SitNet: Discrete Similarity Transfer Network for Zero-shot Hashing
Authors: Yuchen Guo, Guiguang Ding, Jungong Han, Yue Gao
IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on three benchmarks validate the superiority of Sit Net to the state-of-the-arts. |
| Researcher Affiliation | Academia | School of Software, Tsinghua University, Beijing 100084, China School of Computing & Communications, Lancaster University, UK yuchen.w.guo@gmail.com, {dinggg,gaoyue}@tsinghua.edu.cn,jungong.han@northumbria.ac.uk |
| Pseudocode | No | The paper describes an optimization algorithm and how gradients are computed, but it does not provide a formally labeled pseudocode block or algorithm. |
| Open Source Code | No | The paper does not provide any statement about releasing source code for the described methodology or a link to a code repository. |
| Open Datasets | Yes | Animals with Attributes [Lampert et al., 2014]. Aw A dataset consists of 30, 475 images manually labeled by 50 animal categories... CIFAR10 [Krizhevsky et al., 2012]. CIFAR10 dataset contains 10 non-overlapping objects... Image Net [Deng et al., 2009]. Image Net is a large-scale vision dataset organized according to Word Net hierarchy. |
| Dataset Splits | Yes | For Aw A dataset, we split the categories in to 5 groups where each group has 10 categories. We use one group as the unseen concepts and the other four as seen concepts and thus we have 5 different seen-unseen splits. For CIFAR10 dataset, we use one category as unseen and the other nine as seen categories which leads to 10 seen-unseen splits. For Image Net dataset... We use 10 categories as unseen and the other 90 as seen. For all three datasets, we construct the training and query set as follows. From the seen concepts, we randomly select 10, 000 images as the training set. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'Caffe' and 'word2vec tool' but does not specify version numbers for these software components. |
| Experiment Setup | Yes | To train our network, we utilize the Caffe [Jia et al., 2014] tool and adopt the Alex Net as the base network by using its convolution and fc layers. In all experiment, the initial learning rate is set to 10-3 and the momentum is set to 0.9. The weight decay parameter is 0.0005. The mini-batch size is set to 128. The training terminates at the 50, 000-th iteration. In our model, the max-margin parameter is λ = 1, the weights of the regularized center loss are α = β = 10-3, and the weight of the semantic embedding loss is γ = 10-2. |