/fnp_neurips2020

Primary LanguagePythonApache License 2.0Apache-2.0

High level overview of FNP architecture

Examples

  • notebooks/simulation_experiments.ipynb - explore a FNP pretrained on simulated simple/complex cells
  • notebooks/visual_experiments.ipynb - measure performance of a FNP pretrained on real responses

Training the model

PYTHONPATH="." python tuning_manifold/train.py \ 
               --architecture='{"architecture": [[11, 32], [5, 8], [3, 8], [3, 8], 64], "padding": "same", "norm": "batch", "nb_orientations": 8}' \
               --cell_latent_dim=64 --synthetic_rf_dim=8 --image_shape="(16,16,1)"

Contents

The main files and directories for this project are

  • tuning_manifold -- the python project directory
  • tuning_manifold/train.py -- the training script
  • tuning_manifold/fnp_model.py -- the Factorized Neural Process model
  • tuning_manifold/synthetic_sampler.py -- generate TF Datasets of simulated responses
  • tuning_manifold/se2cnn -- the G-CNN library