Yi Zhang
Yi Zhang
你好，我这边给的实验13的[email protected]结果，是只进行稀疏性训练后的结果，并没有进行剪枝，所以我用实验13去做后续试验的原因是，稀疏性训练之后的结果比较好，所以没有进行剪枝之后的效果对比的实验。
稀疏性训练不是越多越好，越多之后，趋于0的权重变多，网络自然会变差，但是100可能没有达到你说的过拟合的状态，具体你可以看下分布图，然后第二个问题，finetuning时间长应该可以更大恢复精度，但是应该还是会有个极限的。
没有具体试过，可能需要稍作修改
You should conduct sufficient sparse training to ensure that there are many BN coefficients that tend to 0, so that this problem will not occur.
1. The distribution should be like the following figure  2.I don't quite understand what this problem is.
1. As you can see from your Tensorboard, this picture represents the distribution of the BN coefficient. 2. Strangely, when I carry out sparse training, the mAP will drop. I...
It's just that it works better.
What is your command to run?
The code is based on [MMSegmentaion v0.24.1](https://github.com/open-mmlab/mmsegmentation/tree/v0.24.1). You can check what version of mmcv is used for this version of mmsegmentation
There is a random reasoning part in [ham_head](https://arxiv.org/abs/2109.04553), which may fluctuate slightly, but it exists.