justinalsing / pydelfi

Density Estimation Likelihood-Free Inference with neural density estimators and adaptive acquisition of simulations

pydelfi

Density Estimation Likelihood-Free Inference with neural density estimators and adaptive acquisition of simulations. The implemented methods are described in detail in Alsing, Charnock, Feeney and Wandelt 2019, and are based closely on Papamakarios, Sterratt and Murray 2018, Lueckmann et al 2018 and Alsing, Wandelt and Feeney, 2018. Please cite these papers if you use this code!

Installation:

The code is in python3. There is a Tensorflow 1 (most stable, see below) and Tensorflow 2 version that can be installed as follows:

Tensorflow 1 (stable)

This can be found on the master branch and has the following dependencies:
tensorflow (<=1.15)
getdist
emcee (>=3.0.2)
tqdm
mpi4py (if MPI is required)

You can install the requirements and this package with,

pip install tensorflow==1.15
pip install git+https://github.com/justinalsing/pydelfi.git

(tensorflow-gpu==1.15 for GPU acceleration instead of tensorflow==1.15)

or alternatively, pip install the requirements and then clone the repo and run python setup.py install

Tensorflow 2

The Tensorflow 2 version can be found on the tf2-tom branch and can be installed as follows. We reccommend you do the install inside a virtual environment to keep version conflicts under control, ie.,

mkdir ~/envs
virtualenv ~/envs/pydelfi
source ~/envs/pydelfi/bin/activate

Followed by a pip install of pydelfi:

pip install git+https://github.com/justinalsing/pydelfi.git@tf2-tom

Note: the Mixture Density Networks (MDN) in the tf2 version are currently not performing as well as in the tf1 version (but the Masked Autoregressive Flows are fine). We are getting ot the bottom of this, and also working on expanding the suite of conditional density estimators in a coming update. Watch this space.

Documentation and tutorials:

Once everything is installed, try out either cosmic_shear.ipynb or jla_sne.ipynb as example templates for how to use the code; plugging in your own simulator and letting pydelfi do it's thing.

If you have a set of pre-run simulations you'd like to throw in rather than allowing pydelfi to run simulations on-the-fly, look at cosmic_shear_prerun_sims.ipynb as a template for how to do this.

If you are interested in using pydelfi with nuisance hardened data compression to project out nuisances (Alsing & Wandelt 2019), take a look at jla_sne_marginalized.ipynb.

Documentation can be found here (work in progress).

If you are interested in applying pydelfi to your problem but need some help getting started, or have an application that requires adaptations of the code, don't hesitate to get in touch with us (at [email protected]) or open an issue - we welcome collaboration!

pydelfi's Issues

fisher bug

In "score/score.py", line 121 should be:

if F is None: self.compute_fisher()

At the moment it is set to none, which, if I remember, was a placeholder once?

use os.path.join to join path and filename

pydelfi/pydelfi/delfi.py

Lines 160 to 161 in 5c4e7ed

	self.graph_restore_filename = results_dir + graph_restore_filename
	self.restore_filename = results_dir + restore_filename

It would be better to use https://docs.python.org/2/library/os.path.html#os.path.join and thus not assume that results_dir has a trailing separator. At present, without a trailing separtor, e.g., results_dir = 'here', the results_dir just becomes a prefix for filenames, not a directory.

not able to install the pydelfi for tensorflow 2

Hi Justin,

I was trying to install the pydelfi compatible with tensorflow 2, following your instructions by running "pip install git+https://github.com/justinalsing/pydelfi.git@tf2-tom". When I import pydelfi.priors, it complains that there's no such module.
I would appreciate it if you could assist.

Thanks,
Sam

Setup file

The current setup.py file has its Tensorflow requirement as
"tensorflow>=v1.1.0",
meaning that when I pip install the package from source in a fresh conda environment it installs the newest TensorFlow version (2.1).

I think what you want here is
"tensorflow>=1.1.0, <=1.15".

Additionally, after having done this, my install threw an error relating to mpi4py. This was fixed by me installing this package separately (not sure whether this is a problem with the setup or just an issue with my environment).

issue for n_procs>1

Hi!

I was running inferences for simplistic linear regression models. This may be my ignorance in understanding the code, but even for simplistic models like

y=mx+c, the NDE takes a lot of time to train. I do not enable any kind of fisher training for now.

additionally, if i want to use the power of n_procs>1:
it shows me an error

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_1564073/1911766202.py in <module>
      1 # Do the SNL training
----> 2 DelfiEnsemble.sequential_training(simulator, compressor, n_initial, n_batch, n_populations, patience=10, save_intermediate_posteriors=True)

~/miniconda3/envs/sbi/lib/python3.7/site-packages/pydelfi/delfi.py in sequential_training(self, simulator, compressor, n_initial, n_batch, n_populations, proposal, simulator_args, compressor_args, safety, plot, batch_size, validation_split, epochs, patience, seed_generator, save_intermediate_posteriors, sub_batch)
    390             ps = np.zeros((safety * n_initial, self.npar))
    391         if self.use_mpi:
--> 392             self.comm.Bcast(ps, root=0)
    393         self.inds_prop = self.allocate_jobs(safety * n_initial)
    394         self.inds_acpt = self.allocate_jobs(n_initial)

AttributeError: 'NoneType' object has no attribute 'Bcast'

I checked my mpi installations are fine, and even works for other codes.

bug with run_mcmc returning State not 3 tuple

I get the "ValueError: too many values to unpack (expected 3)" when I try to use the emcee_sample method at the following line:

pydelfi/pydelfi/delfi.py

Line 348 in 5c4e7ed

pos, prob, state = sampler.run_mcmc(x0, burn_in_chain)

Which I think means it is returning blobs as well from the emcee documentation (https://emcee.readthedocs.io/en/latest/user/sampler/#emcee.State).

With the newest version of emcee it works to instead return the state:
state = sampler.run_mcmc(x0, burn_in_chain)
and then use
sampler.run_mcmc(state.coords, main_chain)
instead of:

pydelfi/pydelfi/delfi.py

Line 352 in 5c4e7ed

sampler.run_mcmc(pos, main_chain)

Reorganization of code

Can we put all python source files into a delfi folder at the root of the project? So it would be, e.g., /delfi/ndes/... , /delfi/delfi/..., etc.

Then, instead of running:

import ndes

one would run

from delfi import ndes

This is in line with other python projects.

Btw, you can now run pip install git+https://github.com/justinalsing/pydelfi.git to install this project and the requirements.

Bugs in examples (after new edits)

The latest "jla_sne.ipynb" example doesn't work.
The line with DelfiEnsemble.fisher_pretraining() now gives an error:

Traceback (most recent call last): File ".../python3.6/site-packages/emcee/ensemble.py", line 519, in __call__ return self.f(x, *self.args, **self.kwargs) TypeError: log_posterior_stacked() missing 1 required positional argument: 'data'

Also, the initial module loads are now wrong (this isn't a huge issue), they should be:

import pydelfi.ndes as ndes
import pydelfi.delfi as delfi
import pydelfi.score as score
import pydelfi.priors as priors