Variational Depth Search in ResNets
One-shot neural architecture search allows joint learning of weights and network architecture, reducing computational cost. We limit our search space to the depth of residual networks and formulate an analytically tractable variational objective that allows for obtaining an unbiased approximate posterior over depths in one-shot. We propose a heuristic to prune our networks based on this distribution. We compare our proposed method against manual search over network depths on the MNIST, Fashion-MNIST, SVHN datasets. We find that pruned networks do not incur a loss in predictive performance, obtaining accuracies competitive with unpruned networks. Marginalising over depth allows us to obtain better-calibrated test-time uncertainty estimates than regular networks, in a single forward pass. Paper appearing at 1st Workshop on Neural Architecture Search at ICLR 2020.
Requirements
Python packages:
- test-tube 0.7.5
- Pytorch 1.3.1, torchvision 0.4.2
- Numpy 1.17.4
- Matplotlib 3.1.2
- scikit-learn 0.22
- scypy 1.3.3
Running Experiments from the Paper
Integers passed as an argument to python scripts represent which cuda device to use. If you only have one GPU, use 0. If you dont have a GPU, pass any integer and your CPU will be used automatically. First change into experiments directory:
cd experiments
Spirals
In order to reproduce the plots from our paper exactly, you will need to run each script multiple times. Different runs of each experiment are automatically saved separately. Train Learnt Depth Networks with every maximum depth:
python scan_max_depth_spirals.py 0 Train deterministic depth networks of every depth:
python scan_deterministic_depth_spirals.py 0 Train Learnt Depth Networks with different dataset sizes:
python scan_data_amount.py 0 Train Learnt Depth Networks with different dataset complexity:
python scan_spiral_complexity.py 0 Train Learnt Depth Networks with different widths:
python scan_width_spirals.py 0
Images
Each script runs experiments on all three datasets (MNIST, Fashion-MNIST and SVHN) and repeats each experiment 4 times. Train Learnt Depth Networks:
python scan_max_depth_images.py 0 Train Deterministic Depth Networks:
python scan_deterministic_depth_images.py 0
Generate Plots from Paper
All plotting code is contained within the notebooks in the ./notebooks/ folder. Once the experiments have been run, running the notebooks will generate the plots.
Citation
If you find this repo useful, please cite:
@misc{antoran2020variational,
title={Variational Depth Search in ResNets},
author={Javier Antorán and James Urquhart Allingham and José Miguel Hernández-Lobato},
year={2020},
eprint={2002.02797},
archivePrefix={arXiv},
primaryClass={stat.ML}
}