Runist

Results 4 issues of Runist

在训练时,验证集的损失能下降到10左右,但是实际测试的时候,计算出来的物体得分数很低,普遍低于0.3.而且画的框位置对的也不是很准。

这边按照默认option的参数从头训练了20个epcoh,lcc只有0.10多,是需要载入预训练权重吗?backbone是你们自己设计的,你们应该也是直接在AVA上训练的吧?

When I install dlib==19.24.9, The error occur: ``` ⅹ => [internal] load build definition from Dockerfile 0.0s ⅹ => => transferring dockerfile: 1.06kB 0.0s ⅹ => resolve image config for...

Hi Yoontae, I have a confusion about the training process. The model parameter didn't update to the next domain in main.py. Is it correct that only the variable E is...