Changlin Li
Changlin Li
By definition the channel is 1 for density map. Why you need a thermal map?
> why multiply by 4? hope to answer my doubts，thanks! This is a trick. Adapt size to multiple of 4
> > 我猜你验证集的代码不对，记得最后一个cell中有一行注释，取消掉就好了 > > 你说的挺对的，我开始也是验证集精度很差，验证集part_B 的MAE是1900多，取消掉最后一个cell中的注释img = transform(Image.open(img_paths[i]).convert('RGB')).cuda()后，我的验证集精度为8.998，感谢 If you do this, then you negate all normalization process before this line.
Yes you should, upsample by factor of 8 and divide pixel by 64
How you are able to set batch_size as 2? Do you have the same images in same batch?
you should not use those value. As the author uses pretrain weights instead of training from zero. However, if you train from zero, you need those values then.
This is given by source data. Check Shanghai dataset for details
> > > > 减少lr > > > > > > > > > but the LR dropped to 1e-8 at the end of training > > > > >...
> ok, let me try 0.25 @jinfagang Did you successfully improve your performance by doing what you listed above?
> No, I tried your tool and dedicatedly design my anchor, result also not full satisfying although it seems work but have a lots of missdetections. @jinfagang I see. I...