Fully Parallel Hyperparameter Search: Reshaped Space-Filling
Authors: Marie-Liesse Cauwet, Camille Couprie, Julien Dehos, Pauline Luc, Jeremy Rapin, Morgane Riviere, Fabien Teytaud, Olivier Teytaud, Nicolas Usunier
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed methods are first validated on artificial experiments and simple real-world tests on clustering and Salmon mappings. Then we demonstrate that they bring performance improvement in a wide range of expensive artificial intelligence tasks, namely attend/infer/repeat, video next frame segmentation forecasting and progressive generative adversarial networks. |
| Researcher Affiliation | Collaboration | 1ESIEE, Universit e Paris-Est, LIGM (UMR 8049), CNRS, ENPC, ESIEE Paris, UPEM, F-77454, Marnela-Vall ee, France 2Facebook AI Research 3Universit e du Littoral Cˆote d Opale. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | See SM for reproduction of all our artificial experiments with one-liners in the Nevergrad platform (Rapin and Teytaud, 2018), for the Nevergrad oneshot experiment (SM, Section 4, or the website (Rapin and Teytaud, 2018))... All methods here are publicly available in the Nevergrad platform: Zero just returns (0, . . . , 0), methods with Avg in the name use sophisticated recommendation methods (i.e. they might recommend a point which was not in the sample). |
| Open Datasets | Yes | The dataset consists in 50000 images from Cifar10 (Krizhevsky et al., 2010) and 50000 object-free patches from COCO (Lin et al., 2014), split into balanced training (80% of the samples) and validation sets. |
| Dataset Splits | Yes | The dataset consists in 50000 images from Cifar10 (Krizhevsky et al., 2010) and 50000 object-free patches from COCO (Lin et al., 2014), split into balanced training (80% of the samples) and validation sets. |
| Hardware Specification | No | The paper mentions "short 10 minutes training on a single GPU" for Progressive GANs, but does not specify the model or any other details about the hardware used. |
| Software Dependencies | No | The paper mentions "Pytorch GAN Zoo (Riviere, 2019)", but it does not specify version numbers for PyTorch or any other software dependencies needed for reproducibility. |
| Experiment Setup | Yes | We optimize 3 continuous hyperparameters, namely leakiness of Relu units in [1e 2, 0.6], the discriminator ϵ parameter in [1e 5, 1e 1], and the base learning rate in [1e 5, 1e 1]... Meta Rctg uses Scrambled-Hammersley and Rctg reshaping with λ = 1 + log(n) / (4 * log(d)). |