Wang, Xiao
Wang, Xiao
I cannot disable either.
这种故意弹出窗口确实是检验脚本的一个好手段，是否有截图可供参考？理论上讲，如果有了坐标和颜色，我们可以添加点击来绕过这个检测。
@Williamongh 你好，感谢你的反馈。我之前考虑过检测“N回目”的方法，但是因为“N回目”的这个截图是动态的，非常难以截取，所以最好的办法就是不停的点怪。而且，如果在N回目开始后延迟2～3s点怪，（我当时玩的时候）会导致没有导航的超星茨木抓瞎然后灭团，现在玩家装备好了应该没有类似问题了。 问题是不管脚本的点怪策略怎么修改，网易那边都可以添加随机的人工检测（就像楼上评论里的那样），这个是没法绕过去的。在有办法绕过人工检测之前，修改点怪策略我感觉意义不大。
Hi, is there still any plan to add this feature, or can I read the timing unit anywhere from the generated json file?
I'm still working on extra benchmark and accuracy test for the new options at this moment.
cc @rwightman @csarofeen @ptrblck @kevinstephano @jjsjann123
I have some ResNet50 **AMP+channels-last** training results with cuda graph and nvfuser that verified the training accuracy (loss, val acc) here https://gist.github.com/xwang233/f3b5b4818762b08d716f969899b6d263. After 10 epochs, V100x8, BS = 128 mode...
ResNet50 **FP32** training results with eager, cuda graph, TS+nvfuser, AOT_autograd+nvfuser https://gist.github.com/xwang233/d5136facb3361af54693081da346fd33 After 10 epochs, A100x8, BS = 128 mode | throughput | eval top1 --- | --- | --- Eager...
@muzimuzhi Thanks! This works for me.
Repro step on V100 ``` TIMM_BENCHMARK_ENABLE_AOT_AUTOGRAD=1 python -u benchmark.py --bench train --model seresnet152d --img-size 224 -b 24 --fuser nvfuser --aot-autograd ```