Ada-Segment: Automated Multi-loss Adaptation for Panoptic Segmentation
Authors: Gengwei Zhang, Yiming Gao, Hang Xu, Hao Zhang, Zhenguo Li, Xiaodan Liang3333-3341
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our Ada-Segment brings 2.7% panoptic quality (PQ) improvement on COCO val split from the vanilla baseline, achieving the state-of-the-art 48.5% PQ on COCO test-dev split and 32.9% PQ on ADE20K dataset. The extensive ablation studies reveal the ever-changing dynamics throughout the training process, necessitating the incorporation of an automated and adaptive learning strategy as presented in this paper. |
| Researcher Affiliation | Collaboration | Gengwei Zhang1, Yiming Gao1, Hang Xu2, Hao Zhang3, Zhenguo Li2, Xiaodan Liang1 1Sun Yat-Sen University 2Huawei Noah s Ark Lab 3Shanghai Jiao Tong University |
| Pseudocode | Yes | Algorithm 1 The Ada-Segment framework. Input: Iterations between two checkpoints q, Initial Loss State l1, Number of checkpoints T, Number of training epochs E Initialize m models {M1, M2, ...Mm} Initialize policy network π l1 best l1 for t 1 to T, do Generate λt+1 by Equation 1 with π and lt best Sample m candidates by Equation 2 with λt+1 Train all models for q iterations with ˆΛt+1 Collect model performances vt and save lt best Obtain policy rewards by Equation 5 Update θ in π via Equation 6 Save πt π Update all models with M t best end for Initialize a model Mp for e 1 to E, do Generate λe+1 by Equation 7 with le Train model Mp with λe+1 end for return M p |
| Open Source Code | No | The paper does not provide any concrete access information to source code (e.g., specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described. |
| Open Datasets | Yes | COCO. Following the competition setting in 2019 Microsoft COCO panoptic segmentation, which consists of 133 classes with 80 things classes and 53 stuff classes. We only use train2017 split with approximately 118k images for training and report the results on val split with 5k images. ... ADE20K. ADE20K is a challenging dataset with densely labeled 22k images, with 100 things classes and 50 stuff classes. |
| Dataset Splits | Yes | For COCO, we randomly sample 10k images from the 118k training set for validation and the rest part as the proxy training set. For ADE20K, we train on 20k training images in which 2k images are randomly sampled images for validation during training. |
| Hardware Specification | No | The paper states, "For each model, we train totally 12 epochs (so called 1x setting) for COCO and 24 epochs for ADE20K on 8 GPUs with 2 images per GPU using Py Torch (Paszke et al. 2017)." However, it does not specify the exact model or type of GPUs used. |
| Software Dependencies | No | The paper mentions "using Py Torch (Paszke et al. 2017)" but does not provide a specific version number for PyTorch or any other software dependencies with version information. |
| Experiment Setup | Yes | We set the initial learning rate as 0.02 and weight decay as 0.0001 with stochastic gradient descent (SGD) for all experiments. ... For each model, we train totally 12 epochs (so called 1x setting) for COCO and 24 epochs for ADE20K on 8 GPUs with 2 images per GPU using Py Torch (Paszke et al. 2017). ... In the weight controller, we set the sampling standard deviation σ = 0.2, and we use three-layer MLP with hidden layer size 16 as the policy network. We use Adam (Kingma and Ba 2014) optimizer with learning rate 5e 2 and weight decay 5e 4 to optimize the policy network. |