Pickling objects

I wanted to pickle the LogisticCalibration() class after I had fit it (for later re-use), but I was getting an error related to can't pickle _thread.Rlock objects.

I was able to find a work-around by setting logger in the class no None. Seems a bit hacky, but it did work. Might be something to think about in future releases

[fabiankueppers]

Hi @VCBE123 thanks for pointing out. Indeed, this is a point I’ve missed to document in the README. You can simply store and reload a model by

lr = LogisticCalibration()

# store on disk
lr.save_model("foo/bar/model.pkl")

# load from disk
lr.load_model("foo/bar/model.pkl")

Hope this helps!

[VCBE123]

Hi @fabiankueppers thank you for your clean demo. I have another question. Since I can save the trained model to a .pkl file, is there any advice for reconstructing the model in C/C++ for development usage?

Finally, I dive into the code and reconstruct the model.

from netcal.scaling import LogisticCalibration
import torch
import numpy as np

lr = LogisticCalibration(detection=True, use_cuda=False)
lr.load_model("lr_5.pkl")

inputs = np.random.randn(10, 5)
out = lr.transform(inputs)


model_weight = torch.load("lr_5.pkl")["_sites"]
inputs = torch.from_numpy(inputs)
weight = torch.from_numpy(model_weight["weights"]['values']).reshape(-1, 1)
bias = torch.from_numpy(model_weight["bias"]['values']).reshape(-1, 1)


def inferse_sigmoid(confidence):
    epsilon = float(torch.finfo(confidence.dtype).eps)
    clipped = torch.clamp(confidence, epsilon, 1. - epsilon)
    inv_clipped = torch.clamp(1. - confidence, epsilon, 1. - epsilon)

    logit = torch.log(clipped) - torch.log(inv_clipped)
    return logit


inputs[:, 0] = inferse_sigmoid(inputs[:, 0])
out_r = torch.sigmoid(torch.mm(inputs, weight) + bias).squeeze_(1).numpy()

assert np.allclose(out, out_r)