Reducing Noise in GAN Training with Variance Reduced Extragradient
Authors: Tatjana Chavdarova, Gauthier Gidel, François Fleuret, Simon Lacoste-Julien
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We observe empirically that SVRE performs similarly to a batch method on MNIST while being computationally cheaper, and that SVRE yields more stable GAN training on standard datasets. ... (iv) we test SVRE empirically to train GANs on several standard datasets, and observe that it can improve SOTA deep models in the late stage of their optimization (see 4). |
| Researcher Affiliation | Academia | Tatjana Chavdarova Mila, Université de Montréal Idiap, École Polytechnique Fédérale de Lausanne Gauthier Gidel Mila, Université de Montréal François Fleuret Idiap, École Polytechnique Fédérale de Lausanne Simon Lacoste-Julien Mila, Université de Montréal |
| Pseudocode | Yes | Algorithm 1 Pseudocode for SVRE. |
| Open Source Code | No | The paper does not provide a direct link to open-source code for the methodology, nor does it state that the code is released or available in supplementary materials. |
| Open Datasets | Yes | Datasets. We used the following datasets: (i) MNIST (Lecun and Cortes), (ii) CIFAR-10 (Krizhevsky, 2009, 3), (iii) SVHN (Netzer et al., 2011), and (iv) Image Net ILSVRC 2012 (Russakovsky et al., 2015), using 28 28, 3 32 32, 3 32 32, and 3 64 64 resolution, respectively. |
| Dataset Splits | No | The paper mentions using datasets like MNIST, CIFAR-10, SVHN, and ImageNet, which typically have standard splits. However, it does not explicitly specify the training/test/validation splits used for these datasets, such as percentages or sample counts. For example, it does not mention '80% training, 10% validation, 10% test' or 'standard splits used'. |
| Hardware Specification | No | The paper acknowledges 'Compute Canada for providing the GPUs used for this research'. However, it does not specify the particular models or types of GPUs (e.g., 'NVIDIA A100', 'Tesla V100'), nor any other hardware details such as CPU or memory specifications. |
| Software Dependencies | No | The paper mentions deep learning frameworks like Adam optimizer but does not specify software dependencies with version numbers (e.g., 'PyTorch 1.9', 'TensorFlow 2.0'). |
| Experiment Setup | Yes | We used the DCGAN architectures (Radford et al., 2016), described in F.2.1. ... For clarity, we refer the former as shallow, and the latter as deep architectures. ... We used = 10 2 for SVRE. SE A with = 10 3 achieves similar IS performances as = 10 2 and = 10 4. ... we observed that to stabilize our baseline when using the deep architectures it was required to use 1:5 update ratio of G:D (cf. G.3), whereas for SVRE we used ratio of 1:1. |