Task 2: Self-supervised training of 2nd stage
eeishaan opened this issue · comments
- We want to evaluate the effects of training the 2nd stage in a completely self-supervised setting.
- Two self-supervised tasks are mentioned in the paper: graph edit distance and graph structure similarity .
- The goal is to implement or reuse existing implementation of both of them. Rest of the pipeline remains same and changes possibly include model changes, dataloader changes (not necessary).
Results :
unsupervised_pretrain_graph_level_from_scratch
58.36488029188599
unsupervised_pretrain_graph_without_graph_level_pretraining
63.143007055658515
unsupervised_pretrain_graph_contextpred_weight
56.02703642285913
unsupervised_pretrain_graph_with_masking_weights
67.88642939178769
unsupervised_pretrain_graph_info_graph
70.1961267598414