Implicit Deep Adaptive Design (iDAD)

This code supports the NeurIPS paper 'Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods'.

@article{ivanova2021implicit,
  title={Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods},
  author={Ivanova, Desi R. and Foster, Adam and Kleinegesse, Steven and Gutmann, Michael and Rainforth, Tom},
  journal={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2021}
}

Computing infrastructure requirements

We have tested this codebase on Linux (Ubuntu x86_64) and MacOS (Big Sur v11.2.3) with Python 3.8. To train iDAD networks, we recommend the use of a GPU. We used one GeForce RTX 3090 GPU on a machine with 126 GiB of CPU memory and 40 CPU cores.

Installation

Ensure that Python and conda are installed.
Create and activate a new conda virtual environment as follows

conda create -n idad_code
conda activate idad_code

Install the correct version of PyTorch, following the instructions at pytorch.org. For our experiments we used torch==1.8.0 with CUDA version 11.1.
Install the remaining package requirements using pip install -r requirements.txt.
Install the torchsde package from its repository: pip install git+https://github.com/google-research/torchsde.git.

MLFlow

We use mlflow to log metric and store network parameters. Each experiment run is stored in a directory mlruns which will be created automatically. Each experiment is assigned a numerical and each run gets a unique . The iDAD networks will be saved in ./mlruns/ / /artifacts, which will be printed at the end of each training run.

Location Finding Experiment

To train an iDAD network with the InfoNCE bound to locate 2 sources in 2D, using the approach in the paper, execute the command

python3 location_finding.py \
    --num-steps 100000 \
    --num-experiments=10 \
    --physical-dim 2 \
    --num-sources 2 \
    --lr 0.0005 \
    --num-experiments 10 \
    --encoding-dim 64 \
    --hidden-dim 512 \
    --mi-estimator InfoNCE \
    --device <DEVICE>

To train an iDAD network with the NWJ bound, using the approach in the paper, execute the command

python3 location_finding.py \
    --num-steps 100000 \
    --num-experiments=10 \
    --physical-dim 2 \
    --num-sources 2 \
    --lr 0.0005 \
    --num-experiments 10 \
    --encoding-dim 64 \
    --hidden-dim 512 \
    --mi-estimator NWJ \
    --device <DEVICE>

To run the static MINEBED baseline, use the following

python3 location_finding.py \
    --num-steps 100000 \
    --physical-dim 2 \
    --num-sources 2 \
    --lr 0.0001 \
    --num-experiments 10 \
    --encoding-dim 8 \
    --hidden-dim 512 \
    --design-arch static \
    --critic-arch cat \
    --mi-estimator NWJ \
    --device <DEVICE>

To run the static SG-BOED baseline, use the following

python3 location_finding.py \
    --num-steps 100000 \
    --physical-dim 2 \
    --num-sources 2 \
    --lr 0.0005 \
    --num-experiments 10 \
    --encoding-dim 8 \
    --hidden-dim 512 \
    --design-arch static \
    --critic-arch cat \
    --mi-estimator InfoNCE \
    --device <DEVICE>

To run the adaptive (explicit likelihood) DAD baseline, use the following

python3 location_finding.py \
    --num-steps 100000 \
    --physical-dim 2 \
    --num-sources 2 \
    --lr 0.0005 \
    --num-experiments 10 \
    --encoding-dim 32 \
    --hidden-dim 512 \
    --mi-estimator sPCE \
    --design-arch sum \
    --device <DEVICE>

To evaluate the resulting networks eun the following command

python3 eval_sPCE.py --experiment-id <ID>

To evaluate a random design baseline (requires no pre-training):

python3 baselines_locfin_nontrainable.py \
    --policy random \
    --physical-dim 2 \
    --num-experiments-to-perform 5 10 \
    --device <DEVICE>

To run the variational baseline (note: it takes a very long time), run:

python3 baselines_locfin_variational.py \
    --num-histories 128 \
    --num-experiments 10 \
    --physical-dim 2 \
    --lr 0.001 \
    --num-steps 5000\
    --device <DEVICE>

Copy path_to_artifact and pass it to the evaluation script:

python3 eval_sPCE_from_source.py \
    --path-to-artifact <path_to_artifact> \
    --num-experiments-to-perform 5 10 \
    --device <DEVICE>

Pharmacokinetic Experiment

To train an iDAD network with the InfoNCE bound, using the approach in the paper, execute the command

python3 pharmacokinetic.py \
    --num-steps 100000 \
    --lr 0.0001 \
    --num-experiments 5 \
    --encoding-dim 32 \
    --hidden-dim 512 \
    --mi-estimator InfoNCE \
    --device <DEVICE>

To train an iDAD network with the NWJ bound, using the approach in the paper, execute the command

python3 pharmacokinetic.py \
    --num-steps 100000 \
    --lr 0.0001 \
    --num-experiments 5 \
    --encoding-dim 32 \
    --hidden-dim 512 \
    --mi-estimator NWJ \
    --gamma 0.5 \
    --device <DEVICE>

To run the static MINEBED baseline, use the following

python3 pharmacokinetic.py \
    --num-steps 100000 \
    --lr 0.001 \
    --num-experiments 5 \
    --encoding-dim 8 \
    --hidden-dim 512 \
    --design-arch static \
    --critic-arch cat \
    --mi-estimator NWJ \
    --device <DEVICE>

To run the static SG-BOED baseline, use the following

python3 pharmacokinetic.py \
    --num-steps 100000 \
    --lr 0.0005 \
    --num-experiments 5 \
    --encoding-dim 8 \
    --hidden-dim 512 \
    --design-arch static \
    --critic-arch cat \
    --mi-estimator InfoNCE \
    --device <DEVICE>

