custom data run

[anastasiiakim]

Hi, how to run it for custom data easy?

Also in src/gbd_utils/precondition.py function "get_discrete_prob" has missing values, how to get those for custom data?

def get_discrete_prob(self, dataset_name, dataset_info, prob=None, num_classes=None):
    if prob is not None:
        return prob

    if dataset_name == 'comm20':
        prob_X = torch.tensor([0.0000, 0.0393, 0.2376, 0.2376, 0.2192, 0.1394, 0.0910, 0.0340, 0.0020])
        prob_E = torch.tensor([0.2914])
    elif dataset_name == 'ego':
        ValueError('Not implenmted. The code will Coming soon.')
    elif dataset_name == 'planar':
        ValueError('Not implenmted. The code will Coming soon.')
    elif dataset_name == 'sbm':
        ValueError('Not implenmted. The code will Coming soon.')
    elif dataset_name == 'qm9':
        prob_X = dataset_info.node_types
        prob_E = dataset_info.edge_types_new
    elif dataset_name == 'zinc250k':
        ValueError('Not implenmted. The code will Coming soon.')
    else:
        raise ValueError('Invalid dataset')

    return (prob_X, prob_E)

[xinyangATK]

Hi @anastasiiakim, thanks for your attention!

We just provide the general framework for graph generation with GBD, along with examples for training public datasets. For "get_discrete_prob" with custom data, you can design your own reasonable strategy depended on your own purpose, e.g., summarize the distribution of node degree and predefine the corresponding precondition module.

btw I will update the continuous version of precondition module later next week so that you can just use this default choice to get mean and std, depending on the assumption on uniform distribution for each graph elements.

[anastasiiakim]

Any update on the continuous version of precondition module?

It seems if prob_X above consists of not exactly 9 elements (channels) the code will give an error:
src/graph_beta_diffusion_general.py", in training_step
noisy_data['X_t'] = (noisy_data['X_t'] - mean_logit_X_t) / std_logit_X_t
RuntimeError: The size of tensor a (9) must match the size of tensor b (19) at non-singleton dimension 2

above (19) is the dimension of my prob_X I calculated based on node degree distribution in my data. Can you please comment on this? Is this expected that there should be 9 channels only for any data? And so I need to pick prob_X with length of 9?

[xinyangATK]

The continuous version of precondition function is called when you feed the prior_prob as None or you can directly call the function get_logit_beta_stats_con in precondition.py.

For the error about node number defined in dataset, I think you may should check your all the scripts about dataset process and define the exact number of nodes for your own data.