Skip to content

Latest commit

 

History

History
88 lines (54 loc) · 3.04 KB

README.md

File metadata and controls

88 lines (54 loc) · 3.04 KB

HklEnv

An environment for a reinforcement learning approach to faster crystallographic measurements using Actor-Critic (A2C) from OpenAI baselines

Installation

Build a new python environment with the usual numpy, matplotlib, scipy, h5py. Using an HPC systems with an anaconda module, this would be something like:

module load anaconda3
conda create -n hklgym numpy matplotlib scipy h5py

Install the appropriate tensorflow / tensorflow-gpu. On IBM power architecture, the base tensorflow packages are much too old, but you maybe can use the IBM powerai anaconda channel:

conda config --prepend channels https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/
source activate hklgym
conda install tensorflow-gpu

Install openai gym into your python enviroment as:

pip install gym

We depend on our own fork of openai baselines, which can be installed from the git repo as:

git clone https://github.com/scattering/baselines.git
(cd baselines && pip install -e .)

or directly from pip as:

pip install "git+https://github.com/scattering/baselines.git#egg=baselines"

Need pycrysfml installed from source:

git clone https://github.com/scattering/pycrysfml.git
cd pycrysfml
# ... may need to modify machine environment in build.sh ...
./build.sh
pip install -e .
cd ..

Finally, change into this source directory and type:

git clone https://github.com/scattering/HklEnv.git
(cd HklEnv && pip install -e .)

which installs hkl as an edittable environment in the gym.

To run HklEnv on a particular problem, define the configuration files in a datapath and set the following os environment variable:

HKL_DATAPATH=/path/to/problem
HKL_STOREPATH=/path/to/save

Default is "pycrysfml/hklgen/examples/sxtal".

Run a training example on the new hkl environment:

python3 -m baselines.run --alg=a2c --env=HklEnv:hkl-v0 --num_timesteps=2e16 --log_path=~/logs/hklstest/ --num_env=1

Normally this command would be part of a slurm batch script, run.sh:

#!/bin/bash
#SBATCH -c 40
#SBATCH --gres=gpu:3
#SBATCH --partition=gpu
#SBATCH --time=96:00:00
module load anaconda3
source activate hklgym
python3 -m baselines.run --alg=a2c --env=HklEnv:hkl-v0 --num_timesteps=2e16 --log_path=~/logs/hklstest/ --num_env=1

submit using

sbatch run.sh

Using DREAM

Currently, the fit in this model uses the Marquardt-Levenburg method. If you'd like to use a more comprehensive fitting method, you'll need to add a patch to the bumps library, which can be found here.

With that patch, you can call dream directly from within the test_bumps_refl.py file. You should update the code on line 33 to look something like this:

   result = fitters.fit(problem, method='dream', name='demo', store='/home/output')

Problems

If you've come to the README because the simple setup.py didn't work, you probably need to uninstall the official baselines, and install our forked version. Use the following to uninstall:

pip uninstall baselines