Imbalanced Semi-supervised Learning with Bias Adaptive Classifier
Authors: Renzhen Wang, Xixi Jia, Quanziang Wang, Yichen Wu, Deyu Meng
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments under various imbalanced semi-supervised setups, and the results demonstrate that our method can be applied to different pseudo-labeling models and is superior to current state-of-the-art methods. |
| Researcher Affiliation | Academia | Renzhen Wang1, Xixi Jia2, Quanziang Wang1, Yichen Wu3, Deyu Meng1,4,5 1Xi an Jiaotong University, 2Xidian University, 3City University of Hong Kong 4Macau University of Science and Technology, 5Peng Cheng Laboratory |
| Pseudocode | Yes | Algorithm 1 learning to adapt classifier during training Input: labeled / unlabeled training data Dl / Du, labeled / unlabeled batch size n / m, max iterations T Output: classification network parameters {θ, ϕ} |
| Open Source Code | Yes | The source code is made publicly available at https://github.com/renzhenwang/bias-adaptive-classifier. |
| Open Datasets | Yes | We evaluate our approach on four benchmark datasets: CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009), STL-10 (Coates et al., 2011) and SUN397 (Xiao et al., 2010) |
| Dataset Splits | Yes | In detail, a labeled set and an unlabeled set are randomly sampled from the original training data, keeping the number of images for each class to be the same. Then both the two sets are tailored to be imbalanced by randomly discarding training images according to the predefined imbalance ratios γl and γu. ... To evaluate the model, we use its exponential moving average (EMA) version, and report the average test accuracy of the last 20 epochs following Berthelot et al. (2019b). |
| Hardware Specification | Yes | We herein measure floating point operations per second (FLOPS) using NVIDIA Ge Force RTX 3090 to quantify the training cost. |
| Software Dependencies | No | All our experiments are implemented with the Pytorch platform (Paszke et al., 2019)... trained with Adam optimizer (Kingma & Ba, 2015) under the default parameter setting... (No specific version numbers for PyTorch or Adam are provided, only citations to the papers introducing them.) |
| Experiment Setup | Yes | The learning rate is set as 2 * 10^-3 and the batch size is set as 64. The total number of training iterations are 2.5 * 10^5 as in Kim et al. (2020a). To evaluate the model, we follow the setting in Berthelot et al. (2019b) and use an exponential moving average (EMA) of its parameters with a decay rate of 0.999 at each iteration. |