Compact Generalized Non-local Network

Authors: Kaiyu Yue, Ming Sun, Yuchen Yuan, Feng Zhou, Errui Ding, Fuxin Xu

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results illustrate the clear-cut improvements and practical applicability of the generalized non-local module on both fine-grained object recognition and video classification.
Researcher Affiliation Collaboration Kaiyu Yue , Ming Sun Yuchen Yuan Feng Zhou Errui Ding Fuxin Xu Baidu VIS Baidu Research Central South University {yuekaiyu, sunming05, yuanyuchen02, zhoufeng09, dingerrui}@baidu.com fxxu@csu.edu.cn
Pseudocode No The paper provides workflow diagrams and formulas in figures (Fig 2, 3, 4), but these are not structured pseudocode or algorithm blocks. No section is explicitly labeled 'Algorithm' or 'Pseudocode'.
Open Source Code Yes Code is available at: https://github.com/Kaiyu Yue/cgnl-network.pytorch.
Open Datasets Yes We evaluate the CGNL network on multiple tasks, including fine-grained classification and action recognition. For fine-grained classification, we experiment on the Birds-200-2011 (CUB) dataset [25]... For action recognition, we experiment on two challenging datasets, Mini-Kinetics [30] and UCF101 [22]. ... We use the models pretrained on Image Net [20] to initialize the weights.
Dataset Splits Yes The Mini-Kinetics dataset... we use 78265 videos for training and 4986 videos for validation. ... Image Net [20] dataset, which has 1.2 million training images and 50000 images for validation in 1000 object categories.
Hardware Specification No The paper does not mention any specific hardware components (e.g., GPU models, CPU types, or memory specifications) used for running the experiments.
Software Dependencies No The paper mentions various neural network architectures and components like 'Res Net', 'Mask R-CNN', 'Batch Normalization', and 'dropout', but it does not specify any software names with version numbers (e.g., 'PyTorch 1.9', 'TensorFlow 2.x').
Experiment Setup Yes We use a weight decay of 0.0001 and momentum of 0.9 in default. The strategy of gradual warmup is used in the first ten epochs. The dropout [23] with ratio 0.5 is inserted between average pooling layer and last fully-connected layer. To keep same with [27], we use zero to initialize the weight and bias of the Batch Norm (BN) layer in both CGNL and NL blocks [6]. To train the networks on CUB dataset, we follow the same training strategy above but the final crop size of 448.