Information-Theoretic GAN Compression with Variational Energy-based Model
Authors: Minsoo Kang, Hyewon Yoo, Eunhee Kang, Sehwan Ki, Hyong Euk Lee, Bohyung Han
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on the standard datasets using various architectures. Since VEM is a model-agnostic algorithm component, we incorporate the module into existing GAN compression approaches and the combined models are expressed in the form of [Algorithm Name] + VEM throughout this section. |
| Researcher Affiliation | Collaboration | Minsoo Kang1 Hyewon Yoo2 Eunhee Kang3 Sehwan Ki3 Hyong-Euk Lee3 Bohyung Han1,2 1ECE & 2IPAI, Seoul National University 3Samsung Advanced Institute of Technology (SAIT) |
| Pseudocode | No | The paper describes procedures in text, but no formal pseudocode or algorithm blocks are provided. |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We attach the code and instructions in the supplementary material. |
| Open Datasets | Yes | Image-to-image translation We adopt Pix2Pix [6] on the Edges Shoes [67] and Cityscapes [68] datasets while taking Cycle GAN [7] on Horse Zebra [7] and Summer Winter [7]. ... Image generation To evaluate the performance of our approach for unconditional GANs, we run experiments with Style GAN2 [66] on Flickr-Faces-HQ (FFHQ) [71] and Self-Attention GAN (SAGAN) [4] on Celeb A [72]. |
| Dataset Splits | Yes | Following the original models [6, 7], we employ generators based on U-Net [69] and Res Net [70] for Pix2Pix and Cycle GAN, respectively. For all the datasets, images are resized to 256 256 before feeding them into the models. |
| Hardware Specification | No | We did not include them. |
| Software Dependencies | No | The proposed algorithm is implemented in Py Torch [76] based on the publicly available code1. |
| Experiment Setup | Yes | Specifically, we set the batch size to 4 for Pix2Pix, 1 for Cycle GAN, 64 for SAGAN, and 16 for Style GAN2 while we adopt the Adam optimizer with an initial learning rate of 0.0002, which decays to zero linearly. When we train our energy-based model, we employ the Adam optimizer with a learning rate of 0.0001. Note that we only run 10 steps of Langevin dynamics to draw samples from the energy-based model to save training time. We set the standard deviation of random noise to 0.005 and use a step size of 100 for each gradient step of Langevin dynamics for all datasets except for Cityscapes and FFHQ. For these two datasets, we use a step size of 50 instead. Details about the hyperparameter settings are provided in the supplementary document. |