GuardSkill

Modify Pad Conv ConvTranspose Slice Add RandomNormal Tile ops

2

1. Modify Conv op support F.Conv() 2. Modify Pad op for general support 2. Modify Slice op support negative step 3 .Add Tile and RandomNormal ops

When I convert the model, I can't get the Conv weight ? I am confused about it.

4


Hi ,Could you provide AVA Pretrained model?

1

Thank you very much

Is there a worng understand in total variation?

5

I find this does not conform to the original paper’s method, I think the sum of the abs value should be taken into the Loss（tv）, and the tv loss is...

Why this code don't apply adversarial training?

1

It didn't train the discriminator for prior distirbution. I am very confused about this...

How to convert the pretrained model to Onnx or TensorRT

3

I found the deployment of StyleGAN is very difficult because of custom ops. Could you provide some help?

非常感谢你的帮助

请教关于关于box score为0的问题。

1

我现在正在将320x320的单类别yolov5移植到rknn上，用的 airockchip 的项目进行训练，修改了C demo中的yolo.cc中的类别数，但是目前我发现我的int8和fp32模型中的box_conf的全是0,请问我是不是漏掉了什么应该需要修改的地方 目前实在RKNN3399Pro平台上进行运行，在docker中进行模型转化，模型转换时提示cos相似度提升达到99%。用python的脚本推理torchscript时，是正常的，score能达到0.95.

hello, thanks for your work, I just want to see . Why does ComfyUI need train a new lora?

Does this support dynamic resolution? thx