Is pretrained model available?

[hqng]

I had read the paper. It is very interesting idea and a step forward from TabPFN.
Now I am looking at the MotherNet API and would like to try it on my internal data. However, it is quite frustrated since I cannot found any pretrained model. Am I missing something? or can you open-source a pretrained model? Thank you so much.

[amueller]

Hey! sorry for the slow response. There's an example here:
https://github.com/microsoft/mothernet/blob/762f63b821ed864eced123e09f7b9d26dc18b0b2/mothernet/tests/test_follow_sklearn_interface.py#L47

Using

from mothernet.prediction import MotherNetClassifier
from mothernet.utils import get_mn_model
model_string = "mn_d2048_H4096_L2_W32_P512_1_gpu_warm_08_25_2023_21_46_25_epoch_3940_no_optimizer.pickle"
    model_path = get_mn_model(model_string)
classifier = MotherNetClassifier(path=model_path)

yields the model from the arXiv paper.

Note that this is still a WIP and see the paper for some of the failure cases. In particular results with categorical variables depend a lot on whether OneHotEncoding is used and useful for a particular dataset.
Note that MotherNetClassifier doesn't do any ensembling and you'd need to use EnsembleMeta to get the ensemble as in the paper.

[hqng]

Thank you for your information. I'll try it later.

[Stanwang218]

Hi, it seemed that the website for the pretrained weight is down! (https://amuellermothernet.blob.core.windows.net/models/tabpfn_nooptimizer_emsize_512_nlayers_12_steps_2048_bs_32ada_lr_0.0001_1_gpu_07_24_2023_01_43_33_epoch_1650.cpkt)