move_data_to_device fails with dataclasses with field(init=False) fields

[ronif]

🐛 Bug

move_data_to_device has special handling of dataclasses, but in case the dataclass has a non-init field, elem_type(**out) in apply_to_collection will fail

To Reproduce

from pytorch_lightning.utilities import move_data_to_device
import torch
from dataclasses import dataclass, field
from typing import NamedTuple, List

@dataclass
class Data:
    example_id: List[str]
    x: torch.Tensor
    y: int = field(init=False)

    def __post_init__(self):
        self.y = 3

batch_size = 5

batch = Data(
    example_id=[f"e-{i}" for i in range(batch_size)],
    x = torch.rand(batch_size),
)

assert batch.y == 3
device = torch.device("cuda:0")

move_data_to_device(batch=batch, device=device)

Expected behavior

Should run without exceptions

Environment

• Packages:
  - numpy: 1.21.2
  - pyTorch_debug: False
  - pyTorch_version: 1.9.1
  - pytorch-lightning: 20211014
  - tqdm: 4.62.3
• System:
  - OS: Linux
  - architecture:
  - 64bit
  -
  - processor: x86_64
  - python: 3.9.7

[ronif]

I think this is an easy fix. Can submit a PR if needed.

[justusschock]

@ronif that would be great!