This Looks Like It Rather Than That: ProtoKNN For Similarity-Based Classifiers
Authors: Yuki Ukai, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on multiple open datasets demonstrate that the proposed method can achieve competitive results with a stateof-the-art method. We conducted experiments on three public datasets: CUB200-2011 (Wah, 2011), Stanford Dogs (Khosla, 2011), and Stanford Cars (Krause, 2013). Tables 1 and 2 show the top-1 accuracy of the proposed method with various model backbones on CUB200-2011 and Stanford Dogs, respectively. To examine the effectiveness of each component of the proposed method, we conducted an ablation study on CUB200-2011 with the full images. |
| Researcher Affiliation | Collaboration | Yuki Ukai 1,2, Tsubasa Hirakawa2, Takayoshi Yamashita2, and Hironobu Fujiyoshi2 1GLORY Ltd., 2Chubu University |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. Algorithms are described in textual paragraphs and mathematical equations. |
| Open Source Code | No | The paper does not provide a direct link to open-source code or an explicit statement of code availability for the described methodology. |
| Open Datasets | Yes | We conducted experiments on three public datasets: CUB200-2011 (Wah, 2011), Stanford Dogs (Khosla, 2011), and Stanford Cars (Krause, 2013). |
| Dataset Splits | No | The paper does not specify explicit training, validation, and test dataset splits by percentages or counts. It mentions 'minibatch' and 'training' and 'testing' phases, but no specific split ratios or methods for creating validation sets are provided. |
| Hardware Specification | Yes | Moreover, the training time of our method (Resnet 50 on the CUB200-2011 dataset) is nearly four hours, which is almost the same as Proto Pool in our experimental settings (one RTX3090 GPU). |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer' and conducting experiments based on 'the implementation of Roth (2020)', but it does not specify version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used. |
| Experiment Setup | Yes | During the training, we used the Adam optimizer and set the learning rate to 1e-5 for the model backbone and to 1e-3 for the other layers. For data augmentation, we used Random Perspective, Color Jitter, Random Horizontal Flip, Random Affine, and Random Crop following Proto Tree (Nauta, 2021). When cropped images were used, we followed previous work (Rymarczyk, 2022) and set the number of prototypes to 195 for Stanford Cars and to 202 for CUB200-2011. We also set the number of prototypes to 512 on all datasets when we used the full images. The number of epochs was set to 140, 140, and 60 for the experiments on Stanford Cars, CUB200-2011, and Stanford Dogs, respectively. We constructed the minibatch so that it contained 56 classes and two images per class. We empirically set λclst = 0.1 and λaux = 1.0 in this paper. |