projectABC

This repository houses code for training deep neural networks to approximate the likelihood function of a family of Sequential Sampling Models. We provide two separate strategies for this: the pointwise and the dataset approach. More details regarding the ideas behind this work can be found at https://elifesciences.org/articles/65074

al-mlp: The pointwise approach

Provides implementations for the "pointwise" approach.

Note that this code base is not self-contained. We suggest to use the tutorial you can find in the hddmnn_tutorial folder to use this method through an extension to the HDDM Python package. You can also check the hddmnn_tutorial under https://github.com/AlexanderFengler/hddmnn_tutorial for recent updates.

al-mlp/basic_simulator.py holds a python function to call the cython simulators (can be called directly as well) defined in al-mlp/cddm_data_simulation.pyx. Run the al-mlp/setup.py to load compile the cython simulators (needs the cython package installed).

al-mlp/kde_class.py holds the class which we use to define KDE based likelihood functions from simulator data.

al-mlp/keras_fit_model.py holds the code used to train the mlps given fully pre-pocessed training data.

al-mlp/full_training_data_generator.py holds the training data generators (internally making use of the basic_simulator.py function mentioned above).

The folder al-mlp/networks holds pre-trained networks for the models shown in the paper.

The folder al-mlp/samplers holds the mcmc-samplers (including DE-MCMC) used in the paper. They were called through al-mlp/method_comparison_sim.py. We however suggest to instead use HDDM at this point.

al-cnn: The dataset approach

al-cnn/src/config.py contains the path to model weights/checkpoints. Also used to specify the various hyperparameter settings, including the type of SSM (and related options such as number of choices, boundary values, etc.), bin size and dataset size.

After simulating the neccesary dataset, we prepare for training by creating tfrecords.

cd al-cnn/src

python create_tfrecords.py

To train the forward model, we run

CUDA_VISIBLE_DEVICES=0 python run_pipeline.py --train

To run the Importance Sampler, we run

CUDA_VISIBLE_DEVICES=0 python sampler_inference.py --model <ddm> --nsample <1> --nbin <256> --N <1024> --proposal <normal/tdist>

Just to run MLE for parameter recovery

CUDA_VISIBLE_DEVICES=0 python inference.py --model <ddm> --nbin <256> --N <1024>

hddmnn_tutorial

Provides a demo for using the above two approaches in the context of hierarchical parameter estimation using the HDDM package.