Generator Born from Classifier
Authors: Runpeng Yu, Xinchao Wang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical validation from various image generation tasks substantiates the efficacy of our strategy. We conduct experiments on commonly used image datasets. Fig. 1 shows some generated images of the MNIST and Celeb A datasets. |
| Researcher Affiliation | Academia | Runpeng Yu Xinchao Wang: National University of Singapore r.yu@u.nus.edu xinchao@nus.edu.sg |
| Pseudocode | Yes | The training procedure is summarized in Alg. 1 in the Appendix. The algorithm for training a generator using multiple classifiers is summarized in Alg. 2 in the Appendix. |
| Open Source Code | No | The paper does not contain any explicit statements about open-sourcing the code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | In this subsection, we showcase the experimental results on the MNIST [Lecun et al., 1998] and Celeb A [Liu et al., 2015] datasets. |
| Dataset Splits | No | The paper describes setting up classification tasks with sampled training data (e.g., "500 training data (50 images per class)" for MNIST), but it does not specify explicit training/validation/test splits or mention a separate validation set. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for conducting the experiments, such as GPU models, CPU specifications, or memory. |
| Software Dependencies | No | The paper mentions components like "Re LU activation function" and "batch normalization" and initialization methods like "Kaiming initialization," but it does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We employed a three-layer fully-connected network with a Re LU activation function and batch normalization as the classifier for the aforementioned classification tasks. The networks were trained until the classification loss converges using full batch gradient descent, which ensures the parameters are close to the convergence point required in the theory of Maximum-Margin Bias. We use generators composed of three fully-connected layers followed by three transposed convolution layers, with Re LU activation function and batch normalization. Network parameters were initialized using Kaiming initialization [He et al., 2015] and trained for 50, 000 epochs. The batch size and learning rate were set as hyperparameters and optimized via random search. In order to control the noise in generated images, we also utilized total variation loss in pixel space as a regularization term. |