qiang chen

@Seperendity 你好，这个是可以通过[广播机制](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/beginner/tensor_cn.html#id7)来实现的

@Seperendity 你好，这样做是为了配合v的维度。举个例子理解一下，假设v的shape是`[B, C, H, W]`，x_cnn2v的shape是`[B, C, 1, 1]`，那么`v = v * x_cnn2v`是一个简单的channel attention。但是，在代码里的第223行，因为后续要准备做window-based self-attention，v的shape是`[B*(H/win)*(W/win), win*win, num_heads, C/num_heads]`，而x_cnn2v的shape是`[B, C, 1, 1]`，这个时候没法直接做channel attention。当然这里可以用不同的实现： 1. 你可以把v再reshape回`[B, C, H, W]`，做完channel attention之后，再变成`[B*(H/win)*(W/win), win*win, num_heads, C/num_heads]`，再进入到下面的self-attention。...

Hi, sorry for the late reply. I don't have log files for multi-scale training now. You can train it for yourself. By following the settings in the paper, you can...

I don't understand that what files did you write. Could you provide more details about how you run a demo with YOLOF?

For the demo, you can use the "DefaultPredictor" directly. Could you debug the output of the predictor? I will help with it when I got time this week.

Hi, sorry for the late reply. It's normal that the performance drops when you set a higher threshold (e.g., 0.3). A higher threshold means that you remove several valid predictions...

There exist many TPs (True Positives) between 0.05 and 0.3. Thus the mAP is lower than the original one when you set the threshold to 0.3. A detailed analysis on...

You should check the whole validation set instead of one single image.

I check several visualizations. There indeed exists low score bounding boxes in some images, which may be wrong predictions. While for the performance calculation, you need to do the counting...

For flops calculation, we follow the steps of DETR. You can check [here](https://github.com/facebookresearch/detr/issues/110).