Kidleyh

and i tried deeprobust==0.2.2 on ubuntu 16.04 LTS ,it has same problem as 0.2.4

yes，i also tried `examples/test_mettack.py` , it is also the problem of `self.inner_train` , and i find this bug again on windows11 when i use torch==1.10.0,so i think maybe it is...

error message may not tell us any useful information, because the error message is about the cuda out of memory due to the `self.inner_train` could not empty its memmory normally.And...

Now i can run the code normally use torch==1.8.0, but I'm afraid the code has some questions with torch=1.10.0. Anyway, i think you can check the `self.inner_train` in `Metaack` ,...

Hi, sorry to bother you again, I also want to know how much cuda memory capacity is enough for training metaApprox to attack pubmed, i have attacked cora,citeseer and polblogs...