Command line interpolations are not being evaluated lazily, leading to changed multirun behaviour

In v0.11, when I wanted to do multirun with certain parameters dependent on each other, I would do something like

./train.py param1=1,2,3 param2=\${param1} -m

and this would launch 3 jobs, with same values for param1 and param2.

With v1.0, this seems to be launching 3*3 = 9 jobs, with all possible combinations of param1 and param2. Wondering if that’s expected behavior? My understanding was that the interpolations are evaluated lazily, so it should have the same behavior as in v0.11.

System information

• Hydra Version : 1.0rc1
• Python version : python 3.7.7
• Virtual environment type and version : conda 4.7.10
• Operating system : Linux

[omry]

@egonuel thanks. the PR does not pass tests and is not ready. you are using it at your own risk 😃

I am not ready to push additional bugs onto this stack just yet. Please wait until I land what I am working on now and if there are still issues you can file new bugs.

[rohitgirdhar]

Here is my actual usecase: basically I have been using this functionality to couple variables, for eg, if I am running a model on 5 folds, and I want to load a pre-trained model corresponding to the specific fold, I would do something like this:

./train.py fold_id=0,1,2,3,4 pretrained=/checkpoint/\${fold_id}