Regularized Frank-Wolfe for Dense CRFs: Generalizing Mean Field and Beyond

Authors: Đ.Khuê Lê-Huu, Karteek Alahari

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate this in our empirical results on standard semantic segmentation datasets, where several instantiations of our regularized Frank-Wolfe outperform mean field inference, both as a standalone component and as an end-to-end trainable layer in a neural network.
Researcher Affiliation Academia Ð.Khuê Lê-Huu Karteek Alahari Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK 38000 Grenoble, France {khue.le,karteek.alahari}@inria.fr
Pseudocode Yes Algorithm 1 Generic regularized Frank-Wolfe for (approximately) solving MAP inference (6).
Open Source Code Yes Our source code is made publicly available under the GNU general public license for this purpose.1 1https://github.com/netw0rkf10w/CRF
Open Datasets Yes We first pretrain Deep Labv3 and Deep Labv3+ on the COCO dataset [46] and then finetune them on PASCAL VOC (trainaug) and Cityscapes (train) to obtain similar results to previous work [16, 17] (Table 1, CNN column).
Dataset Splits Yes We first pretrain Deep Labv3 and Deep Labv3+ on the COCO dataset [46] and then finetune them on PASCAL VOC (trainaug) and Cityscapes (train) to obtain similar results to previous work [16, 17] (Table 1, CNN column). ...Table 1 shows the performance on the validation sets of PASCAL VOC and Cityscapes...
Hardware Specification No The paper states: 'The experiments were performed using HPC resources from GENCI-IDRIS (Grants 2020-AD011011321 and 2020AD011011881).' However, it does not specify concrete hardware details such as specific GPU or CPU models, memory sizes, or detailed cloud instance types used for the experiments.
Software Dependencies Yes Our implementation builds on PyTorch 1.7.0 and mmsegmentation [2].
Experiment Setup Yes We train the model for 20 epochs with 5 CRF iterations, using the same poly schedule as before. ...We set its learning rate to a small value of 0.0001. For the CRF, we tried 4 different values of initial learning rates 0 2 {1.0, 0.1, 0.01, 0.001}...