly state what these do.
| if BOARD == "nrf5340dk_nrf5340_cpuapp": | ||
| config_main_stack_size = 4000 | ||
| elif BOARD == "nucleo_l4r5zi": | ||
| config_main_stack_size = 4000 |
There was a problem hiding this comment.
Can we just set config_main_stack_size=4000?
There was a problem hiding this comment.
done, added a note that they might need to adjust it based on their board
| # If you like to build the submission with CMSIS-NN, modify USE_CMSIS variable. | ||
| # | ||
|
|
||
| MODEL_URL = "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite" |
There was a problem hiding this comment.
I'm with @mehrdadh - I worry having multiple URLS might be confusing. Adding a micro API call would be a good way to clean this up in a future PR.
gromero
changed the title
mehrdadh
force-pushed
the
micro/mlperftiny_tutorial
branch
from
000b551 to
faf3b56
Compare
…ache#13783) This PR adds a tutorial on how to generate an MLPerftiny submission on Zephyr OS using microTVM.
This PR adds a tutorial on how to generate an MLPerftiny submission on Zephyr OS using microTVM.
cc @alanmacd @gromero @mkatanbaf