Skip to content

Latest commit

 

History

History
67 lines (48 loc) · 2.02 KB

DEVELOPMENT.md

File metadata and controls

67 lines (48 loc) · 2.02 KB
Raw

Setup project

Requirements

- TensorFlow
- XFoil
  1. Clone the repository and activate virtual-env there.
git clone [email protected]:heytitle/blackbox-optimization-using-rnns.git
  1. Once virtal-env activated, install dependencies
(neural-network-project) $  pip install -r requirements.txt
  1. Setup Python's path
# in bin/activate (virtualenv's activate)
export PYTHONPATH="./src:$PYTHONPATH"

Train a lstm model

Each experiment is corresponding to training RNN on training data determined with --dimension with combinations of hyperparameter specifid in config.yaml. The experiment-manager.py will write training logs to a directory(--log-dir) under LOG_BASE, for example ./log/2d. The example below shows how to run the command.

TF_CPP_MIN_LOG_LEVEL=3 python ./scripts/train-lstm-model.py run --dimension 2 --log_dir 2d --kernel rbf --epochs 2

NOTE : Before running any experiment

1. Make sure that the resposity is up-to-dated, otherwise the results might not be consistent
2. Make sure that `--log_dir` is empty.

Useful commands

  1. tmux

    If you want to leave the experiment running while you're logging out from the server. You have to use tmux. You can do this by running tmux after log-in to the server.

  2. nvidia-smi

    Monitor GPU usage. Combining this command with watch to get near realtime updated.

    $ watch -n 2 nvidia-smi #update every 2 secs

Experiments

SKOpt's optimizers on GP test data

python ./scripts/skopt-test-data-experiment.py run --kernel rbf --dim 6 --no_testing_func 2000 --optimizer gp --dataset prior0 --n_steps 21

Airfoil Optimization

This experiment requires xfoil.

# Make sure XFOIL_PATH set properly
python ./scripts/airfoil-experiment.py run --optimizer basinhopping  --normalization 100 --dim 6

MNIST Hyperparameters Optimization

python ./scripts/mnist-experiment.py run --optimizer basinhopping  --no_runs 5