search

ChinaYi / ASFormer

Official repo for BMVC2021 paper ASFormer: Transformer for action segmentation
MIT License
94 stars 19 forks source link

about the randomness of code #10

Closed wlsh1up closed 2 years ago

wlsh1up commented 2 years ago

Hi, Thank you for your code https://github.com/ChinaYi/ASFormer/blob/3940443a7dca336f879c43a42bcf91c5bf7c790f/model.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L370 When I change the test interval from 10 to 20, 30, etc., different results(such as training loss ) are obtained under the same seed. What do you think is the reason？ best regards

ChinaYi commented 2 years ago

Each test operation will call reset() function, see https://github.com/ChinaYi/ASFormer/blob/3940443a7dca336f879c43a42bcf91c5bf7c790f/model.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L393

where reset function will call random() function, which results different randomness even started with the same seed. https://github.com/ChinaYi/ASFormer/blob/3940443a7dca336f879c43a42bcf91c5bf7c790f/batch_gen.py#L21-L47