Invariant Feature Coding using Tensor Product Representation

Authors: YUSUKE Mukuta, Tatsuya Harada

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The effectiveness of our method is demonstrated on several image datasets. ... In this section, the accuracy and invariance of the proposed method are evaluated using the pretrained features in Section 5.1. In Section 5.2, we evaluate the method on an end-to-end case.
Researcher Affiliation Academia Yusuke Mukuta EMAIL The University of Tokyo, RIKEN
Pseudocode Yes Algorithm 1 Calculation of Invariant PCA ... Algorithm 2 Calculation of invariant k-means
Open Source Code No The pretraining code was implemented using Pytorch (Paszke et al., 2019) with the group equivariant convolution layers implemented using Groupy (Cohen & Welling, 2016). ... The training code was implemented with Pytorch, Groupy and fast-MPN-COV libraries (Li et al., 2018).
Open Datasets Yes These methods were evaluated using the Flickr Material Dataset (FMD) (Sharan et al., 2013), describable texture datasets (DTD) (Cimpoi et al., 2014), UIUC material dataset (UIUC) (Liao et al., 2013), Caltech UCSD Birds (CUB) (Welinder et al., 2010)) and Stanford Cars (Cars) (Krause et al., 2013).
Dataset Splits Yes We used given training test splits for DTD, CUB, and Cars. We randomly split 10 times such that the sizes of the training and testing data would be the same for each category for FMD and UIUC.
Hardware Specification Yes As for computation time, using an Intel Xeon E5-2698v4 x2 20 Core, 2.2 GHz CPU it takes 13 seconds to extract the training features and 61 seconds to learn SVM to train BP on UIUC... when using 8 A100 GPUs it takes 0.17 seconds/batch to train i SQRT-COV (Resnet50)...
Software Dependencies No The pretraining code was implemented using Pytorch (Paszke et al., 2019) with the group equivariant convolution layers implemented using Groupy (Cohen & Welling, 2016). ... The linear SVM implemented in LIBLINEAR (Fan et al., 2008) was used to evaluate the average test accuracy.
Experiment Setup Yes All the models were learned, including the feature extractor, using a momentum grad with an initial learning rate of 0.1, momentum of 0.9, and weight decay rate of 1e-4 for 65 epochs with a batch size of 160. The learning rate was multiplied by 0.1 at 30, 45, and 60 epochs.