Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
First-order Convergence Theory for Weakly-Convex-Weakly-Concave Min-max Problems
Authors: Mingrui Liu, Hassan Rafique, Qihang Lin, Tianbao Yang
JMLR 2021 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments verify the convergence theory and also demonstrate the effectiveness of the proposed methods on training GANs. |
| Researcher Affiliation | Academia | Mingrui Liu EMAIL Department of Computer Science The University of Iowa Iowa City, IA, 52242, USA Hassan Rafique EMAIL Department of Mathematics The University of Iowa Iowa City, IA, 52242, USA Qihang Lin EMAIL Business Analytics Department The University of Iowa Iowa City, IA, 52242, USA Tianbao Yang EMAIL Department of Computer Science The University of Iowa Iowa City, IA, 52242, USA |
| Pseudocode | Yes | Algorithm 1 Inexact Proximal Point (IPP) Method for Weakly-Monotone SVI; Algorithm 2 Stochastic Subgradient Method for SVI(F, Z): SG(F, Z, z(0), η, T); Algorithm 3 GD(F, Z, z(0), η, T); Algorithm 4 EG(F, Z, w(0), η, T); Algorithm 5 SVRG Method for SVI(F, Z): SVRG(F, Z, w(0), T); Algorithm 6 Nesterov s Accelerated Method for SVI(F, Z): NA(F, Z, w(0), T) |
| Open Source Code | No | The paper does not provide an explicit statement of code release or a link to a code repository. It mentions using existing neural network structures from original implementations for GANs. |
| Open Datasets | Yes | We conduct the experiments using the same neural network structures for the discriminator and generator as in the original implementations on two datasets (CIFAR10 and LSUN Bedroom). |
| Dataset Splits | No | The paper does not explicitly provide details about training/test/validation splits. While it mentions datasets like CIFAR10 and LSUN Bedroom which have standard splits, it does not specify which splits were used in their experiments. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run its experiments, such as GPU models, CPU types, or memory. |
| Software Dependencies | No | The paper does not provide specific version numbers for key software components. It mentions activation functions like ReLU and ELU, but not the software environment or libraries with versions. |
| Experiment Setup | Yes | The step size of SGD and SGD (5d) is set as η0/√t with t being the number of performed updates as in standard theory for convex-concave problems. For IPP-SGD, the step size and the number of inner iterations at k-th stage are set to be η0/(k + 1) and T0(k + 1)2 respectively, where T0 is tuned from {5000 : 5000 : 60000}. In our experiment on CIFAR10 data, the tuned initial stepsize is 0.1 for WGAN (ReLU and ELU) and 0.01 for WGAN-GP (ReLU). For Adam(5d), Adam and IPP-Adam, we choose the initial stepsize according to the original implementations for WGAN and WGAN-GP. For IPP-Adam, the stepsize and the number of passes of CIFAR10 data (with batchsize 64) at k-th stage are set to be η 0/(k + 1) and c respectively, where η 0 is the initial stepsize for Adam (5d) and c is tuned from {50 : 50 : 400}. Note that the number of iterations for each stage is T = 50000c /64. For our proposed methods (IPP-Adam and IPP-SGD), we tune the weakly monotone parameter from {10−3, 10−4}. |