ConR: Contrastive Regularizer for Deep Imbalanced Regression

Authors: Mahsa Keramati, Lili Meng, R. David Evans

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our comprehensive experiments show that Con R significantly boosts the performance of all the state-of-the-art methods on four large-scale deep imbalanced regression benchmarks.
Researcher Affiliation Collaboration Mahsa Keramati1,2 , Lili Meng1, R. David Evans1 1 Borealis AI, 2 School of Computing Science, Simon Fraser University
Pseudocode Yes Algorithm 1 Con R: Contrastive regularizer for deep imbalanced regression
Open Source Code Yes Our code is publicly available in https://github.com/Borealis AI/Con R.
Open Datasets Yes We use three datasets curated by Yang et al. (2021) for the deep imbalanced regression problem: Age DB-DIR is a facial age estimation benchmark, created based on Age DB (Moschoglou et al., 2017). IMDB-WIKI-DIR is an age estimation dataset originated from IMDB-WIKI (Rothe et al., 2018). NYUD2-DIR is created based on NYU Depth Dataset V2 (Silberman et al., 2012) to predict the depth maps from RGB indoor scenes. Moreover, we create MPIIGaze-DIR based on MPIIGaze, which is an appearance-based gaze estimation benchmark.
Dataset Splits Yes IMDB-WIKI (Rothe et al., 2018) has 191.5K images for training, and 11.0K images for validation and testing, respectively.
Hardware Specification Yes We use four NVIDIA Ge Force GTX 1080 Ti GPU to train all models.
Software Dependencies No The paper mentions software components like Resnet50, Adam optimizer, and Le Net, but does not provide specific version numbers for any libraries or frameworks used.
Experiment Setup Yes The batch size is 64 and the learning rate is 2.5 10 4 and decreases by 10 at epoch 60 and epoch 80. We use the Adam optimizer with a momentum of 0.9 and a weight decay of 1e-4. Following the baselines (Yang et al., 2021) the loss function for regression LR is Mean Absolute Error(MAE). All the models are trained for 90 epochs.