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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
MobileODE: An Extra Lightweight Network
Authors: Le Yu, Jun Wu, Bo Gou, Xiangde Min, Lei Zhang, Zhang Yi, Tao He
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across various tasks, including image classification, object detection, and semantic segmentation, demonstrate the effectiveness of the proposed Mobile ODEV1 and Mobile ODEV2 architectures. |
| Researcher Affiliation | Academia | 1Machine Intelligence Laboratory, Sichuan University 2Tongji Medical College, Huazhong University of Science and Technology 1EMAIL 1EMAIL 2mxianade@tih,timu.edu.cn |
| Pseudocode | Yes | Algorithm 1 Channelwise ODE Solver (COS) with DDE |
| Open Source Code | Yes | The code is available at https://github.com/cashily/Mobile ODE. |
| Open Datasets | Yes | We validate our proposed methods on four different datasets, namely CIFAR-10/CIFAR-100 [Krizhevsky et al., 2009], Image Net-R (INR) [Hendrycks et al., 2021a] and Tiny Imagenet (IN-tiny) [Le and Yang, 2015]. ... We integrate Mobile ODE with Deep Lab V3 [Chen, 2017], training Mobile ODE on PASCAL VOC 2012 [Everingham et al., 2015] dataset and ADE20K dataset [Zhou et al., 2019] from scratch... We train and validate our model s detection performance on the publicly available BUSI dataset [Al-Dhabyani et al., 2020] and Facial Feature Extraction (FFE) dataset [kag, 2025]. |
| Dataset Splits | Yes | IN-tiny has 200 classes, and each class contains 500 training images, 50 validation images at a resolution of 64 64. Additionally, we assess the robustness of our methods on IN-R dataset, which consists of 30K images from 200 Image Net classes... We divide IN-R into training and testing sets at a 4 : 1 ratio, using a resolution of 256 256. ... The FFE dataset is a labeled dataset designed for the detection of various facial features... with 457 images allocated for training and 126 images designated for validation. |
| Hardware Specification | Yes | For hardware,We use the Lenovo Legion R9000P laptop, equipped with an NVIDIA Ge Force RTX 4060 GPU (8 GB VRAM), AMD Ryzen 9 7945HX CPU, 15.3 Gi B DDR5 memory. It runs Ubuntu 20.04.4 LTS (Linux 5.15 kernel). |
| Software Dependencies | No | The paper mentions models are implemented in Py Torch and runs on Ubuntu 20.04.4 LTS (Linux 5.15 kernel), but does not provide specific version numbers for PyTorch or other machine learning libraries used for the experiments. |
| Experiment Setup | Yes | The batch size is set to 32 images for CIFAR-10/CIFAR-100, while for IN-tiny and IN-R, it is set to 16 images. For CIFAR-10/CIFAR-100, as suggested in Haase and Amthor [2020], we remove the first and second pooling operations of Mobile Nets to obtain a final feature map of size 4 4. Our experimental setup is consistent with Mehta and Rastegari [2021]. Basic data augmentation techniques, including random resized cropping and horizontal flipping, are applied during training. ... The loss function used cross-entropy with label smoothing (smoothing = 0.1), and optimization was performed with Adam W [Loshchilov, 2017], and a weight decay of 0.01. The learning rate started at 0.0002, increased to 0.002 for the first 2k iterations, and then decayed to 0.0002 using a cosine schedule [Loshchilov and Hutter, 2016]. The model was trained for a maximum of 200 epochs. |