Reproduction Code of House-GAN: Relational Generative Adversarial Networks for Graph-constrained House Layout Generation Original code can be found in the Nelson Nauata's (first author) github page.
Specific goals regarding the reproduced parts will be added soon.
Well, first of all it's a great paper and an interesting read. Secondly, it is a project for the course CS4240 Deep Learning in TU Delft.
- Get the original data from here
- Make sure the dataset is in the correct path.
- The entry point is
train.py. Check that the arguments are correct. - Enjoy your 24h waiting time.
The original author defines one epoch as one pass on the whole dataset and as iteration the pass on a single batch (whose size is adjustable). According to the paper the training was done on 50-70 epochs or 200k iterations.
Run python -m unittest discover test
In order to evaluate diversity, the FID score from https://github.com/mseitzer/pytorch-fid is used. One way to calculate the FID score is to generate on esample for each graph (5k fake) and compare it with the corrresponding GT (5k real). So in order to evaluate all 5 metrics provided, to evaluate fro each group the model should be trained with the data of the other groups. Ex. to evaluate 1-3, we need to train a model on 4-6, 7-9, 10-12, 13+.
- Set correct data path on Argument parser of compute_FID.py
- Set correct checkpoint path
- run compute_FID.py
- use https://github.com/mseitzer/pytorch-fid to calculate FID providing the two folder paths created from compute_FID.py