Skip to content

Reproduce Results on dsprites problem. #1

@ThomasMrY

Description

@ThomasMrY

Hi MattPainter01,
Hope this email finds you well.
I am intersted in your work "Linear Disentangled Representations and Unsupervised Action Estimation" and build my experiments with your code. However, when I reproduce the results (dsprites) reported in your paper, I encountered the following problem:
I train the model on dsprites dataset with argments: "--model rgrvae --epochs 200 --gvae-gamma 10 --group-structure c+ c- --dataset dsprites --seed 0 --data-path datadir" and get the following results:

'dmetric/hig_acc':0.412
'dmetric/val_hig_acc':0.369
'dmetric/discrete_mig':0.008665972419410495
'dmetric/informativeness_train':0.9212
'dmetric/informativeness_test':0.1476
'dmetric/disentanglement':0.0029809895753678396
'dmetric/completeness':0.00398547160125029
'dmetric/modularity_score':0.6027582870941948
'dmetric/explicitness_score_train':0.6246885653134273

which is much lower than scores reported in paper.
Is there anything wrong with my experiment? Could you please help me to fix it?
Thank you very much,
Thomas

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions