iterative.ai mlem.api.save() | AttributeError: __name__. Did you mean: '__ne__'?

[danielbellsa]

Goal: Save a model as a mlem model, without sample_data.

I plan on loading in mlem model files, in deployment, like so:

from mlem.api import load

model = load('models/mlem-model')

To do this, I need to load in existing models and then save() them as a mlem model.

save() creates a model binary and a mlem metadata file.

models/
    ├── mlem-model
    └── mlem-model.mlem

I've trained models, so don't pass sample_data as a paramter to save().

---

Code

from fastai.vision.all import *
import mlem

model = load_learner('models/model.pkl', cpu=True)
model.load('models/weights.pth')

mlem.api.save(model, 'models/mlem-model')  #, sample_data=df)

Traceback

(venv) me@laptop:~/BitBucket/project$ python3 tools/export_pickle_as_mlem.py 
Traceback (most recent call last):
  File "/home/me/BitBucket/project/tools/export_pickle_as_mlem.py", line 18, in <module>
    mlem.api.save(model, 'models/my-model-mlem')
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/mlem/telemetry.py", line 50, in inner
    return f(*args, **kwargs)
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/mlem/core/metadata.py", line 122, in save
    meta = get_object_metadata(
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/mlem/core/metadata.py", line 53, in get_object_metadata
    return MlemModel.from_obj(
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/mlem/core/objects.py", line 740, in from_obj
    model_type = model_hook.process(model)
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/mlem/contrib/callable.py", line 213, in process
    s = Signature.from_method(
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/mlem/core/model.py", line 236, in from_method
    name=override_name or method.__name__,
  File "/home/me/miniconda3/envs/venv/lib/python3.10/site-packages/fastcore/basics.py", line 497, in __getattr__
    raise AttributeError(k)
AttributeError: __name__. Did you mean: '__ne__'?

[aguschin]

Hi @danielbellsa! Thanks for reporting! I know @mike0sv was working on implementing fastai support, so I think he can help us figure this out!

[mike0sv]

FastAI is very hard to integrate with because it's authors are very "creative". You can try exporting to torch or maybe provide a callable function instead of learner (maybe something learner.predict or declare a custom function).
I will try to investigate it in the meantime

[mike0sv]

@ open code in question
@ see this

@ close the code in question

[aguschin]

Sorry @mike0sv - didn't get your point about the screenshot. Do you mean it's an example of a "creative" code? Do we plan to ever integrate with FastAI then? 🤔

[aguschin]

@danielbellsa - sorry - based on what Mike said, it looks like we can't just support that for now. Did you use some workaround or decided to skip using MLEM in this case?

Btw, about sample_data - to serve model with FastAPI or Streamlit (and deploy the model somewhere), you'll have to pass sample_data - so MLEM can figure out data schema to build an API for your model.

[mike0sv]

The point is that fastai is a lot of pain :). I will get back to this in #268