Generative Modeling for Multi-task Visual Learning
Authors: Zhipeng Bao, Martial Hebert, Yu-Xiong Wang
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental evaluation on challenging multi-task benchmarks, including NYUv2 and Taskonomy, demonstrates that our MGM framework improves the performance of all the tasks by large margins, consistently outperforming state-of-the-art multi-task approaches in different sample-size regimes. |
| Researcher Affiliation | Academia | 1Carnegie Mellon University 2University of Illinois at Urbana-Champaign. Correspondence to: Zhipeng Bao <zbao@cs.cmu.edu>, Martial Hebert <hebert@cs.cmu.edu>, Yu-Xiong Wang <yxw@illinois.edu>. |
| Pseudocode | Yes | Algorithm 1 Training procedure of MGM |
| Open Source Code | Yes | The code of this work is available at https://github.com/zpbao/multi-task-oriented_generative_modeling. |
| Open Datasets | Yes | We evaluate our approach on standard multi-task benchmarks, including the NYUv2 (Nathan Silberman & Fergus, 2012) and Taskonomy (Zamir et al., 2018) datasets. |
| Dataset Splits | Yes | For NYUv2, we randomly select 1,049 images as the full training set and 200 images each as the validation/test set. For Tiny-Taskonomy, we randomly pick 80% of the whole set as the full training set and 10% each as the validation/test set. |
| Hardware Specification | No | The paper mentions "GPU memory constraints" when discussing model architecture, but it does not specify any exact hardware details such as specific GPU models, CPU models, or server specifications used for experiments. |
| Software Dependencies | No | The paper mentions software components like PyTorch, Adam optimizer, ResNet-18, SAGAN, SimCLR, and DCGAN, but it does not provide specific version numbers for these libraries, frameworks, or the programming language itself, which are necessary for reproducible dependency descriptions. |
| Experiment Setup | Yes | The learning rates are set to 0.001 for the multi-task, self-supervision, and refinement networks, 0.0001 for the SAGAN generator, and 0.0004 for the SAGAN discriminator. The batch size is set to 32. |