Realistic evaluation of transductive few-shot learning

Authors: Olivier Veilleux, Malik Boudiaf, Pablo Piantanida, Ismail Ben Ayed

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate experimentally state-of-the-art transductive few-shot methods over 3 widely used datasets, and observe that the performances decrease by important margins, albeit at various degrees, when dealing with arbitrary class distributions.
Researcher Affiliation Academia Olivier Veilleux ÉTS Montreal Malik Boudiaf * ÉTS Montreal Pablo Piantanida L2S, Centrale Supélec CNRS Université Paris-Saclay Ismail Ben Ayed ÉTS Montreal
Pseudocode No The paper describes methods and procedures in text but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is publicly available at https://github.com/oveilleux/Realistic_Transductive_Few_Shot.
Open Datasets Yes Datasets We use three standard benchmarks for few-shot classification: mini-Imagenet [46], tiered Imagenet [30] and Caltech-UCSD Birds 200 (CUB) [47].
Dataset Splits Yes We used different Dirichlet s concentration parameter a for validation and testing. The validation-task generation is based on a random sampling within the simplex (i.e Dirichlet with a = 1K). Testing-task generation uses a = 2 1K to reflect the fact that extremely imbalanced tasks (i.e., only one class is present in the task) are unlikely to happen in practical scenarios; see Figure 1 for visualization. ... For all methods, hyper-parameter tuning is performed on the validation set of each dataset, using the validation sampling described in the previous paragraph.
Hardware Specification Yes All the experiments have been executed on a single GTX 1080 Ti GPU.
Software Dependencies No The paper mentions software components like PyTorch and libraries used implicitly by the described models (e.g., Res Net-18, WRN28-10), but it does not specify explicit version numbers for these software dependencies (e.g., 'PyTorch 1.9').
Experiment Setup Yes The feature extractors are trained for 90 epochs, using a learning rate initialized to 0.1 and divided by 10 at epochs 45 and 66. We use a batch size of 256 for Res Net-18 and of 128 for WRN28-10. During training, color jitter, random croping and random horizontal flipping augmentations are applied.