direct-pretraining
direct-pretraining copied to clipboard
wxzs5
Fast and accurate pipeline for training object detector
Hi, thanks for your work. I see that in the P1x configs, the bbox_head is changed to `Shared4Conv1FCBBoxHead`. I'm curious if the baseline models (pretrained on ImageNet) uses the same...
Hi, Thank you for your work. I was wondering what is the motivation for using fixed learning rate during pretraining? Is it purely from empirical results or are there any...
In retinenet's direct pre-train, I observed that loss becomes nan or inf during warmup when some seeds are specified for train.py. Wrt default seed, No problem. I hope to re-explore...
Isn't the learning rate of direct-pretraining for RetinaNet 0.08 ? Wrt the paper, 0.04 is the setting on the less data. If I'm mistaken, my apologies.
hoping to add or steps to follow with repo https://github.com/google/automl/tree/master/efficientdet to use the mentioned methods, helpful to all
