Dual Adversarial Semantics-Consistent Network for Generalized Zero-Shot Learning
Authors: Jian Ni, Shanghang Zhang, Haiyong Xie
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our approach achieves significant improvements over the state-of-the-art approaches. |
| Researcher Affiliation | Academia | Jian Ni1 nj1@mail.ustc.edu.cn Shanghang Zhang2 shanghaz@andrew.cmu.edu Haiyong Xie3,4,1 haiyong.xie@ieee.org 1University of Science and Technology of China, Anhui 230026, China 1 2Carnegie Mellon University, Pittsburgh, PA 15213, USA 2 3Advanced Innovation Center for Human Brain Protection, Capital Medical University, Beijing 3 100054, China 4 4National Engineering Laboratory for Public Safety Risk Perception and Control by Big Data 5 (NEL-PSRPC), Beijing 100041, China 6 |
| Pseudocode | Yes | The training procedure of our framework is summarized in Algorithm 1. In each iteration, the discriminators DV , DS are optimized for n1, n2 steps using the loss introduced in Eq. (6) and Eq. (8) respectively, and then one step on generators with Eq. (7) and Eq. (9) after the discriminators have been trained. According to [30], such a procedure enables the discriminators to provide more reliable gradient information. The training for traditional GANs suffers from the issue that the sigmoid cross-entropy is locally saturated as discriminator improves, which may lead to vanishing gradient and need to balance discriminator and generator carefully. Compared to the traditional GANs, the Wasserstein distance is differentiable almost everywhere and demonstrates its capability of extinguishing mode collapse. We put the detailed algorithm for training DASCN model in the supplemental material. |
| Open Source Code | No | The paper does not provide any concrete access to source code or explicitly state that it is open-source. |
| Open Datasets | Yes | To test the effectiveness of the proposed model for GZSL, we conduct extensive evaluations on four benchmark datasets: CUB [27], SUN [21], AWA1 [15], a PY [6] and compare the results with state-of-the-art approaches. Statistics of the datasets are presented in Table 1. |
| Dataset Splits | Yes | For fair comparison, we follow the training/validation/testing split as described in [28]. |
| Hardware Specification | No | The paper states: "Our implementation is based on Py Torch." but does not specify any hardware details like GPU, CPU, or memory. |
| Software Dependencies | No | The paper mentions "Py Torch" but does not specify its version number or any other software dependencies with version numbers. |
| Experiment Setup | Yes | Both the generators and discriminators are MLP with Leaky Re LU activation. In the primal GAN, GSV has a single hidden layer containing 4096 nodes and an output layer that has a Re LU activation with 2048 nodes, while the discriminator DV contains a single hidden layer with 4096 nodes and an output layer without activation. GV S and DS in the dual GAN have similar architecture with GSV and DV respectively. We use λ1 = λ4 = 10 as suggested in [9]. For loss term contributions, we cross-validate and set λ2 = λ3 = λ6 = 0.01, λ5 = 0.1. We choose noise z with the same dimensionality as the class embedding. Our model is optimized by Adam [12] with a base learning rate of 1e 4. |