theislab/batchglm

Model initialisation with reference model

davidsebfischer opened this issue · 4 comments

davidsebfischer commented

Why is this

batchglm/batchglm/train/tf/nb_glm/estimator.py

Line 934 in 154e871

init_loc = np.random.uniform(
a random initialisation within a very small interval rather than a constant? I do not see the benefit of introducing this stochasticity but it can break reproducibility. As this additive in log space, I would just set the parameters that did not occur in the full model to zero?

davidsebfischer commented

I can modify it if we decide to change this.

Hoeze commented

I always used this to make sure that not all values in a tensor are equal, I just copied this over.
This can also be set to 0 I think...

davidsebfischer commented

Ok, I will try it and merge if it works.