Requirements

When you have a model saved via MLEM and want to use it, the first step is to make sure you have the right dependencies and packages. Given a model, MLEM can get the list of requirements of different types (including Python, Unix, as well as some others).

This complements the checkenv command.

Pip based requirements

MLEM can export the installable requirements needed for a model using the mlem build command, where model is the path to model saved via mlem

$ mlem build requirements -m model
⏳️ Loading model from sk-model.mlem
scikit-learn==1.0.2 pandas==1.4.2 numpy==1.22.3

Now, it can be used to install the requirements like this:

$ pip install $(mlem -q build requirements -m model)

where -q stands for the quiet mode which disables the emoji output.

One can also save these requirements to generate a requirements.txt file using the --target option which allows us to pass a path of the file i.e.

$ mlem build requirements -m model --target requirements.txt
⏳️ Loading model from model.mlem
💼 Materializing requirements...
✅  Materialized to requirements.txt!

and now the contents of requirements.txt can be checked using

$ cat requirements.txt
scikit-learn==1.0.2
pandas==1.4.2
numpy==1.22.3

This is different from creating a python package for the model.

Unix based requirements

Some python libraries require unix based packages underneath to function correctly. An example is the libgomp1 package required by the lightgbm library. MLEM can figure out unix based packages for some supported libraries and these can be used as below:

To get a list of unix based requirements, use the --req_type option such as

mlem build requirements -m model --req_type unix.

The output of above can be used in conjunction like:

$ apt-get install $(mlem -q build requirements -m model --req_type unix)

The --target option is not supported for unix based requirements.

Custom requirements

Custom requirements represent local python code such as files, zipped sources, etc. Custom requirements always need the --target option since they are materialized at the target.

For instance, a function can be saved via MLEM at a location mlem-f

# func.py
def f(txt):
    print(txt)

# save.py
from mlem.api import save
from func import f

saved = save(f, 'mlem-f')

and the following command could be used to materialize this custom requirement at a target:

$ mlem build requirements -m mlem-f --req_type custom --target ./dir
⏳️ Loading model from mlem-f.mlem
💼 Materializing requirements...
✅  Materialized to ./dir!

The contents of dir can be checked using

$ ls dir
func.py

and with

$ cat dir/func.py
def f(txt):
    print(txt)

Conda requirements

Conda based requirements specify conda packages and cannot be determined automatically as of now. In any case, the --target option is not supported for it.

One can manually pass conda requirements to create conda based virtual environments using the conda builder as discussed here.