Add `params_to_tune` for new NNs (`RNNModel`, `MLPModel`)

[Mr-Geekman]

🚀 Feature Request

Add params_to_tune for new NNs (rnn, mlp).

Proposal

Sources:

• pycaret
  • mlp
• kats
  • lstm

Suggested grid for RNNModel:

{
    "num_layers": IntUniformDistribution(1, 3, step=1),
    "hidden_size": IntUniformDistribution(4, 64, step=4),
    "lr": LogUniformDistribution(1e-5, 1e-2),
}

Suggested grid for MLPModel:

{
    "num_layers": IntUniformDistribution(1, 3, step=1),
    "hidden_size": IntUniformDistribution(4, 64, step=4),
    "lr": LogUniformDistribution(1e-5, 1e-2),
}

To discuss:

• Is tuning lr necessary?
• It is unclear how to parameterize MLPModel parameter hidden_size because it accepts list of sizes for the layers. Even if we come up with parametrization it is unclear how to implement it because it is expected that we have a complete set of parameter values. We can use CategoricalDistribution but it seems not very effective.

Test cases

Look at https://github.com/tinkoff-ai/etna/issues/1184.

Additional context

No response

[martins0n]

• Is tuning lr necessary? - I think it's ok for now but may be can make lr a little bit wider from the left
• we can add context size for rnn

[Mr-Geekman]

Closed by #1218.