Corrected Noisy Expected Improvement function

This is the code for "A Corrected Expected Improvement Acquisition Function Under Noisy Observations"[arxiv].

Key dependencies

(excluding commonly used packages such as scipy, numpy, torch etc.)

• python>=3.8
• torch=2.0.1 (install using pip rather than conda)
• gpytorch

  pip install gpytorch

• botorch (https://github.com/pytorch/botorch)

  pip install botorch

• chainer (https://github.com/chainer/chainer)

   pip install chainer

Toy example

The following example compares our proposed acquisition function with expected improvement under noisy observations on a simple synthetic function.

python toy_example.py

Synthetic function

python benchmark.py --output_dir OUTPUT_DIR --acq {acq_name}

acq can be 'q_NEI', 'NEI', 'PI', 'UCB', 'EI_C', 'PI_C' or 'EI'.

Model compression

Prepare dataset

• ImageNet (ILSVRC2012) The dataset can be found on the official website if you are affiliated with a research organization. It is also available on Academic torrents. Download the ILSVRC2012_img_train.tar and extract those images under the folder './data/ILSVRC2012'. Then run the following code to process the ImageNet dataset.

cd ./compression/imagenet
python extract_image.py

• MNIST (https://github.com/datapythonista/mnist) Download the minst dataset and extract those images under the folder './data/mnist'.

Run compression task

python compress_task.py --output_dir OUTPUT_DIR --acq {acq_name} --model {model_name}

```model``` can be 'Resnet50', 'VGG16' or 'FC3'.