This repo contains a set of optional, extra datasets for training sketch-rnn, a generative model for vector drawings. You can learn more about the model by reading this blog post or the paper. An open-source TensorFlow implementation of sketch-rnn is available here. Although the datasets had been created in the format customized for training sketch-rnn, it can, and should be used for training newer and better models to advance the state of generative vector image modelling.

We have organized 3 datasets in this repo:

• Aaron Koblin Sheep Market
• Kanji Stroke Data
• Omniglot Stroke Data

In each dataset, each sample is stored as list of coordinate offsets: ∆x, ∆y, and a binary value representing whether the pen is lifted away from the paper. This format, we refer to as stroke-3, is described in Alex Graves' paper on sequence generation. Note that the data format described in the sketch-rnn paper has 5 elements (stroke-5 format), and we need to perform live conversion of the data to the stroke-5 format during mini-batch construction.

We show a single example to illustrate this data format in the figure above. This below-average turtle drawing is sketched by the author. The turtle is represented as a sequence of (∆x, ∆y, binary pen state) points. In the rendered form, the line color corresponds to the sequential stroke ordering to illustrate the ordering.

Each dataset is stored as a python list of examples. Each example is represented as a np.array with np.int16 datatype, and can be easily loaded using the following commands in a Jupyter notebook:

filename = "sketch-rnn-datasets/kanji/short_kanji.npz"
load_data = np.load(filename)
train_set = load_data['train']
valid_set = load_data['valid']
test_set = load_data['test']

print len(train_set)
  Output: 10000
print len(valid_set)
  Output: 600
print len(test_set)
  Output: 500
  
# draw a random example (see draw_strokes.py)
draw_strokes(random.choice(train_set))

For some of the datasets, we also performed simple stroke simplification to preprocess the data, using the Ramer-Douglas-Peucker (RDP) algorithm. We choose the epsilon parameter to be a value between 1.0 to 3.0, depending on how aggressively we want to simply the lines.

If you have a large set of simple SVG images, and want to construct a similar dataset, there are some available libraries to convert subsets of SVGs into line segments. You can then apply RDP on the line segments before converting the data to stroke-3 format. There is some easy-to-use open source code for applying this algorithm here.

Please check the license and copyright information for each dataset in their respective directories.