Towards Calibrated Model for Long-Tailed Visual Recognition from Prior Perspective

Authors: Zhengzhuo Xu, Zenghao Chai, Chun Yuan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments verify that our strategies contribute to a better-calibrated model and their combination achieves state-of-the-art performance on CIFAR-LT, Image Net-LT, and i Naturalist 2018.
Researcher Affiliation Academia Zhengzhuo Xu1 , Zenghao Chai1 , Chun Yuan1,2 1Shenzhen International Graduate School, Tsinghua University 2Peng Cheng Laboratory
Pseudocode Yes Algorithm 1 Integrated training manner towards calibrated model.
Open Source Code No The paper does not contain an explicit statement or link providing open-source code for the described methodology.
Open Datasets Yes The imbalanced datasets CIFAR-10-LT and CIFAR-100-LT are constructed via suitably discarding training samples following previous works [64, 12, 44, 6].
Dataset Splits Yes In our experiment, we utilize the balanced validation set constructed by Cui et al. [12] for fair comparisons. The official splits of train and validation images [6, 64, 30] are adopted for fair comparisons.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper does not explicitly list software dependencies with specific version numbers.
Experiment Setup Yes Input: Dtrain, Batch Size N, Stop Steps T1, T2, Random Sampler R, Uni Mix Sampler R Output: Optimized , i.e., feature extractor parameters , classifier parameters W , b 1: Initialize the parameters (0) randomly and calculate By via Eq.14 2: for t = 0 to T1 do 3: Sample a mini-batch B = {xi, yi}N i=1 R(Dtrain, N) 4: Sample a mini-batch B = {x j=1 R (Dtrain, N) 5: Calculate Uni Mix factor via Eq.6 6: Construct VRM dataset B = {exk, eyk}N k=1 via Eq.7 7: Calculate LB = E[ i,j LB(yi, (ex; (t))) + (1 j , (ex; (t)))] via Eq.11,15 8: Update (t+1) (t) r (t)LB 9: end for 10: for t = T1 to T2 do 11: Sample a mini-batch B = {xi, yi}N i=1 R(Dtrain, N) 12: Calculate LB = E[LB(yi, (xi; ))] via Eq.15 13: Update (t+1) (t) r (t)LB 14: end for