How to compute and store the Hessian for last layer LA offline and use it for future predictions about laplace

[runame]

Hi,

Unfortunately the library currently has no native support for saving and loading Laplace instances. However, you should be able to simply use pickle to store and load the whole class instance, i.e. by using something like

import pickle

def save_laplace(la, filepath):
    with open(filepath, 'wb') as output:
        pickle.dump(la, output)

def load_laplace(filepath):
    with open(filepath, 'rb') as input:
        la = pickle.load(input)
    return la

where la is your instance of last-layer Laplace after fitting and filepath is the path to the file in which you want to store the object. Just let me know if you run into issues with this approach.

We are planning to make this process more convenient in a future release, I have just opened an issue for this (#45).

[roysubhankar]

Thank you for this workaround.

[roysubhankar]

Hi @runame , I am getting this error when I am using your suggestion.

AttributeError: Can't pickle local object 'FeatureExtractor._get_hook..hook'

[runame]

Ah I didn't test this approach with last-layer Laplace which uses the FeatureExtractor with nested functions and pickle can't handle this. Can you try using dill instead of pickle? Simply install it via pip install dill.

import dill

def save_laplace(la, filepath):
    with open(filepath, 'wb') as outpt:
        dill.dump(la, outpt)

def load_laplace(filepath):
    with open(filepath, 'rb') as inpt:
        la = dill.load(inpt)
    return la

[runame]

@roysubhankar Did it work?

[panos108]

I had the same issue. It seems that the dill solved thanks!

[runame]

Perfect, thanks for letting me know!