除以七

I want to use `r100` pretrained model to finetune the custom face rec dataset, but the default `num_classes=93431` , I have 50 people img folder, should I set the `num_classes=50`?...

I drew the yolov9 structure graph with some introduction to help everyone understand the model better with English and Chinese.

I have a question About the eq.7 in paper

1

Eq7: $\left.\Delta s=\left\{\begin{array}{cc}\Delta i_c\cdot\Delta i&\quad\text{if}|\Delta i_c\cdot\Delta i|\geq\beta\\0&\quad\text{otherwise}\end{array}\right.\right.$ In my mind, the $\Delta i_c$ is $\Delta i[c]$, that's right? So the $\Delta i_c$ is a number, and the $\Delta i$ is...

`class Model(nn.Module): ······ def forward(self, x, x_target, x_last, A_act, lamda_act): print("A_act.shape = {}, lamda_act = {}".format(A_act.shape) >>> A_act.shape=2,2,25,25 What's the A_act?

For the specified number of key points custom dataset debug（针对自定义关键点数量的yolopose的debug）

3

我总结了一下需要改的地方： https://zhuanlan.zhihu.com/p/603799078 如果对您有帮助给个点赞收藏哈哈哈

onnx模型及导出的rknn模型如下网盘： ``` 链接：https://pan.baidu.com/s/1qm7Q6Dr1yD8CBq9yzoYsvQ?pwd=rknn 提取码：rknn ``` onnx导出fp16的rknn模型没有出现报错；在do_quantization的时候出现Error，但仍能导出int8模型，报错如下：  但是报错仍然能导出int8的模型； 在3588上进行推理，发现该模型似乎能够推理出正确的输出形状，但速度相当慢，并报错算子问题：