A Closer Look at Smoothness in Domain Adversarial Training

Authors: Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, Arihant Jain, Venkatesh Babu Radhakrishnan

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We extensively verify the empirical efficacy of SDAT over DAT across various datasets for classification (i.e., Domain Net, Vis DA-2017 and Office-Home) with Res Net and Vision Transformer (Dosovitskiy et al., 2020) (Vi T) backbones. We also show a prototypical application of SDAT in DA for object detection, demonstrating it s diverse applicability.
Researcher Affiliation Collaboration 1Video Analytics Lab, Indian Institute of Science, Bengaluru, India 2PES University, Bengaluru 3Amazon, India (Work done at Indian Institute of Science, Bengaluru).
Pseudocode Yes L. Py Torch Pseudocode for SDAT
Open Source Code Yes The source code used for experiments is available at: https://github.com/val-iisc/SDAT.
Open Datasets Yes We evaluate our proposed method on three datasets: Office Home, Vis DA-2017, and Domain Net... Office-Home (Venkateswara et al., 2017):... Domain Net (Peng et al., 2019):... Vis DA-2017 (Peng et al., 2017):
Dataset Splits Yes We split the target data into train and validation sets and report the best m AP on validation data.
Hardware Specification Yes All the above experiments were run on Nvidia V100, RTX 2080 and RTX A5000 GPUs.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al., 2019)', 'Detectron2 (Wu et al., 2019)', and 'Wandb (Biewald, 2020)' with citations indicating their year of publication/release, but it does not provide explicit version numbers (e.g., PyTorch 1.9) for these software components.
Experiment Setup Yes We use a learning rate of 0.01 with batch size 32 in all of our experiments with Res Net backbone... The ρ value is set to 0.02 for the Office-Home experiments, 0.005 for the Vis DA-2017 experiments and 0.05 for the Domain Net experiments. ...We train it for a total of 30 epochs with 1000 iterations per epoch. The momentum parameter in SGD is set to 0.9 and a weight decay of 0.001 is used.