sketchrnn_tf2's Introduction

SketchRNN Implementation In Tensorflow 2

This repo contains a simple implementation of the SketchRNN model with Tensorflow 2, following the best practice as much as possible.

Setup

Prerequisite

pyenv
pipenv

Installation

After cloning the repo, run the following commands to setup

pipenv install --dev --skip-lock
pipenv shell

Usage

Training

See the colab notebook for an example train on the rabbit dataset.

Results

Here are some results from a model trained on the rabbit dataset.

See complete examples in the usage.ipynb.

References

A Neural Representation of Sketch Drawings - Original Paper
Teaching Machines to Draw - Official blog post
googlecreativelab/quickdraw-dataset - The QuickDraw dataset
magenta/models/sketch_rnn - Official Implementation
hardmaru/sketch-rnn - Original Implementation from the author of the paper @hardmaru
eyalzk/sketch_rnn_keras - Implementation based on Keras by @eyalzk
zzh8829/yolov3-tf2 - Great source for some TF2 features in practice.

sketchrnn_tf2's People

Contributors

Stargazers

Watchers

sketchrnn_tf2's Issues

Negative loss

Hello, first of all I wanted to thank your for your awesome work. Then, I had a question about the training loss. After several epochs, the reconstruction loss becomes negative. Is it OK to have negative reconstruction loss?
thanks.

Plot with a gap and flipped in both axes

I propose this modification to avoid gap at the start of the path and unnecessary flipping on both axes.

def make_mpl_path(strokes):
    strokes = np.row_stack(([0, 0, 0], strokes))
    vertices = strokes[:, :-1].cumsum(axis=0, dtype=np.float32) 
    vertices[:,1] = vertices[:,1] * -1
<rest of the code>

Thank you for great work!
Best regards
Jacek

Recommend Projects