/Sculpture-GAN

3D-DCGAN trained on a corpus of 3D printable objects - as a result, the generations are usually 3D printable

Primary LanguagePython

To Understand the Sculpture-Gan!

  • I just added the comments in Korean about the code without any fixations
  • All codes are from robbiebarrat!
  1. 코드 구성

  • binvox_rw.py -> binvox 파일을 python에서 읽을 수 있게 도와주는 귀여운 친구이다. binvox 파일에 대한 설명은 다음에 있다. -> http://dimatura.net/misc_projects/binvox_rw_py/
    이 코드 역시 저 링크에서 따온듯 하다!

  • generate_dataset.py -> 데이터 파일을 만들어야한다! stl 파일을 binvox 파일로 변환하고 저장하는 역할을 한다. render가 True 면 그림을 그려준다.
    특이 사항은 파이썬 2.7을 사용했다는 점 -> 이 코드를 사용하려면 파이썬 3.x로 수정해주는 것이 좋을 것 같다.

  • Train을 할 때 DCGAN을 사용 -> 모양을 만들 때 Mnist -DCGAN을 사용하였다.

  • DCGAN - Discriminator

  • 4 layer CNN, each depth is increasing by 50, the window is 5,5,5 and the stride is 2,2,2
  • The specific description of one layer is CNN - ReakyRelu(alpha = 0.2) - Dropout
  • The last activation function is sigmoid
  • DCGAN - Generator
  1. Dense (dimdimdim*depth) - batchnormalization - relu - reshape - dropout
  2. one layer consists of sequence that upsample - conv3dTranspose - batchnormalization - relu
  3. 4 layer and end with sigmoid

  • upsample - Repeats the 1st, 2nd and 3rd dimensions of the data by size[0], size[1] and size[2] respectively.

  • Mnist-gan - for training the DCGAN

  • Visualize.py -> 그림을 그려준다.

Sculpture-GAN

3D DCGAN inspired GAN trained on 32x32x32 voxelizations of Thingi10k - a corpus of 10,000 3D printable objects and sculptures - as a result, the generated sculptures are almost always 3D-printable, but usually do not have any real 'meaning', and are just abstract sorts of shapes.

I originally began this project as an attempt at generative architecture; but lack of an appropriate dataset held me back from that. Currently working with trying to get something usable from the google sketchup 3d workshop

Example abstract shape generations

3D-printed generation

1

Computer visualized generations

1 2 3

Usage

There are really only 3 files you need to use to generate your own 3D shapes.

generate_dataset.py

You need to download Thingi10k, and in line 8 of generate_dataset.py, set the variable path_to_dataset to be the path to your raw_meshes folder from your thingi10k download.

Next, just run the file with python generate_dataset.py, and it will start generating numpy arrays that contain all of the now-voxelized models in thingi10k and save them to your data/ folder.

train.py

On line 23 of train.py, point it to the desired (usually the largest) .npy file in your data/ folder by setting numpy_array_saved equal to its path.

Then, you can just run python train.py and trained network checkpoints will start to populate your network_checkpoints/ folder.

visualize.py

Point visualize.py to an array of generations in your generations folder (an array is created every 50 epochs) by setting array_to_visualize equal to its path.

This script will allow you to see the models that network has generated in 3D, with the option to save as PNG, obj, etc.

Issues / Future efforts

Right now, the main issue is the fact that it's very hard to get data for this sort of project; the only two datasets I have encountered are shapnet and thingi10k; currently; i am trying to train a conditional version of the 3d GAN with shapenet so that you can select that you want to generate a table, chair, etc, and the network will move past just generating "shapes"...