pip install .
Download dataset here.
civl5220-group13 housegan visualize_dataset valid_data.npy
We have nodes (room types) and the boxes (bounding boxes of the room), we don't have the connection information, hence we need to extract the edges of the graph.
civl5220-group13 housegan extract_edges valid_data.npy
Download model here.
civl5220-group13 housegan inference valid_data.npy --ckpt exp_demo_D_500000.pth
civl5220-group13 housegan constraint_inference valid_data.npy --ckpt exp_demo_D_500000.pth --constraint-mask toy_masks/0.txt
civl5220-group13 housegan animate snapshots/0.pkl
Toggle to show old stuff
Clone the houseganpp repo,
git clone https://github.com/ennauata/houseganpp
or clone this repo recursively.
git clone --recurse-submodules https://github.com/enhuiz/civl5220-group13
Then go to the houseganpp folder, install requirements
pip intall -r requirements.txt
(or just run it and install what is missing, version seems not important for houseganpp tested on nov. 6, 2021)
After this, you need to change the path in data/sample_list.txt like this:
-/home/nelson/Workspace/autodesk/autodesk/rplan_2/json_data_update/7513.json
+data/json/7513.json
Run python test.py, you will get the dump folder, which contains something like:
Clone the housegan repo,
git clone https://github.com/ennauata/housegan
or clone this repo recursively.
git clone --recurse-submodules https://github.com/enhuiz/civl5220-group13
For HouseGAN, you actually need to download the dataset and pretrained model according to their document (they are not contained in the repo :|). Requirements are also needed to be installed, latest version seemed fine, except for networkx==2.4 (nov. 6, 2021).
After downloading the model and processed dataset (yes, for this one they provide a processed dataset, train_data.npy and valid_data.npy), modify those lines of variation_bbs_with_target_graph_segments_suppl.py since they named their experiment as exp_demo and run for 500_000 iterations.
-numb_iters = 200000
-exp_name = 'exp_with_graph_global_new'
+numb_iters = 500000 # change the iters, to make the ckpt path right
+exp_name = 'exp_demo' # change the name
target_set = 'D'
phase='eval'
checkpoint = './checkpoints/{}_{}_{}.pth'.format(exp_name, target_set, numb_iters)
@@ -177,7 +177,7 @@ generator.load_state_dict(torch.load(checkpoint))
cuda = True if torch.cuda.is_available() else False
if cuda:
generator.cuda()
-rooms_path = '/local-scratch4/nnauata/autodesk/FloorplanDataset/'
+rooms_path = './FloorplanDataset/' # i put those xxx_data.npy here
Then run python variation_bbs_with_target_graph_segments_suppl.py, this perhaps needs cuda, you got a folder output, with output like this:
By default, consider HouseGAN++.
- Bubble diagram: A graph, whose nodes represent room type (including the room "outside") and edge represent doors.
- RPLAN dataset: A dataset from USTC, used by HouseGAN++: http://staff.ustc.edu.cn/~fuxm/projects/DeepLayout/index.html
- Raster-to-Vector, Vectorization: The algorithm used to convert raster images to vector graphics: http://art-programmer.github.io/floorplan-transformation.html
- Room Mask: 64x64 (32x32 in HouseGAN) signle-channel segment map (like those in instance segmentation), foreground (value > 0, perhaps) means this pixel is within the region of given room. The final generation result you see from the website and paper is a aggregation of the maps (i.e. room masks) of all input rooms.
An example training data from: 3rdparty/houseganpp/data/json/7513.json
{
// 14 rooms, each has a type
"room_type": [3, 3, 2, 5, 4, 5, 1, 17, 17, 17, 17, 17, 17, 15],
// 14 box coord, range from [0, 256) (exclusive?)
"boxes": [
[51.0, 56.0, 96.0, 107.0],
[78.0, 113.0, 124.0, 168.0],
[159.0, 56.0, 205.0, 78.0],
[192.0, 84.0, 205.0, 107.0],
[159.0, 84.0, 186.0, 107.0],
[130.0, 183.0, 186.0, 200.0],
[102.0, 56.0, 186.0, 177.0],
[195.0, 79.0, 205.0, 83.0],
[114.0, 108.0, 126.0, 112.0],
[97.0, 96.0, 101.0, 105.0],
[154.0, 66.0, 158.0, 76.0],
[154.0, 96.0, 158.0, 108.0],
[137.0, 178.0, 179.0, 182.0],
[137.0, 51.0, 149.0, 55.0]
],
// box edge coord, together with room type and ?
