[joss-review] Error while trying to reproduce results from paper

[swouf]

openjournals/joss-reviews#9093 (comment)
Hi, while trying to reproduce the results from the paper, I encountered an error and I don't find my way around it. I applied the steps for the runnable example, but I can't manage to get around this. I have the following error.

Could someone help me with this? That would be great.

vscode ➜ /workspaces/python/emlearn-micropython/paper (master) $ python create_plot.py
missing values 0.0033332962967078145 %
window duration 256 2.56 s
run-cmd micropython /workspaces/python/emlearn-micropython/paper/example.py /tmp/tmppsdbb0u8/input.npy /tmp/tmppsdbb0u8/output.npy
Traceback (most recent call last):
  File "/workspaces/python/emlearn-micropython/paper/example.py", line 13, in <module>
ValueError: native code in .mpy unsupported
Traceback (most recent call last):
  File "/workspaces/python/emlearn-micropython/paper/create_plot.py", line 119, in <module>
    main()
    ~~~~^^
  File "/workspaces/python/emlearn-micropython/paper/create_plot.py", line 106, in main
    combined, smooth, estimator = train_and_run(sub, class_columns)
                                  ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^
  File "/workspaces/python/emlearn-micropython/paper/create_plot.py", line 49, in train_and_run
    features = process_data(sensor_data)
  File "/workspaces/python/emlearn-micropython/paper/processing.py", line 80, in process_data
    subprocess.check_output(args)
    ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^
  File "/usr/local/lib/python3.13/subprocess.py", line 472, in check_output
    return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
           ~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
               **kwargs).stdout
               ^^^^^^^^^
  File "/usr/local/lib/python3.13/subprocess.py", line 577, in run
    raise CalledProcessError(retcode, process.args,
                             output=stdout, stderr=stderr)
subprocess.CalledProcessError: Command '['micropython', '/workspaces/python/emlearn-micropython/paper/example.py', '/tmp/tmppsdbb0u8/input.npy', '/tmp/tmppsdbb0u8/output.npy']' returned non-zero exit status 1.

[jonnor]

Hi @swouf - as far as I understand, you are on Mac OS with an arm64 CPU, right?

If so, then you are affected by dynamic loadable MicroPython modules (.mpy) files not being supported there. Ref micropython/micropython#5500 - and the ValueError: native code in .mpy unsupported indicates it as well.
So I need to create some other instructions for building the modules into a custom micropython binary, or maybe a pre-built binary for Mac OS via Gitlab CI.

[swouf]

Hi, yes, I'm on Mac OS with an arm64. Yes, it looks like I'm running into the issue you mentionned. If I understand it correctly, it's a particular case. Maybe I missed something from your docs that warns about that, or, if not, maybe you could put a disclaimer until dynamic modules are supported in this context. I see you are already well aware of this issue. I'll try to reproduce your results on a linux machine in the meantime to finalize the review. Thanks!

[jonnor]

I have added the other method of running it now, to paper/README: https://github.com/emlearn/emlearn-micropython/blob/master/paper/README.md#running-using-custom-micropython-executable
We already have a prebuilt executable for arm64, so it should hopefully be easy to download and run that (per the instructions).

[swouf]

Hi @jonnor, thanks for adding these instructions. Unfortunately, this doesn't work for me still. I have downloaded the micropython binary with the embedded emlearn modules according to your instructions. I also pulled the latest branch as I saw you updated the Makefile. But now, I run into this issue when python create_plot.py:

missing values 0.0033332962967078145 %
window duration 256 2.56 s
run-cmd micropython /Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/example.py /var/folders/7y/0tz8nnv17hqbfrpz_s5xn7rw0000gn/T/tmpku6v7lxx/input.npy /var/folders/7y/0tz8nnv17hqbfrpz_s5xn7rw0000gn/T/tmpku6v7lxx/output.npy
Traceback (most recent call last):
  File "/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/example.py", line 132, in <module>
  File "/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/example.py", line 91, in process_file
AttributeError: module 'npyfile' has no attribute 'Reader'
Traceback (most recent call last):
  File "/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/create_plot.py", line 119, in <module>
    main()
  File "/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/create_plot.py", line 106, in main
    combined, smooth, estimator = train_and_run(sub, class_columns)
                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/create_plot.py", line 49, in train_and_run
    features = process_data(sensor_data)
               ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/processing.py", line 80, in process_data
    subprocess.check_output(args)
  File "/opt/homebrew/anaconda3/lib/python3.11/subprocess.py", line 466, in check_output
    return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/anaconda3/lib/python3.11/subprocess.py", line 571, in run
    raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['micropython', '/Users/jlascala/EPFL/PhD/3️⃣ - 🗺️ Areas/Reviewing/emlearn-micropython/emlearn-micropython/paper/example.py', '/var/folders/7y/0tz8nnv17hqbfrpz_s5xn7rw0000gn/T/tmpku6v7lxx/input.npy', '/var/folders/7y/0tz8nnv17hqbfrpz_s5xn7rw0000gn/T/tmpku6v7lxx/output.npy']' returned non-zero exit status 1.
(base) 

Apparently, this there is something with the npyfile module:

AttributeError: module 'npyfile' has no attribute 'Reader'

Is it possible that something is missing in micropython binary?

[Puneetha-Ramachandra]

I vaguely remember encountering such errors when I used micropython prebuilts. Have you tried building micropython from source - https://emlearn-micropython.readthedocs.io/en/latest/getting_started_host.html as mentioned in #50 (comment)?

[jonnor]

Hi @swouf and thanks for testing again. The error refers to npyfile.py, which should have been downloaded by the curl command. But it turns out curl does not follow redirects by default :o - so it ended up with an empty file and no error... I have updated the command now. So hopefully create_plot will work after you re-run that

[swouf]

Hi @jonnor, thank you for your patience and your help. It finally worked ! I could reproduce the plot on MacOS, and with that, I'll close my review.

Before I close this issue, there is a tiny mistake in your documentation. In the command to install the packages, I think you forgot plotly. I had to install it to make the plot.

---

Aside from that, maybe you would be interested in packaging your script with an env management tool such as uv. I find very convenient to manage dependencies to scripts and whole projects. I used it to run your script, first by applying the dependencies with

uv add --script create_plot.py "numpy" "pandas" "scikit-learn" "emlearn" "setuptools" "pyarrow" "kaleido" "plotly"

which adds a section with dependencies at the top of the file. This can be read by uv to manage the deps automatically. Then, you can run it simply with

uv run create_plot.py

which is the only command your users would need to run the script (uv fetches the deps and run the script automatically).

This comment is just an idea. Take what you want from it.

[jonnor]

That is great to hear that it worked now @swouf - thank you for the quick iterations on testing and reviewing.

Good point about plotly, that might be missing. Will check and fix that.

Regarding uv - that would not handle the MicroPython part, right? Which is a key dependency here. In the future I want to see MicroPython packaged as a PyPi package - but that work should happen upstream. And that we would have support for .mpy modules on Mac OS, that way not needing a custom micropython build.

I see the benefit with uv and I will consider it in the future. I would need to get a bit familiar with the tool myself first - as anything the documentation recommends, I (as a maintainer) need at the very least to be able to debug basics of.