sketch-pix2seq

This is the reimplementation code of paper Sketch-pix2seq: a Model to Generate Sketches of Multiple Categories.

Input	Generated examples

Requirements

Python 3
Tensorflow (>= 1.4.0)
InkScape or CairoSVG (For vector sketch rendering. Choose one of them is ok.)
```
sudo apt-get install inkscape
# or
pip3 install cairosvg
```

Data Preparations

Follow these steps:

First download the .npz data from QuickDraw dataset. Choose one or more categories as you like.
Place the .npz packages under datasets/npz/ dir.
Modify the hyper params in get_default_hparams() in model.py
- Set the name(s) of the downloaded packages in data_set
- Set the size of the raster image in img_H / img_W
We provide two approaches of rendering sequential data into raster images:
- Using InkScape: this approach is slower but accurate all the time
```
python render_svg2bitmap.py --data_base_dir='datasets' --render_mode='v1'
```
- Using CairoSVG: this approach is faster, but will have one-pixel misalignment problem sometimes (when setting image-width to 256, it will turn out to be 255 sometimes)
```
python render_svg2bitmap.py --data_base_dir='datasets' --render_mode='v2'
```
The raster images can be found under datasets/ dir.

Training

Run this command for default training:

python sketch_pix2seq_train.py

You can also change the settings, e.g. image size, batch size, in get_default_hparams() in model.py. For multi-category training, set the data_set with a list of more than one packages' names.

Training procedure

The following figure shows the reconstruction loss during training within 60k iterations. The orange line belongs to sketch-rnn and the blue one belongs to sketch-pix2seq.

Sampling

With trained model placed under outputs/snapshot/ dir, run this command for sampling:

python sketch_pix2seq_sampling.py

And results in .svg format can be found under outputs/sampling/ dir.

Credits

This code is modified from repo of Sketch-RNN.

sgybupt / sketch-pix2seq