Hierarchical Granularity Transfer Learning
Authors: Shaobo Min, Hongtao Xie, Hantao Yao, Xuran Deng, Zheng-Jun Zha, Yongdong Zhang
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on three benchmarks with hierarchical granularities show that Big SPN is an effective framework for Hierarchical Granularity Transfer Learning. |
| Researcher Affiliation | Academia | 1University of Science and Technology of China, Hefei, China 2 Institute of Automation, Chinese Academy of Sciences. Beijing, China |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | we construct three datasets with hierarchical categories and semantic descriptions, i.e., CUB-HGTL, AWA2-HGTL, and Flower-HGTL, which are based on the existing datasets of Caltech-USCD Birds200-2011 [36], Animals with Attributes 2 [39], and Flower [26], respectively. |
| Dataset Splits | Yes | The CUB...Finally, we construct the CUB-HGTL dataset whose training set consists of three components: 1) images along with basic-level category annotations...To split the train/val sets for AWA2-HGTL and Flower-HGTL, we randomly divide the images of each sub-category by 3 : 2, and report the averaged performance for multiple splits. The final data structures of AWA2-HGTL and Flower-HGTL are consistent with CUB-HGTL. Table 1: Some statistics of the experimental datasets. Train Test CUB-HGTL 5,994 5,794 AWA2-HGTL 22,392 14,930 Flower-HGTL 4917 3272 |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts) used for running its experiments. It only mentions the backbone network used: 'The backbone network uses the Res Net-101 [11].' |
| Software Dependencies | No | The paper mentions 'Res Net-101' as the backbone network and 'SGD optimizer' but does not provide specific version numbers for any software dependencies like programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | The backbone network uses the Res Net-101 [11]. MSRA random initializer is used to initialize the Big SPN. In terms of data augmentation, 448 448 random cropping and horizontal flipping are applied to the input images. Specifically, Lsa and Lse are alternately optimized for each data batch. The batch size is N = 12, and reduction channel is D = 256. The SGD optimizer is used with initial lr = 0.001, momentum=0.9, and 180 training epoch. The hyper-parameter is set by K = 4 and λ = 1, which will be analyzed later. |