Unofficial implementation of PConv paper in TF2.0
Original paper: https://arxiv.org/pdf/1804.07723.pdf
This is a minimalistic implementation with all the code in one ipython notebook that is (hopefully) easy to read and understand.
The code in the notebook uses Places2 dataset to train, although I did not finish training it (that would take way to much time).
Random examples of outputs are shown below. These are not cherry-picked and represent more or less average performance of the model that was trained for about a week on my laptop with GTX1070 with 8Gb of GPU memory and 32 GB of RAM.