Using Language to Extend to Unseen Domains
Authors: Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E. Gonzalez, Aditi Raghunathan, Anna Rohrbach
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we discuss our main experiments and results. We defer dataset details, the remainder of the experiments and their discussion to the Appendix (B, D, E). We evaluate LADS on two domain adaptation benchmarks, Domain Net (Peng et al., 2019) and CUBPaintings (Wang et al., 2020), as well as two benchmarks exhibiting color and contextual bias, Colored MNIST (Arjovsky et al., 2021) and Waterbirds (Sagawa et al., 2019). |
| Researcher Affiliation | Academia | Lisa Dunlap, Clara Mohri UC Berkeley {lisabdunlap,cmohri}@berkeley.edu Aditi Raghunathan Carnegie Mellon University raditi@cmu.edu Han Zhang, Devin Guillory, Trevor Darrell, Joseph E. Gonzalez, Anna Rohrbach UC Berkeley {pariszhang,dguillory,trevordarrell,jegonzal,anna.rohrbach}@berkeley.edu |
| Pseudocode | No | The paper describes the method and uses figures to illustrate the process, but it does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/lisadunlap/LADS. |
| Open Datasets | Yes | CUB-Paintings (one new domain) is composed of 2 datasets, CUB-200 (Wah et al., 2011), a fine-grained bird classification benchmark containing 200 different bird species and CUB-200Paintings (Wang et al., 2020) ... Domain Net (Peng et al., 2019) ... Colored MNIST (Arjovsky et al., 2021) was made by taking the original MNIST Digits (Deng, 2012) ... Waterbirds (Sagawa et al., 2019) is a synthetically created dataset which creates contextual bias by taking species of landbirds and waterbirds from the CUB-200 Wah et al. (2011) dataset and pasting them on forest and water backgrounds from the Places (Zhou et al., 2017) dataset. |
| Dataset Splits | Yes | For each baseline, we do a hyperparameter sweep across learning rate and weight decay and choose the parameters with the highest class-balanced validation accuracy. For LADS we also do a sweep across the parameters of the augmentation network, namely learning rate, weight decay, and α, and select a checkpoint based on the validation loss. In the training and validation sets, even numbers are red and odd numbers are blue, while in the test set digits are colored randomly. |
| Hardware Specification | Yes | We train on 10 GeForce RTX 2080 Ti GPUs. |
| Software Dependencies | No | The paper mentions using the 'Open AI CLIP model with a Vi T-L backbone' but does not provide specific version numbers for other key software dependencies like Python, PyTorch, or CUDA, which are necessary for full reproducibility. |
| Experiment Setup | Yes | The augmentation network faug used in LADS is a 2-layer MLP with input and output dimensions of 768 and a hidden dimension of 384. ... In general, we set α = 0.5, lr = 0.001, wd = 0.05. Our hyperparameter search spaces and final choice of hyperparameters are listed in Table 4. |