[Feature Request] `instantiate` support for ListConfig at top level

[Jasha10]

The below text is copied directly from a post by a user of the internal Meta Hydra users group:

---

It seems that Hydra doesn't support recursive instantiation for list objects yet. For example, I would expect the following code to work, but it's raising an InstantiationException: Top level config has to be OmegaConf DictConfig, plain dict, or a Structured Config class or instance.

hydra.utils.instantiate([
    {"_target_": "torch.nn.Linear", "in_features": 3, "out_features": 4},
    {"_target_": "torch.nn.Linear", "in_features": 4, "out_features": 5},
])
# hydra.errors.InstantiationException: Top level config has to be OmegaConf DictConfig, plain dict, or a Structured Config class or instance

As another example, we often instantiate a list of callbacks and pass them to a trainer. Although the same thing can be done with a list comprehension, you can probably imagine its usefulness when composed with other structures. This looks like a simple and intuitive feature, so I'm curious if we plan to support it, if it hasn't been done already.

[Jasha10]

Workaround 1: wrapping in a toplevel dict:

A workaround is to wrap the top-level list in a dictionary before calling instantiate:

>>> cfg = OmegaConf.create([
...     {"_target_": "torch.nn.Linear", "in_features": 3, "out_features": 4},
...     {"_target_": "torch.nn.Linear", "in_features": 4, "out_features": 5},
... ])
>>> hydra.utils.instantiate({"foo": cfg}).foo
[Linear(in_features=3, out_features=4, bias=True), Linear(in_features=4, out_features=5, bias=True)]

Workaround 2: list comprehension

>>> [hydra.utils.instantiate(elt) for elt in cfg]
[Linear(in_features=3, out_features=4, bias=True), Linear(in_features=4, out_features=5, bias=True)]