Improving Bi-level Optimization Based Methods with Inspiration from Humans’ Classroom Study Techniques
Authors: Pengtao Xie
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on various datasets demonstrate the effectiveness of our methods. (Abstract) |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, USA. Correspondence to: Pengtao Xie <p1xie@ucsd.edu>. |
| Pseudocode | Yes | Algorithm 1 Optimization algorithm for FAS (Section C.2) and Algorithm 2 Optimization algorithm for PQDR (Section C.3) |
| Open Source Code | No | The paper states: "We build our method upon official python packages for different differentiable search approaches, such as DARTS1 , P-DARTS2 and PC-DARTS3 ." and provides links to these *third-party* repositories, but it does not contain an explicit statement or link for *their own* implementation code. |
| Open Datasets | Yes | We used three image classification datasets: CIFAR-10, CIFAR-100, and Image Net (Deng et al., 2009)... (Section 6.1.1) and We used the Path VQA (He et al., 2021) dataset... (Section 6.2.1) and We applied FAS on six text classification datasets in the GLUE collection (Wang et al., 2018). (Section I.1) and We performed the experiments on five graph classification datasets6, including PROTEINS, D&D, NCI1, NCI109, and Mutagenicity. Datasets can be downloaded from https://ls11-www.cs.tu-dortmund.de/staff/morris/graphkerneldatasets (Section I.2) |
| Dataset Splits | Yes | For CIFAR-10 and CIFAR-100, each of them is split into a 25K training set, a 25K validation set, and a 10K test set. (Section 6.1.1) and For architecture search on Image Net, following (Xu et al., 2020), we randomly sample 10% of the 1.2M images as a training set, and randomly sample 2.5% of the 1.2M images as a validation set. (Section 6.1.1) |
| Hardware Specification | Yes | Experiments were conducted on Nvidia 1080Ti GPU. (Section 6.1.2) and We implement the methods using Py Torch and perform training on four GTX 1080Ti GPUs. (Section D.3) and Eight Tesla V100 GPUs were used. (Section L.2 for ImageNet PC-DARTS) |
| Software Dependencies | Yes | We use Py Torch to implement all models. The version of Torch is 1.4.0 (or above). (Section D.1) |
| Experiment Setup | Yes | The number K of augmented images is set to 5. Augmentation operations (Perez & Wang, 2017) include random rotation, flipping, cropping, and color jitter. The tradeoff parameters γ and λ are set to 1 and 0.5 respectively. (Section 6.1.2) and further detailed in Additional Experimental Settings (Section D.1, D.2, D.3, etc.). For example, in D.1: The search algorithm runs for 50 epochs with a batch size of 64. Model weights are optimized using SGD, with an initial learning rate of 0.025.... |