WEN Hao
WEN Hao
Check your bounding box annotations to see if there are zero (or even negative) area boxes, which might have been created by mistake. Several months ago, I encountered this NaN...
> Thank you! This looks like a great improvement. > > I'm not entirely familiar with `pathlib` so I'm a little unsure of some of the details. > > One...
> One thing to be wary of is that "file names" or "record names" might be either filesystem paths or URL paths. Some existing functions permit record_name to include a...
I think there's a lot of work to do to better deal with the paths. For example, I found currently `wfdb` detects whether a `pn_dir` contains the version number using...
> Thanks very much for identifying the bugs. We're going to think a bit more about how we want the API to be in the next major version, such that...
第二个是说做evaluate的图片不在validation set里。会不会是`drop_last`引起的呢？我训练的时候没有遇到过。另外，你说的报错这句我看是在第266行。
> 1. dataset.py 需要加上import torch > 2. coco_eval.py 222行 `assert set(annsImgIds) == (set(annsImgIds) & set(self.getImgIds())), \ 'Results do not correspond to current coco set'` 这里引发了异常 > 3. 多GPU训练时，`torch.load()` 均报错，加个判断是否使用多gpu更好 第三个问题的话，是不是需要像`tool.tv_reference.train.py`里面的`main(args)`函数一样改写`train.py`里的函数呢？我没有用过多GPU训练，暂时还不知道。
> @wenh06 你有训coco吗？我用单卡2080试了一下，bs=2, 608输入，显存占了7G不到，要12h/epoch，我看作者说是单卡训的，不知道他要训多长时间 没试呢。我搞完目标检测的小项目之后现在在搞别的，目前这个已经delay很久了。。。。有时间我去试一下。
> > > 1. dataset.py 需要加上import torch > > > 2. coco_eval.py 222行 `assert set(annsImgIds) == (set(annsImgIds) & set(self.getImgIds())), \ 'Results do not correspond to current coco set'` 这里引发了异常 >...
> 有大佬遇见过 训练了两三轮，AP AR 都是0的问题嘛? 有可能标签有问题？或者你开tensorboard看看tf events里面记录的那些图？