Update FAIR1M dataset and datamodule #1275
Conversation
isaaccorley
force-pushed
the
datasets/fair1mv2
branch
from
1d0d25b
to
e74ecad
Compare
…nto datasets/fair1mv2
github-actions
bot
added
datamodules
isaaccorley
force-pushed
the
datasets/fair1mv2
branch
3 times, most recently
from
56ffb9c
to
6cdd0a0
Compare
There was a problem hiding this comment.
@adamjstewart can comment on the versionadded stuff, else LGTM
|
Yes, it does need versionadded |
isaaccorley
force-pushed
the
datasets/fair1mv2
branch
from
6cdd0a0
to
2ef0cad
Compare
| don't match | ||
|
|
||
| .. versionchanged:: 0.5 | ||
| Added *split* and *download* parameters. |
There was a problem hiding this comment.
The split and download parameters.
There was a problem hiding this comment.
This seems like more of a nitpick. Added is more clear. The is just a statement.
There was a problem hiding this comment.
It should also be versionadded, not versionchanged. The versionadded template already says "New in version 0.5:"
| from .utils import dataset_split | ||
|
|
||
|
|
||
| def collate_fn(batch: list[dict[str, Tensor]]) -> dict[str, Any]: |
There was a problem hiding this comment.
How is this different from unbind_samples?
There was a problem hiding this comment.
Because of the torch.stack on line 27
| .. versionadded:: 0.5 | ||
| """ | ||
| output: dict[str, Any] = {} | ||
| output["image"] = torch.stack([sample["image"] for sample in batch]) |
There was a problem hiding this comment.
Does this line do anything?
There was a problem hiding this comment.
This was mostly copied from nasa_marine_debris.py. I'm assuming it's there for mypy reasons.
There was a problem hiding this comment.
But doesn't this just unstack and restack so that the output is identical?
There was a problem hiding this comment.
No, this takes the batch (a list of sample dicts) and grabs each image and stacks it into a single tensor along a new batch dimension.
adamjstewart
added
the
backwards-incompatible
Originally we created the FAIR1M dataset when only train/part1 images and labels were available.
This PR updates the FAIR1M dataset and datamodule to work with the latest train/val/test sets.