Gaël Alguero

Here it is : [requirements.txt](https://github.com/scikit-hep/pyhf/files/5990266/requirements.txt) (also editing the main description) so `torch` version is `1.6.0`

I've updated `torch` to 1.7.1 and it's even worth : the observed CLs is a NaN at least up to `poi_test = 1.2`.

Using `test_stat="q"` indeed solved the NaN problem, however changing the lower bound to a negative number doesn't seem to be necessary but it slightly change the returned CLs. Thanks for...

> For some more progress, seems like some of the normfactors do hit the parameters at the limit during the non-gradient version of the optimization (`do_grad=False`). > > This corresponds...

Indeed it works now with the latest `iminuit` version. However, the (`pytorch, minuit`) duo is very very slow (~20 secs per hypotest call) and I also have to use `test_stat...