SketchODE: Learning neural sketch representation in continuous time

Authors: Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, Yi-Zhe Song

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our models are validated on several datasets including Vector MNIST, Di Di and Quick, Draw!.
Researcher Affiliation Collaboration Ayan Das1,2, Yongxin Yang1,3, Timothy Hospedales1,3, Tao Xiang1,2 & Yi-Zhe Song1,2 1Sketch X, CVSSP, University of Surrey, UK 2i Fly Tek-Surrey Joint Research Centre on Artificial Intelligence 3School of Informatics, University of Edinburgh, UK
Pseudocode No The paper describes its methods using prose and mathematical equations, but does not include explicit pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or link to its own open-source code for the methodology described.
Open Datasets Yes Quick Draw:1 Released by Ha & Eck (2018), Quick, Draw! is the largest collection of publicly available free-hand doodling dataset. ... 1https://github.com/googlecreativelab/quickdraw-dataset. Di Di Dataset2: Released by Gervais et al. (2020), Digital Ink Diagram dataset is a large collection of synthetically generated flowcharts ... 2https://github.com/google-research/google-research/tree/master/didi dataset
Dataset Splits No We use a 80% 20% train-test split for each. While the paper shows "Validation error" in Figure 5, it does not explicitly provide details about a specific validation dataset split (e.g., as a percentage or count) distinct from the training and test sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running the experiments.
Software Dependencies No The paper mentions using "Adam W (Loshchilov & Hutter, 2019) optimizer" and "RK4 (The classic Runge-Kutta method)" but does not specify software versions for any libraries, frameworks, or programming languages.
Experiment Setup Yes The dimensionality of the hidden state in CDE/ODE is denoted d and the latent dimension as l. For experimentation, we choose d = 64, l = 48 for Vector MNIST, d = 96, l = 84 for Di Di and d = 144, l = 96 for Quick, Draw!. Dynamics networks (for both CDE and ODE) are MLPs with few (2 for full-sequence, 1 for multi-stroke) hidden layers of size 1.2d . ... We empirically found the best value of the upper limit of time range to be T = 5. All models are trained with Adam W (Loshchilov & Hutter, 2019) optimizer, learning rate of 1 × 10−4 for RNN and 3 × 10−3 for CDE-ODE, annealed using cosine scheduler (100 epoch period) with decreasing (by factor of 4) amplitude. We also used 10−2 as weight decay (L2 penalty) regularizer on the weight of the dynamics networks and gradients are clipped at a magnitude of 0.01.