StrokeNet: A Neural Painting Environment
Authors: Ningyuan Zheng, Yifan Jiang, Dingjiang Huang
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To prove the effectivess of our neural environment, we trained an agent to perform drawing task on several popular datasets, from characters to drawings, with the generator part frozen. For MNIST and Omniglot, we trained an agent to draw the characters within one stroke. We later trained the recurrent Stroke Net on more complex datasets like Quick Draw and Kanji VG (Ofusa et al., 2017). We resized all the input images to 256 256 with anti-alias and paddings. |
| Researcher Affiliation | Academia | Ningyuan Zheng, Yifan Jiang & Dingjiang Huang School of Data Science and Engineering, East China Normal University {10165101164, 10153903133}@stu.ecnu.edu.cn, djhuang@dase.ecnu.edu.cn |
| Pseudocode | No | The paper describes the architecture and processes in detail, but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code for the model at: https://github.com/vexilligera/strokenet. Code for the web application available at: https://github.com/vexilligera/drawwebapp. |
| Open Datasets | Yes | For the MNIST (Le Cun & Cortes, 2010) digits, we evaluated the quality of the agent with a classifier trained solely on the original MNIST dataset, and tested the classifier on generated images. We also compared our method with others to show the efficiency. We explored the latent space of the agent as well...For MNIST and Omniglot, we trained an agent to draw the characters within one stroke. We later trained the recurrent Stroke Net on more complex datasets like Quick Draw and Kanji VG (Ofusa et al., 2017). |
| Dataset Splits | No | The paper mentions training, testing, and evaluation, but does not explicitly provide specific dataset split information (e.g., percentages or counts for train/validation/test splits). |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA Tesla P40 GPU. |
| Software Dependencies | No | The paper mentions building a painting software using Java Script and Web GL, but does not provide specific version numbers for these or other ancillary software components. |
| Experiment Setup | Yes | It can be found that smaller batch size results in more accurate images. We trained the generator with a batch size of 64 until the loss no longer improves. We then set the batch size to 32 to sharpen the neural network. To train the agent, we freeze the generator. Denote the agent loss as lagent, the generated image and ground-truth image as igen and igt respectively, the loss is defined as: lagent = igen igt 2 2 + λ n 1 k=1 Pk Pk+1 2 2 , (9) where Pk = [xk, yk, pk]T is the data describing the kth anchor point on the stroke. Here the summation term constrains the average distance between neighbouring points, where λ denotes the penalty strength...For (c) and (d) the learning rate is set to 10 4, batch size equals to 64. |