To run the adaptive (explicit likelihood) DAD baseline, use the following

python3 pharmacokinetic.py \
    --num-steps 100000 \
    --lr 0.0001 \
    --num-experiments 5 \
    --encoding-dim 32 \
    --hidden-dim 512 \
    --mi-estimator sPCE \
    --design-arch sum \
    --device <DEVICE>

To evaluate the resulting networks run the following command

python3 eval_sPCE.py --experiment-id <ID>

To evaluate a random design baseline (requires no pre-training):

python3 baselines_pharmaco_nontrainable.py \
    --policy random \
    --num-experiments-to-perform 5 10 \
    --device <DEVICE>

To evaluate an equal interval baseline (requires no pre-training):

python3 baselines_pharmaco_nontrainable.py \
    --policy equal_interval \
    --num-experiments-to-perform 5 10 \
    --device <DEVICE>

To run the variational baseline (note: it takes a very long time), run:

python3 baselines_pharmaco_variational.py \
    --num-histories 128 \
    --num-experiments 10 \
    --lr 0.001 \
    --num-steps 5000 \
    --device <DEVICE>

Copy path_to_artifact and pass it to the evaluation script:

python3 eval_sPCE_from_source.py \
    --path-to-artifact <path_to_artifact> \
    --num-experiments-to-perform 5 10 \
    --device <DEVICE>

SIR experiment

For the SIR experiments, please first generate an initial training set and a test set:

python3 epidemic_simulate_data.py \
    --num-samples=100000 \
    --device <DEVICE>

To train an iDAD network with the InfoNCE bound, using the approach in the paper, execute the command

python3 epidemic.py \
    --num-steps 100000 \
    --num-experiments 5 \
    --lr 0.0005 \
    --hidden-dim 512 \
    --encoding-dim 32 \
    --mi-estimator InfoNCE \
    --design-transform ts \
    --device <DEVICE>

To train an iDAD network with the NWJ bound, execute the command

python3 epidemic.py \
    --num-steps 100000 \
    --num-experiments 5 \
    --lr 0.0005 \
    --hidden-dim 512 \
    --encoding-dim 32 \
    --mi-estimator NWJ \
    --design-transform ts \
    --device <DEVICE>

To run the static SG-BOED baseline, run

python3 epidemic.py \
    --num-steps 100000 \
    --num-experiments 5 \
    --lr 0.005 \
    --hidden-dim 512 \
    --encoding-dim 32 \
    --design-arch static \
    --critic-arch cat \
    --design-transform iid \
    --mi-estimator InfoNCE \
    --device <DEVICE>

To run the static MINEBED baseline, run

python3 epidemic.py \
    --num-steps 100000 \
    --num-experiments 5 \
    --lr 0.001 \
    --hidden-dim 512 \
    --encoding-dim 32 \
    --design-arch static \
    --critic-arch cat \
    --design-transform iid \
    --mi-estimator NWJ \
    --device <DEVICE>

To train a critic with random designs (to evaluate the random design baseline):

python3 epidemic.py \
    --num-steps 100000 \
    --num-experiments 5 \
    --lr 0.005 \
    --hidden-dim 512 \
    --encoding-dim 32 \
    --design-arch random \
    --critic-arch cat \
    --design-transform iid \
    --device <DEVICE>

To train a critic with equal interval designs, which is then used to evaluate the equal interval baseline, run the following

python3 epidemic.py \
    --num-steps 100000 \
    --num-experiments 5 \
    --lr 0.001 \
    --hidden-dim 512 \
    --encoding-dim 32 \
    --design-arch equal_interval \
    --critic-arch cat \
    --design-transform iid \
    --device <DEVICE>

Finally, to evaluate the different methods, run

python3 eval_epidemic.py \
    --experiment-id <ID> \
    --device <DEVICE>

Implicit Deep Adaptive Design (iDAD)

Related tags

Overview

Implicit Deep Adaptive Design (iDAD)

Computing infrastructure requirements

Installation

MLFlow

Location Finding Experiment

Pharmacokinetic Experiment

SIR experiment

Owner

Desi

Video Autoencoder: self-supervised disentanglement of 3D structure and motion

Contains code for Deep Kernelized Dense Geometric Matching

Portfolio analytics for quants, written in Python

基于深度强化学习的原神自动钓鱼AI

Neural Turing Machines (NTM) - PyTorch Implementation

🔎 Monitor deep learning model training and hardware usage from your mobile phone 📱

Source code for paper: Knowledge Inheritance for Pre-trained Language Models

Deep Learning segmentation suite designed for 2D microscopy image segmentation

Towards Part-Based Understanding of RGB-D Scans

《LightXML: Transformer with dynamic negative sampling for High-Performance Extreme Multi-label Text Classiﬁcation》(AAAI 2021) GitHub:

Source code for our EMNLP'21 paper 《Raise a Child in Large Language Model: Towards Effective and Generalizable Fine-tuning》

Code for our NeurIPS 2021 paper Mining the Benefits of Two-stage and One-stage HOI Detection

NeuroLKH: Combining Deep Learning Model with Lin-Kernighan-Helsgaun Heuristic for Solving the Traveling Salesman Problem

Team Enigma at ArgMining 2021 Shared Task: Leveraging Pretrained Language Models for Key Point Matching

Official code for paper "Demystifying Local Vision Transformer: Sparse Connectivity, Weight Sharing, and Dynamic Weight"

Designing a Minimal Retrieve-and-Read System for Open-Domain Question Answering (NAACL 2021)

Simple codebase for flexible neural net training

Code for Towards Unifying Behavioral and Response Diversity for Open-ended Learning in Zero-sum Games

Implementation of GGB color space

Label Mask for Multi-label Classification