Custom dataset, dataloader returns list[list[dict]]

I am having a curious problem with a custom dataset, registered with a custom dataset function.

The Detectron2 setup and `cfg` is copied from [Detic Colab tutorial](https://colab.research.google.com/drive/1QtTW9-ukX2HKZGvt0QvVGqjuqEykoZKI)

A snippet of my code:
```
from detectron2.modeling import build_model
from detectron2.data import build_detection_test_loader
import torch

def my_dataset_generator():
data_list = []
for index, observation in df.iterrows():
data_dict = {}
data_dict["file_name"] = observation['file_path']
data_dict['image_id'] = observation['image']
data_list.append(data_dict)
return data_list

DatasetCatalog.register("my_whales3", my_dataset_generator)

model = build_model(cfg) # returns a torch.nn.Module
test_loader = build_detection_test_loader(cfg, 'my_whales3')
```
which returns
```
[03/01 12:05:54 d2.data.dataset_mapper]: [DatasetMapper] Augmentations used in inference: [ResizeShortestEdge(short_edge_length=(800, 800), max_size=1333, sample_style='choice')]
[03/01 12:05:54 d2.data.common]: Serializing 51033 elements to byte tensors and concatenating them all ...
[03/01 12:05:54 d2.data.common]: Serialized dataset takes 6.42 MiB
```
Note that the output format of ` my_dataset_generator()` is `list[dict]`. Also, the `df` in the function is the actual image metadata Pandas dataframe.

Running
```
model.eval()
with torch.no_grad():
outputs = model(test_loader)
```
produces this traceback:
```
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
/tmp/ipykernel_3448/2278559883.py in
1 model.eval()
2 with torch.no_grad():
----> 3 outputs = model(test_loader)

/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1101 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102 return forward_call(*input, **kwargs)
1103 # Do not call functions when jit is used
1104 full_backward_hooks, non_full_backward_hooks = [], []

~/Detic/detic/modeling/meta_arch/custom_rcnn.py in forward(self, batched_inputs)
113 """
114 if not self.training:
--> 115 return self.inference(batched_inputs)
116
117 images = self.preprocess_image(batched_inputs)

~/Detic/detic/modeling/meta_arch/custom_rcnn.py in inference(self, batched_inputs, detected_instances, do_postprocess)
94 assert detected_instances is None
95
---> 96 images = self.preprocess_image(batched_inputs)
97 features = self.backbone(images.tensor)
98 proposals, _ = self.proposal_generator(images, features, None)

/opt/conda/lib/python3.7/site-packages/detectron2/modeling/meta_arch/rcnn.py in preprocess_image(self, batched_inputs)
222 Normalize, pad and batch the input images.
223 """
--> 224 images = [x["image"].to(self.device) for x in batched_inputs]
225 images = [(x - self.pixel_mean) / self.pixel_std for x in images]
226 images = ImageList.from_tensors(images, self.backbone.size_divisibility)

/opt/conda/lib/python3.7/site-packages/detectron2/modeling/meta_arch/rcnn.py in (.0)
222 Normalize, pad and batch the input images.
223 """
--> 224 images = [x["image"].to(self.device) for x in batched_inputs]
225 images = [(x - self.pixel_mean) / self.pixel_std for x in images]
226 images = ImageList.from_tensors(images, self.backbone.size_divisibility)

TypeError: list indices must be integers or slices, not str
```

So the `x` in `batched_inputs` is a list, not a dict as expected. When I try
```
data_iter = iter(test_loader)
i = 0
for x in data_iter:
if i % 1000 == 0:
print(type(x))
i = i + 1
```
the output is
```

(...)

```
In other words, my data loader seems to return `list[list[dict]]`, instead of `list[dict]`. Have I goofed, or is this error due to something else?

Start by running the shown custom dataset registration and comparing one batch from build_detection_test_loader with the input expected by detic/modeling/meta_arch/custom_rcnn.py. Read detectron2/modeling/meta_arch/rcnn.py around preprocess_image and the loader entry point; done means the documented inference path uses the expected batch structure without the traceback.

Written by the indexing model from the issue text.