Skip to content

pretrained_checkpoint=None causes inconsistent predictions #168

@schnamo

Description

@schnamo

I noticed that when I use the fix for the pre-trained checkpoint issue, the model will output different attentions every time i run it on the same SMILES. now I took it out (bc this fix wasn't needed when i am working on our server where I pre-trained everything with the same setup myself) and now it seems to be consistent for each run. But that behaviour is quite weird no? After training a model the output should be consistent, anyone any idea what's going on there? running
kind='test'
model = model_class.load_from_checkpoint(ckpt_opt, input_dim = data_module.feature_vector_size, pretrained_checkpoint=None)
vs
model = model_class.load_from_checkpoint(ckpt_opt, input_dim = data_module.feature_vector_size)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions