singleline_models

Install

pip install singleline_models

Models

Sketch RNN

PyTorch implementation of the SketchRNN paper, A Neural Representation of Sketch Drawings.

Sketch RNN learns to reconstruct stroke-based drawings, by predicting a series of strokes. It uses a sequence-to-sequence LSTM model, with gaussian mixture heads to produce a sequence of stroke coordinates.

Datasets

• data/quickdraw/: Sample data from Quick, Draw! Dataset
• data/look/: Custom dataset of single-line drawings by @andrewlook

All data is stored in stroke-3 format, meaning a list with three columns:

• delta_x
• delta_y
• lift_pen (if 1, “lift the pen” and start a new stroke; otherwise 0)

Acknowledgements

• PyTorch Sketch RNN project by Alexis David Jacq
• Annotated Sketch RNN in PyTorch by LabML
• Tensorflow SketchRNN by Magenta Team and David Ha
• sketch-rnn-datasets by David Ha
• SketchRNN-Pytorch by OhataKenji

Improvements

• Log epoch and learning rate
• LR decay
• ETA decay (for KL loss)
• Dropout
• Layer Normalization
• Recurrent Dropout