[Models: conf.pth] Wrong `data path` in `conf.pth`-files of the provided models

[GregorKobsik]

Description:

The provided conf.pth-files seem inconsistent with the provided dataset (cmp. Example below). The dataloader PartGraphShapesDataset uses the provided argument data_path for self.data_dir. Later, in the function __getitem__(...) the geometric data geo_data is loaded given the provided data path. Unfortunately, all .npz-files are located in Chair_dgeo directory and not in the Chair_dhier directory.

Is my assumption correct, that we need to manually update the configuration files to make it consistent with the provided dataset (or vice versa)?

Example (exp_ae_Chair):

{'exp_name': 'PartNet_ae_Chair_4000_ccc',
 'category': 'Chair',
 'device': 'cuda:0',
 'seed': 3124256514,
 'model_path': '../data/models',
 'data_path': '../data/partnetdata/Chair_dhier',     <-- inconsistent with dataset, no ".npz"-files
 'floor_data_path': '',
 'train_dataset': 'chair.txt',
...
 'max_part_per_parent': 10,
 'X': 12,
 'Y': 2}

https://github.com/IGLICT/DSG-Net/blob/6c6d117525686fcca14b3efe477dfdb37737b55a/code/datav1.py#L464-L465

[GregorKobsik]

It seems like the provided models need to be called with different evaluation scripts. eval.py uses the Chair_dgeo dir and eval_recon_sn.py uses Chair_dhier.

I start to suspect, that the derivation sn is used for StructureNet and the default evaluation scripts eval_vae_chair.sh runs a StructureNet example and not the model described in the published paper.

P.S. I think it is the described SN-Mesh model, which uses a different dataset encoding/format.