"edges": [
[51.0, 56.0, 51.0, 107.0, 3, 0],
[51.0, 107.0, 96.0, 107.0, 3, 0],
[96.0, 107.0, 96.0, 56.0, 3, 1],
[96.0, 56.0, 51.0, 56.0, 3, 0],
[78.0, 113.0, 78.0, 168.0, 3, 0],
[78.0, 168.0, 124.0, 168.0, 3, 0],
[124.0, 168.0, 124.0, 113.0, 3, 0],
[124.0, 113.0, 78.0, 113.0, 3, 1],
[159.0, 56.0, 159.0, 78.0, 2, 1],
[159.0, 78.0, 205.0, 78.0, 2, 5],
[205.0, 78.0, 205.0, 56.0, 2, 0],
[205.0, 56.0, 159.0, 56.0, 2, 0],
[192.0, 84.0, 192.0, 107.0, 5, 0],
[192.0, 107.0, 205.0, 107.0, 5, 0],
[205.0, 107.0, 205.0, 84.0, 5, 0],
[205.0, 84.0, 192.0, 84.0, 5, 2],
[159.0, 84.0, 159.0, 107.0, 4, 1],
[159.0, 107.0, 186.0, 107.0, 4, 0],
[186.0, 107.0, 186.0, 84.0, 4, 0],
[186.0, 84.0, 159.0, 84.0, 4, 0],
[130.0, 183.0, 130.0, 200.0, 5, 0],
[130.0, 200.0, 186.0, 200.0, 5, 0],
[186.0, 200.0, 186.0, 183.0, 5, 0],
[186.0, 183.0, 130.0, 183.0, 5, 1],
[102.0, 56.0, 102.0, 107.0, 1, 3],
[102.0, 107.0, 130.0, 107.0, 1, 3],
[130.0, 107.0, 130.0, 177.0, 1, 0],
[130.0, 177.0, 186.0, 177.0, 1, 5],
[186.0, 177.0, 186.0, 113.0, 1, 0],
[186.0, 113.0, 153.0, 113.0, 1, 0],
[153.0, 113.0, 153.0, 56.0, 1, 4],
[153.0, 56.0, 102.0, 56.0, 1, 0],
[195.0, 79.0, 195.0, 83.0, 17, 0],
[195.0, 83.0, 205.0, 83.0, 17, 5],
[205.0, 83.0, 205.0, 79.0, 17, 0],
[205.0, 79.0, 195.0, 79.0, 17, 2],
[114.0, 108.0, 114.0, 112.0, 17, 0],
[114.0, 112.0, 126.0, 112.0, 17, 3],
[126.0, 112.0, 126.0, 108.0, 17, 0],
[126.0, 108.0, 114.0, 108.0, 17, 1],
[97.0, 96.0, 97.0, 105.0, 17, 3],
[97.0, 105.0, 101.0, 105.0, 17, 0],
[101.0, 105.0, 101.0, 96.0, 17, 1],
[101.0, 96.0, 97.0, 96.0, 17, 0],
[154.0, 66.0, 154.0, 76.0, 17, 1],
[154.0, 76.0, 158.0, 76.0, 17, 0],
[158.0, 76.0, 158.0, 66.0, 17, 2],
[158.0, 66.0, 154.0, 66.0, 17, 0],
[154.0, 96.0, 154.0, 108.0, 17, 1],
[154.0, 108.0, 158.0, 108.0, 17, 0],
[158.0, 108.0, 158.0, 96.0, 17, 4],
[158.0, 96.0, 154.0, 96.0, 17, 0],
[137.0, 178.0, 137.0, 182.0, 17, 0],
[137.0, 182.0, 179.0, 182.0, 17, 5],
[179.0, 182.0, 179.0, 178.0, 17, 0],
[179.0, 178.0, 137.0, 178.0, 17, 1],
[137.0, 51.0, 137.0, 55.0, 15, 0],
[137.0, 55.0, 149.0, 55.0, 15, 1],
[149.0, 55.0, 149.0, 51.0, 15, 0],
[149.0, 51.0, 137.0, 51.0, 15, 0]
],
// ?
"ed_rm": [
[0],
[0],
[0, 6],
[0],
[1],
[1],
[1],
[1, 6],
[2, 6],
[2, 3],
[2],
[2],
[3],
[3],
[3],
[3, 2],
[4, 6],
[4],
[4],
[4],
[5],
[5],
[5],
[5, 6],
[6, 0],
[6, 1],
[6],
[6, 5],
[6],
[6],
[6, 4],
[6],
[7],
[7, 3],
[7],
[7, 2],
[8],
[8, 1],
[8],
[8, 6],
[9, 0],
[9],
[9, 6],
[9],
[10, 6],
[10],
[10, 2],
[10],
[11, 6],
[11],
[11, 4],
[11],
[12],
[12, 5],
[12],
[12, 6],
[13],
[13, 6],
[13],
[13]
]
}Each room has a room type.
According to here, room types are:
{
'living_room': 1,
'kitchen': 2,
'bedroom': 3,
'bathroom': 4,
'balcony': 5,
'entrance': 6,
'dining room': 7,
'study room': 8,
'storage': 10,
'front door': 15, # the door connects room and outside
'unknown': 16,
'interior_door': 17 # the door connects rooms
}🤔, some numbers are somehow skipped, there is only 12 classes, including 10 room types and 2 door types.
Each room has a axes-aligned (i.e. either horizontal or vertical) box, represented by pixels within 256x256 canvas.
By drawing the boxes from the above sample on a 256x256 canvas, we can get:
.
├── 3rdparty
│ ├── housegan # ennauata's housegan, we may take some code from it
│ └── houseganpp # ennauata's newer version
├── civl5220_group13 # our code, python only, I'm gonna start from here
├── overleaf # put our overleaf documents
└── README.md
- For Python formatter, I prefer black, it is easy to use and consistent.