AttributeError: module 'pymc' has no attribute 'save_trace' when trying to save trace

Questions v5
1

Hi All,

I am trying to save a model with:
pm.save_trace(trace, "path", overwrite=True)
but get the error in the title.

Is this not the way to do it?

Many thanks for your help!

2

For storing the mcmc samples, you can interface through Arviz. See InferenceData — ArviZ dev documentation

2 Likes
3

Thanks! What is save_trace doing then that I’ve been reading and finding about in all the old threads?

4

Also: sorry for the thousand questions, is the best practice then to save to dataframe and then load and convert a dataframe again? Or what is the best way to save a trace and load it for use again later?

There isn’t much written here :sweat_smile: arviz.InferenceData.load — ArviZ dev documentation

5

I just throw everything into a cloudpickle:

import cloudpickle
model_fpath = '/Your/Path/Here.pkl'
with open(model_fpath, 'wb') as buff:
    cloudpickle.dump({'model': model, 'trace': idata,'y_obs':y_obs}, buff)

and then to read:

with open(model_fpath, 'rb') as buff:
      saved_model = cloudpickle.load(buff)
idata = saved_model['trace']
1 Like
6

Many thanks! It was my initial thought too, I read here in discussions and on github where everyone says not to do that but instead use save_trace. Thanks so much, I find that much easier and straightforward too.

7

Usually I do something like:

import arviz as az
model_fpath = '/Your/Path/Here.nc'
idata.to_netcdf(filepath)
# then to read:
idata = az.from_netcdf(filepath)