One-dimensional Path Convolution
Authors: Xuanshu Luo, Martin Werner
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results further validate the efficacy of our method, establishing the proposed 1D Path Conv as a viable backbone for efficient vision models. The paper includes sections such as '5. Experiments' and '5.3. The impact of path selection', detailing evaluations on datasets like CIFAR-10, SVHN, and Image Net-64, and presenting performance metrics like top-1 accuracy in tables (e.g., Table 3, Table 4, Table 5) and ablation studies. |
| Researcher Affiliation | Academia | 1Professorship of Big Geospatial Data Management, Technical University of Munich, Germany. Correspondence to: Xuanshu Luo <EMAIL>, Martin Werner <EMAIL>. |
| Pseudocode | No | The paper describes methods and model architectures in textual form and through diagrams (e.g., Figure 5 for overall architecture, Figure 7 for building blocks), but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper mentions implementing a CUDA kernel ('We implement a CUDA kernel to perform fast image traversal by leveraging NVIDIA GPU parallelism, demonstrating up to 73x acceleration compared to a single-thread CPU implementation. We provide more details in Appendix E.') but does not explicitly state that the source code for the entire methodology or the CUDA kernel itself is open-source, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate Path Conv models using three datasets: CIFAR-10 (Krizhevsky et al., 2009), SVHN (Netzer et al., 2011), and Image Net-64 (Chrabaszcz et al., 2017), a downsampled version of Image Net-1K (Russakovsky et al., 2015) with the same image count and dataset splits with a resolution of 64 64 across 1000 classes. |
| Dataset Splits | Yes | We evaluate Path Conv models using three datasets: CIFAR-10 (Krizhevsky et al., 2009), SVHN (Netzer et al., 2011), and Image Net-64 (Chrabaszcz et al., 2017), a downsampled version of Image Net-1K (Russakovsky et al., 2015) with the same image count and dataset splits with a resolution of 64 64 across 1000 classes. |
| Hardware Specification | Yes | Table 7 compares runtimes between our CUDA kernel implementation and CPU version for varying batch sizes, s, and the number of paths P on 1 NVIDIA A100 GPU. |
| Software Dependencies | No | The paper mentions various optimizers (Adam W, SGD), data augmentation techniques (Rand Augment, Mixup, Cut Mix), and regularization methods (Stochastic depth, label smoothing) along with their respective citations, but it does not provide specific version numbers for any programming languages, libraries, or frameworks used (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | Training. Path Conv models employs the Adam W optimizer (Loshchilov & Hutter, 2019), while Res Net models utilize stochastic gradient descent with Nesterov momentum (Sutskever et al., 2013). All models are trained for 300 epochs with a 10-epoch linear warmup. The learning rate follows the cosine annealing schedule (Loshchilov & Hutter, 2017). We adopt Rand Augment (Cubuk et al., 2020) for data augmentation. For Image Net-64, we additionally apply Mixup (Zhang et al., 2018) and Cut Mix (Yun et al., 2019). Stochastic depth (Huang et al., 2016) and label smoothing (Szegedy et al., 2016) are also used for regularization. More experimental settings are provided in Appendix F. Tables 8 and 9 in Appendix F provide detailed training settings including initial learning rate, weight decay, optimizer, batch size, scheduler, and warmup parameters. |