Revisiting Sparse Convolutional Model for Visual Recognition
Authors: xili dai, Mingyang Li, Pengyuan Zhai, Shengbang Tong, Xingjian Gao, Shao-Lun Huang, Zhihui Zhu, Chong You, Yi Ma
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we provide experimental evidence for neural networks with CSC-layers as discussed in Sec. 3. Through experiments on CIFAR-10, CIFAR-100, and Image Net, Sec. 4.1 shows that our networks have equally competitive classification performance as mainstream architectures such as Res Net. Furthermore, we show in Sec. 4.2 that our network is able to handle input perturbations with the robust inference technique. Finally, we demonstrate in Sec. 4.3 that our network is able to handle adversarial perturbations as well. |
| Researcher Affiliation | Academia | 1The Hong Kong University of Science and Technology (Guangzhou) 2Tsinghua-Berkeley Shenzhen Institute (TBSI), Tsinghua University 3 Harvard University 4 University of California, Berkeley 5 Ohio State University |
| Pseudocode | Yes | Algorithm 1 Robust inference with neural networks constructed from CSC-layers |
| Open Source Code | Yes | Source code can be found at https://github.com/Delay-Xili/SDNet. |
| Open Datasets | Yes | Datasets. We test the performance of our method using the CIFAR-10 and CIFAR-100 [51] datasets... CIFAR-10 and CIFAR-100 are available at https://www.cs.toronto.edu/~kriz/cifar.html. We also use the Image Net dataset [52] that contains 1,000 classes and a total number of around 1 million images. Image Net is a publicly available dataset: https://www.image-net.org |
| Dataset Splits | No | Each dataset contains 50,000 training images and 10,000 testing images where each image is of size 32 32 with RGB channels. The paper explicitly mentions training and testing set sizes for CIFAR-10/100 but does not specify a validation set split. |
| Hardware Specification | Yes | All the experiments are conducted on a single NVIDIA GTX 2080Ti GPU. For Image Net, ... All experiments on Image Net are conducted on 4 NVIDIA RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions 'pytorch-cifar' indirectly by linking to a GitHub repository for ResNet implementation, but it does not specify any software names with version numbers for reproducibility. |
| Experiment Setup | Yes | Network training. For CIFAR-10 and CIFAR-100, we use a cosine learning rate decay schedule with an initial learning rate of 0.1, and train the model for 220 epochs. We use the SGD optimizer with 0.9 momentum and Nestrov. The weight decay is set to 5e-4, and batch size is set to 128... For Image Net, we use multi-step learning rate decay schedule with an initial learning rate of 0.1 decayed by a factor of 0.1 at the 30th, 60th, and 90th epochs, and train the model for 100 epochs. The batch size is 512, and the optimizer is SGD with 0.9 momentum and Nestrov. Forward propagation through each CSC-layer is performed via unrolling two iterations of FISTA. |