Improving Differentiable Neural Architecture Search by Encouraging Transferability
Authors: Parth Sheth, Pengtao Xie
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on several datasets demonstrate the effectiveness of our method. In this section, we present experimental results. We compare our method with state-of-the-art methods 1) on search spaces designed by (Zela et al., 2019) for measuring generalizability and stability of differentiable NAS methods; 2) on CIFAR-100, CIFAR-10, and Image Net under search spaces of DARTS, P-DARTS, PC-DARTS, and PR-DARTS. Please refer to the appendix for detailed hyperparameter settings and additional results. |
| Researcher Affiliation | Academia | Parth Sheth University of Pennsylvania parthfour@gmail.com Pengtao Xie University of California San Diego p1xie@eng.ucsd.edu |
| Pseudocode | Yes | Algorithm 1 shows the algorithm. |
| Open Source Code | No | The paper states: "We build our LPPDT upon official python packages for different differentiable search approaches, such as DARTS, P-DARTS and PC-DARTS." It provides links to these third-party repositories but does not explicitly state that the authors are releasing their own code for the TETLO framework described in the paper. |
| Open Datasets | Yes | Three datasets are used in the experiments, including Image Net (Deng et al., 2009), CIFAR100 (Krizhevsky et al., 2009), and CIFAR-10 (Krizhevsky & Hinton, 2010). |
| Dataset Splits | Yes | Image Net is split into a training set and a test set with 1.2M and 50K images respectively. CIFAR-100 is split into training, validation, and test sets, with 25K, 25K, and 10K images respectively. So is CIFAR-10. |
| Hardware Specification | Yes | Search cost is measured by GPU days on a Nvidia 1080Ti. All GPU costs are measured on 1080Ti GPUs. Eight Tesla V100 GPUs were used [for ImageNet PC-DARTS and P-DARTS evaluation]. |
| Software Dependencies | Yes | We use Py Torch to implement all models. The version of Torch is 1.4.0 (or above). |
| Experiment Setup | Yes | Please refer to the appendix for detailed hyperparameter settings and additional results. Major training details including hyperparameters, parameter tuning strategies, data splits, etc. are specified in Section 4.2, Section M, Section O, Section P, Section Q, Section T, and Section V. Tables 26-33 show the hyperparameter settings used in architecture search experiments. Tables 30-33 show the hyperparameter settings used in architecture evaluation experiments. |