FeLMi : Few shot Learning with hard Mixup
Authors: Aniket Roy, Anshul Shah, Ketul Shah, Prithviraj Dhar, Anoop Cherian, Rama Chellappa
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach on several common few-shot benchmarks FC-100, CIFAR-FS, mini Image Net and tiered Image Net and obtain improvements in both 1-shot and 5-shot settings. Additionally, we experimented on the cross-domain few-shot setting (mini Image Net CUB) and obtain significant improvements. |
| Researcher Affiliation | Collaboration | Aniket Roy Johns Hopkins University Baltimore, USA aroy28@jhu.edu Anshul Shah Johns Hopkins University Baltimore, USA ashah95@jhu.edu Ketul Shah Johns Hopkins University Baltimore, USA kshah33@jhu.edu Prithviraj Dhar Reality Labs, Meta Sunnyvale, USA prithvirj95@meta.com Anoop Cherian MERL Cambridge, MA cherian@merl.com Rama Chellappa Johns Hopkins University Baltimore, USA rchella4@jhu.edu |
| Pseudocode | Yes | Code will be released later. To ease reproducibility, we include pseudocode in Appendix. |
| Open Source Code | Yes | Code: https://github.com/aniket004/Felmi |
| Open Datasets | Yes | We experiment on four widely used few-shot benchmarks: FC-100 [17], CIFAR-FS [4], mini Image Net [33] and tiered Image Net [20]. Our approach achieves significant performance improvements for both the 5-shot and 1-shot setting in this setting as shown in Table 4. Since CUB performs fine-grained classification, we hypothesize that hard mixup samples are even more useful in this case. |
| Dataset Splits | Yes | FC-100 is a subset of CIFAR-100, containing 60 classes for meta-training, 20 classes for meta-validation and 20 classes for meta-testing. CIFAR-FS is also obtained from CIFAR-100, containing 64 classes for meta-training, 16 classes for meta-validation and 20 classes for meta-testing. mini Image Net is derived from Image Net with images downsampled to a resolution of 84 84 pixels. It has 64 classes for meta-training, 16 classes for meta-validation and 20 classes for meta-testing. tiered Image Net [20] is another subset of Image Net, containing total of 608 classes, from which 351 classes are used for meta-training, 97 classes for meta-validation and 160 classes for meta-testing. |
| Hardware Specification | No | The main paper text refers to supplementary material for this information: 'Total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Supplementary material'. |
| Software Dependencies | No | The paper mentions software components like 'Res Net-12 architecture' and 'SGD optimizer' but does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | Learning rate of backbone and classifier are set to 0.025 and 0.05 respectively with weight decay of 5e-4. Following [11], we use the temperature coefficient of 4.0 for our KD loss. We use a minibatch size of 250 for training. Using standard setting, we also perform data augmentation using color jittering, random crop and horizontal flip. Further details of the hyperparameters and training are provided in the supplementary material. |