ccr-cheng / infgcn-pytorch Goto Github PK

Hello.

First of all, thank you for sharing your great work. I was impressed by your work and tried to reproduce your work to get insights.

But in the reproducing sequence, I had some troubles and wanted to ask about it.

When I tried to download the QM9 EDF dataset in the figshares, it was about 1T and ETA was about 6 days in my Linux systems.

In the downloading sequence, I used wget in the terminal but it was irregularly disconnected and started from the bottom.

Do you have any tips for downloading the QM9 EDF dataset?

And I downloaded the Cubit datasets from Figshares but the number of the samples in the whole dataset(ECD-database-part1,2,3) was 17418.

However, in your paper, you used the 16421 dataset as shown in Appendix C.

Could you provide the list of the train/val/test split name or code to divide and read the dataset? I was stuck in the dataset split process.

Finally, I got the reproduced results of ethane and malonaldehyde of the QM9. The results show that MAE as shown below:

Ethane with seed 42

Malonaldehyde with seed 42

In the training metric, the final inference MAEs were about 0.12. Are these NMAE which was reported in the paper? I think it is quite different from the results in your main results.

So can you provide the seeds that you used in the paper or the pre-trained model for the MD? Or some detailed configuration of your model may be helpful, I am confused that I missed something.

Thank you.

# md_malon.yml
train:
  seed: 42 # default seed 42
  train_samples: 1024
  val_samples: 2048
  max_iter: 2000 # maximum iteration
  batch_size: 64
  log_freq: 10000
  val_freq: 20
  save_freq: 100
  max_grad_norm: 100.
  optimizer:
    type: adam
    lr: 5.e-3
    weight_decay: 0.
    beta1: 0.9
    beta2: 0.999
  scheduler:
    type: plateau
    factor: 0.5
    patience: 5
    min_lr: 1.e-5

test:
  batch_size: 16
  inf_samples: 4096
  num_infer: null
  num_vis: 2

datasets:
  type: small_density
  root: ../dataset
  mol_name: malonaldehyde

model:
  type: infgcn
  n_atom_type: 3
  num_radial: 16
  num_spherical: 7
  radial_embed_size: 64
  radial_hidden_size: 128
  num_radial_layer: 2
  num_gcn_layer: 3
  cutoff: 3.
  grid_cutoff: 3.
  is_fc: false
  gauss_start: 0.5
  gauss_end: 5.

Hello, I am interested in your works and try to reproduce your project but I found some typo in ReadMe.

https://github.com/ccr-cheng/InfGCN-pytorch/blame/dfe402bd2798c7482a75e3e89686c0987be427db/README.md#L71

https://github.com/ccr-cheng/InfGCN-pytorch/blame/dfe402bd2798c7482a75e3e89686c0987be427db/README.md#L79

In the training and evaluation command code, I think main.py is missing

Change

python configs/qm9.yml --savename test

to

python main.py configs/qm9.yml --savename test

may help.

Thank you.

Dear sir,
I am interested in experimenting with the model trained in your project. Would it be possible for you to share the weights of the model? This would be immensely helpful for my work. Thanks.