Chengjian Feng

> 作者你好，请问这个空间概率图和空间特征图是如何学习得到的？如果没有TAL策略是不是这个TAP就没有效果？那么为什么说T型头是一个独立的模块，可以在没有TAL的情况下很好地工作？ 1. 空间概率图的获取可以参考https://github.com/fcjian/TOOD/issues/13 2. 没有TAL策略，T-head/TAL也有一定的效果，具体可以参考论文Table 1的实验。而有了TAL策略，可以使得T-head/TAL更好地工作，取得更高的性能。

> 这个交互特征又是如何得到的呢？辛苦作者解答一下谢谢 通过N层共享的卷积：

> 但是论文中说道空间概率图和偏移图的学习是在TAL中实现的所以有点不理解，论文中原话为“where conv1 and conv3 are two 1×1 conv layers for dimen- sion reduction. The learning of M and O is performed by using the proposed Task Alignment Learning (TAL)...

Hi @tanghui2017108, the Equilibrium Loss function is [here](https://github.com/fcjian/LOCE/blob/master/mmdet/models/losses/equilibrium_loss.py).

The mean_score is defined as the mean classification score of each category. We adopt an online manner to compute the mean classification score, and you can refer to the [code](https://github.com/fcjian/LOCE/blob/0a0194d1b42c89ac062bed5e5fdb7a86ce9b70e4/mmdet/models/roi_heads/loce_roi_head.py#L234).

@zwenyu Hi, you can refer to the [tutorials](https://github.com/fcjian/InstaGen/blob/main/mmdetection/docs/en/user_guides/inference.md) provided by MMDetection.

@ohheysherry66 You can also fine-tune SD using LoRA. We tried it at the time and it achieves similar performance.

What is your training script?

--ckpt_path means the path of the model checkpoint, and you should remove it, namely: ` sudo /home/ww/.conda/envs/aedet2/bin/python /home/ww/Coding/AeDet/exps/aedet/aedet_lss_r101_512x1408_256x256_24e_2key.py --amp_backend native -b 8 --gpus 1 `

'- conv_offset' is used for adapting to the 'deform_conv2d' function. The 'deform_conv2d' function carries out the offset based on the sampling grid of the typical convolution. On the other hand,...