SinIR: Efficient General Image Manipulation with Single Image Reconstruction
Authors: Jihyeong Yoo, Qifeng Chen
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | With quantitative evaluation, we show that Sin IR has competitive performance on various image manipulation tasks. Moreover, with a much simpler training objective (i.e., re construction), Sin IR is trained 33.5 times faster than Sin GAN (for 500 500 images) that solves similar tasks. Our code is publicly available at github.com/Yoo Ji Hyeong/Sin IR. |
| Researcher Affiliation | Academia | Jihyeong Yoo 1 Qifeng Chen 1 1Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong. Correspondence to: Qifeng Chen <cqf@ust.hk>. |
| Pseudocode | No | The paper describes the training process and model architecture in text and with diagrams (e.g., Figure 2) but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is publicly available at github.com/Yoo Ji Hyeong/Sin IR. |
| Open Datasets | Yes | Table 3 shows 4X super-resolution scores using BSD100 dataset (Martin et al., 2001). ... We conducted a user study using randomly sam pled images from a dedicated dataset provided by (Luan et al., 2017). ... We conducted a user study using randomly sampled images from a dedi cated dataset provided by (Luan et al., 2018). |
| Dataset Splits | No | The paper uses established datasets like BSD100 and datasets from cited works, but it does not explicitly state the specific training, validation, and test splits used for these datasets within the paper. It only mentions evaluating on them. |
| Hardware Specification | Yes | All experiments are conducted on the same machine with a single NVIDIA RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions using the Adam optimizer (Kingma & Ba, 2015) and specifies its hyperparameters (β1 = 0.5, β2 = 0.999), but it does not list any other specific software dependencies or libraries with version numbers (e.g., PyTorch, TensorFlow, Python version). |
| Experiment Setup | Yes | We use the Adam optimizer (β1 = 0.5, β2 = 0.999) (Kingma & Ba, 2015) and the learning rate is 1e-4, unless mentioned otherwise. For the reconstruction loss, we use MSE (mean squared error) loss. However, it is well-known that MSE loss produces blurry images (Zhao et al., 2017), thus we combine MSE loss and SSIM (structural similarity) loss (Wang et al., 2004). ... Sin IR is trained for 500 iterations at every scale, whereas Sin GAN is trained for 6,000 itera tions at every scale following the authors best practice. ... The percent age of random pixel shuffing is set to 5e-4. ... For super-resolution, we ... set the number of kernels to 256, the iteration number at each scale to 1000 (2000 iterations in total), and the learning rate to 0.001 so that the overall model can still be trained suffciently with enough capacity. |