xxxmy

减低学习率，loss降到15，但是检测效果比1.x版本效果loss20多的差太多，几乎没有对的框

你可以试着按照你那样的计算方式算一下看看，b_off和t_off会变成负值

不懂你想说啥

if you use large batch size, you can enlarge your learing rate accordingly such as batch size 32, lr 1e-3.

eval() is not equivalent to requires_grad.

sorry，the preprocessing of demo.py is for pre-trained model converted from official repo. You should uncomment some code in demo.py.

You didn't provide any information. I couldn't help you. Training curve, demo output or something else.

> > Thank you for your reply. Now there are often problems with the remote connection of our school's server. Many of the results don't show up. I'm looking into...

@zhengdq99 创建模型时没有fc层，也就不用管池化层。在resnet中将strict参数设为False即可。_gen_level_targets中那些可以注释掉。分类不需要进行正样本筛选。 不给出训练参数无从评价。可能原因：batch size太小，学习率太大等等。centerness在batch size较小时难以收敛。 ``` if if_include_top: self.fc = nn.Linear(512 * block.expansion, num_classes) ```

@zhengdq99 如果没看错的话，你才训练了一个epoch不到就说不收敛，还在warm up阶段-_-|||