Haian Huang(深度眸)
Haian Huang(深度眸)
@nijkah I found that even after changing data augmentation, the effect has not improved. Now I cannot determine which part is the problem, my suggestion is that you can copy...
@rydenisbak Thank you. I'm now facing some issues reproducing the results, and it would be greatly appreciated if you could help identify the problems.
@AItairs Thank you. Can you help fix it?
@Wwupup Can you also create a PR to the `dev-3.x` branch?
@yokosyun Because our data flow is transmitted through the `datasampler` object, enabling pin_memory does not have any acceleration effect.
@20231211 Please use version 3.x of mmdet. It will be much more convenient for visualization purposes.
@initiater Can you try the mmdet 3.x branch?
1 refcoco 是可以拼起来一起训练的，但是有可能会超过 256 token 的限制，为了通用性和统一，我们没有进行合并。 2 open grounding 训练的时候并不是采用 phrase grounding 训练方式，而是类似 rec，这个和我们做法也是不一样的。我们是严格统一了所有数据的范式 3 而且 flickr30k 和 gqa 数据集在原始处理时候就已经拼接了，但是也存在一张图片多个 instance 情况。 按照目前的做法，refcoco 有部分数据确实可以合并，但是合并时候也要注意一些特殊情况
按类别拆分 coco 是啥意思？如果你改成并行，那训练效果就更低了
@chinabing `not NoneType` is mean the image path is incorrect.