BECLR: Batch Enhanced Contrastive Few-Shot Learning

Authors: Stylianos Poulakakis-Daktylidis, Hadi Jamali-Rad

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We then present a suite of extensive quantitative and qualitative experimentation to corroborate that BECLR sets a new state-of-the-art across ALL existing U-FSL benchmarks (to the best of our knowledge), and significantly outperforms the best of the current baselines (codebase available at Git Hub).
Researcher Affiliation Collaboration Stylianos Poulakakis-Daktylidis1 & Hadi Jamali-Rad1,2 1Delft University of Technology (TU Delft), The Netherlands 2Shell Global Solutions International B.V., Amsterdam, The Netherlands
Pseudocode Yes The pretraining pipeline of BECLR is summarized in Algorithm 1, and a Pytorch-like pseudo-code can be found in Appendix E. [...] Algorithm 3: Unsupervised Pretraining of BECLR: Py Torch-like Pseudocode [...] Algorithm 4: Dynamic Clustered Memory (Dy CE): Py Torch-like Pseudocode
Open Source Code Yes We then present a suite of extensive quantitative and qualitative experimentation to corroborate that BECLR sets a new state-of-the-art across ALL existing U-FSL benchmarks (to the best of our knowledge), and significantly outperforms the best of the current baselines (codebase available at Git Hub).
Open Datasets Yes We evaluate BECLR in terms of its in-domain performance on the two most widely adopted few-shot image classification datasets: mini Image Net (Vinyals et al., 2016) and tiered Image Net (Ren et al., 2018). Additionally, for the in-domain setting, we also evaluate on two curated versions of CIFAR-100 (Krizhevsky et al., 2009) for FSL, i.e., CIFAR-FS and FC100.
Dataset Splits Yes mini Image Net. It is a subset of Image Net (Russakovsky et al., 2015), containing 100 classes with 600 images per class. We randomly select 64, 16, and 20 classes for training, validation, and testing, following the predominantly adopted settings of Ravi & Larochelle (2016).
Hardware Specification Yes All training and evaluation experiments are conducted on 2 A40 NVIDIA GPUs.
Software Dependencies No We use Py Torch (Paszke et al., 2019) for all implementations. The paper mentions 'PyTorch' and cites its paper, but does not provide a specific version number for PyTorch or any other software libraries used.
Experiment Setup Yes We use a batch size of B = 256 images for all datasets, except for tiered Image Net (B = 512). [...] We use the SGD optimizer with a weight decay of 10 4, a momentum of 0.995, and a cosine decay schedule of the learning rate. [...] The initial learning rate is set to 0.3 for the smaller mini Image Net, CIFAR-FS, FC100 datasets and 0.1 for tiered Image Net, and we train for 400 and 200 epochs, respectively. The temperature scalar in the loss function is set to τ = 2.