add prototypes for Caltech(101|256) datasets
#4510
Conversation
| packages=find_packages(exclude=('test',)), | ||
| package_data={ | ||
| package_name: ['*.dll', '*.dylib', '*.so'] | ||
| package_name: ['*.dll', '*.dylib', '*.so', '*.categories'] |
There was a problem hiding this comment.
We need to some file type here, so that these file types are packaged with everything else. They are plain text files, but I've opted to give them a "custom" extension to not accidentally include anything else.
| from typing import Any, Callable, Dict, List, Optional, Tuple, Union | ||
| import re | ||
|
|
||
| import numpy as np |
There was a problem hiding this comment.
Do we need numpy here? You only use it to cast the array to int64, which you can also do in numpy with a string (or directly add a dtype in torch.as_tensor / torch.tensor)
There was a problem hiding this comment.
Unfortunately, we do. The box_coord array has dtype uint16, which torch.as_tensor cannot handle. So either we use numpy to convert to a common dtype first, or we could also convert the values to a list and build a tensor from that. What would you prefer? In any case I'll add a comment explaining what is going on.
|
|
||
| image = decoder(image_buffer) if decoder else image_buffer | ||
|
|
||
| ann = read_mat(ann_buffer) |
There was a problem hiding this comment.
Does ann contain anything else or just box_coord and obj_contour?
There was a problem hiding this comment.
Just box_coord and obj_contour
| anns_dp = TarArchiveReader(anns_dp) | ||
| anns_dp = Filter(anns_dp, self._is_ann) | ||
|
|
||
| dp = KeyZipper( |
There was a problem hiding this comment.
For my understanding: is this efficient compared to our current Caltech implementation based on the old-style datasets?
There was a problem hiding this comment.
I'm not sure how to answer if this is efficient or not.
Without the Shuffler we could use Zipper here and this is basically what we are doing now. If both datapipes, i.e. images_dp and anns_dp, are already aligned, KeyZipper only adds the overhead of evaluating key_fn and ref_key_fn for every sample.
As an alternative we could put the I was wrong: the two archives are not perfectly aligned so we need the Shuffler after we use a Zipper and remove the KeyZipper all together.KeyZipper anyway.
| create_categories_file(HERE, self.name, categories) | ||
|
|
||
|
|
||
| if __name__ == "__main__": |
There was a problem hiding this comment.
Question: do we want to keep this for all datasets?
There was a problem hiding this comment.
The idea behind this is that we should keep track of how we generated the category files. With this it is easy to regenerate if we need to. Just call python -m torchvision.prototype.datasets._builtin.caltech. Maybe we could also do an aggregation to be able to generate "everything", but I would keep that for the future.
pmeier
changed the title
Caltech256 datasetCaltech(101|256) datasets
Summary: * add prototype for `Caltech256` dataset * silence mypy Reviewed By: prabhat00155, NicolasHug Differential Revision: D31309545 fbshipit-source-id: b1b597c2ac152a77fa5c80f361359d993212ce47
* add prototype for `Caltech256` dataset * silence mypy
* add prototype for `Caltech256` dataset * silence mypy
3 participants
cc @pmeier @mthrok @bjuncek