A Continuous Mapping For Augmentation Design
Authors: Keyu Tian, Chen Lin, Ser Nam Lim, Wanli Ouyang, Puneet Dokania, Philip Torr
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Results over multiple benchmarks demonstrate the efficiency improvement of this work compared with previous methods. |
| Researcher Affiliation | Collaboration | Keyu Tian Software College Beihang University tiankeyu.00@gmail.com Chen Lin University of Oxford chen.lin@eng.ox.ac.uk Ser-Nam Lim Facebook AI Wanli Ouyang University of Sydney Puneet K. Dokania University of Oxford & Five AI Ltd. Philip H.S. Torr University of Oxford philip.torr@eng.ox.ac.uk |
| Pseudocode | Yes | Algorithm 1 A Reference Method for ADA : MCMC-Aug |
| Open Source Code | No | Both source codes and checkpoints will be released to the public, and future work may concern more image modalities like infrared/X-rays/ultrasound imaging, more diverse augmentations, or wider applicants, e.g., in other computer vision areas or the times-series processing. |
| Open Datasets | Yes | We evaluate our method on CIFAR-10/100 [18] and Image Net [7]. |
| Dataset Splits | Yes | For each datasets, a validation set is split from the training set and the testing set is only used for evaluating the final performance (not involved in the search phase). |
| Hardware Specification | Yes | Table 4: Comparison of the efficiency (GPU hours). GPU Device: V100 (for MCMC method). |
| Software Dependencies | No | No specific software dependencies with version numbers are mentioned in the paper. |
| Experiment Setup | Yes | All batch sizes for augmentation search mentioned in this work are set to 256 for CIFARs and 1024 for Image Net as standard. Then we run our MCMC-Aug with 200 epochs, the same as [16] over the training set. Step size and noise scale of SGLD is set according to [25]. The experiments reported have a fixed 0.4 step size, and a noise rate of 2 10 5. Other hyperparameters not mentioned here are directly imported from [28], provided in Supplementary C. |