[BUG] Retinaface implementation have different behavior compared to retinaface-pytorch repo - Githubissues

nodefluxio / vortex

A Deep Learning Model Development Framework for Computer Vision

27 stars 6 forks source link

[BUG] Retinaface implementation have different behavior compared to retinaface-pytorch repo #8

Open alphinside opened 4 years ago

alphinside commented 4 years ago

Describe the bug Using same parameter, on same dataset, got different loss behavior,

Vortex training reference : https://www.comet.ml/hyperion-rg/vortex-face-detection/a74a56c4d0d541fb96cfacfa402f4595?experiment-tab=chart&showOutliers=true&smoothing=0&view=Default%20view&xAxis=step

torch-retinaface training-reference : https://www.comet.ml/hyperion-rg/torch-retinaface/ab07ab8c730b4daf8e81f94b01131f93?experiment-tab=chart&showOutliers=true&smoothing=0&view=Default%20view&xAxis=step

Expected behavior The behaviour should be similar