Adapting Neural Architectures Between Domains
Authors: Yanxi Li, Zhaohui Yang, Yunhe Wang, Chang Xu
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental evaluation shows that only a small part of Image Net will be sufficient for Adapt NAS to extend its architecture success to the entire Image Net and outperform state-of-the-art comparison algorithms. We perform extensive experiments on various domains to demonstrate the practical generalisability of Adapt NAS. |
| Researcher Affiliation | Collaboration | Yanxi Li 1, Zhaohui Yang 2,3, Yunhe Wang 2, Chang Xu 1 1 School of Computer Science, University of Sydney, Australia 2 Noah s Ark Lab, Huawei Technologies, China 3 Key Lab of Machine Perception (MOE), Department of Machine Intelligence, Peking University, China |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a specific link or explicit statement about the release of its source code. |
| Open Datasets | Yes | The source dataset, CIFAR-10, contains 50,000 samples in the training set. For target domain samples, we construct a subset of 50,000 samples from Image Net, containing 50 samples from each category, as target samples that we have access to during searching. We use three pairs of source and target domain of digits to demonstrate the generalisability of Adapt NAS. The first pair is MNIST [15] and MNIST-M [9]... The second pair...target domain is SVHN [18]... The last pair still includes SVHN as the target domain, but the source domain is the more divergent MNIST-M. |
| Dataset Splits | Yes | The source dataset, CIFAR-10, contains 50,000 samples in the training set. For target domain samples, we construct a subset of 50,000 samples from Image Net, containing 50 samples from each category, as target samples that we have access to during searching. We use 5 different values for α from 0 to 1 with an interval of 0.25. ... We also use 3 different values for β, including 0.50, 0.83 and 0.98. Table 1 shows the validation error during search and the test error of retraining on CIFAR-10 and the full Image Net. |
| Hardware Specification | No | The paper mentions 'GPU days' and 'GPU hours' as a measure of search cost but does not provide specific details about the hardware (e.g., GPU or CPU models) used for its own experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | Following many previous works [6, 7, 17, 25, 26], we use the NASNet search space [26]. There are 2 kinds of cells, including normal cells and reduction cells, and each cell has 7 nodes, including 2 input nodes, 1 output node and 4 computation nodes. We use a set of 8 different candidate operations. We use 5 different values for α from 0 to 1 with an interval of 0.25. ... We also use 3 different values for β, including 0.50, 0.83 and 0.98. An exponential scheduler is proposed by Ganin et al.[8], which updates γ by: γp = 2 / (1 + exp( -10 p)) - 1, ... We also test a cosine-based scheduler, which rises slower: γp = (1 - cos(p π)) / 2. |