Auto-Panoptic: Cooperative Multi-Component Architecture Search for Panoptic Segmentation
Authors: Yangxin Wu, Gengwei Zhang, Hang Xu, Xiaodan Liang, Liang Lin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct all experiments on 8 Tesla V100 GPUs using Py Torch [33]. We randomly select 5k images and 2k images from the training set of COCO and ADE20K as our validation set in the architecture search phase. Before architecture search, the supernet is trained for 12 epochs and 24 epochs on COCO and ADE20K, respectively, using mini-batch SGD with a weight decay of 0.0001 and a momentum of 0.9. The initial learning rate is 0.02 and is divided by 10 at the 8th and 11th epoch for COCO, 16th and 22th epoch for ADE20K, respectively. Our Auto-Panoptic achieves new state-of-the-art results on two challenging benchmarks, i.e., COCO and ADE20K and we conduct extensive experiments to demonstrate the robustness of the searched architecture and the effectiveness of our framework. |
| Researcher Affiliation | Collaboration | 1Sun Yat-sen University, 2Huawei Noah s Ark Lab, 3Dark Matter AI Research |
| Pseudocode | Yes | Algorithm 1 Auto-Panoptic Supernet Training Strategy and details of Path-Priority Search Policy. |
| Open Source Code | Yes | Codes and models are available at: https://github.com/Jacobew/Auto Panoptic. |
| Open Datasets | Yes | We conduct experiments on MS-COCO [25] and ADE20K [45]. MS-COCO is one of the most challenging datasets consisting of 115k images for the training set, 5k images for the validation set, and 20k images for the test-dev, and there are 80 things and 53 stuff classes in total. ADE20K is a dataset with more than 20k scene-centric images annotated with objects and object parts. It consists of 20k images for training and 2k images for validation, with 100 things and 50 stuff classes. |
| Dataset Splits | Yes | We randomly select 5k images and 2k images from the training set of COCO and ADE20K as our validation set in the architecture search phase. |
| Hardware Specification | Yes | We conduct all experiments on 8 Tesla V100 GPUs using Py Torch [33]. |
| Software Dependencies | No | The paper mentions 'Py Torch [33]' but does not provide specific version numbers for PyTorch or any other software dependencies needed for reproducibility. |
| Experiment Setup | Yes | Before architecture search, the supernet is trained for 12 epochs and 24 epochs on COCO and ADE20K, respectively, using mini-batch SGD with a weight decay of 0.0001 and a momentum of 0.9. The initial learning rate is 0.02 and is divided by 10 at the 8th and 11th epoch for COCO, 16th and 22th epoch for ADE20K, respectively. In our settings, L1 = 40, L 1 = 7, L2 = 9, and the overall search space includes over 440 67 39 6.7 1033 candidate architectures. δ1 and δ2 are Re LU and 1-Sigmoid respectively. For architecture search, we find T = 5, E = 12 suffices to find a satisfactory model, which makes the total number of enumerated models in the Path-Priority Search Policy roughly equals to that of one generation in EA (60 vs 50). |