[transforms] Handle samples without boxes in RandomRotate and RandomCrop

[raashish1601]

RandomRotate and RandomCrop did not handle a sample without any box (a page with no text, or a negative sample):

import numpy as np, torch
from doctr.transforms import RandomRotate, RandomCrop
from doctr.utils.common_types import Sample

img = torch.rand(3, 40, 100)
RandomRotate(10)(Sample(image=img, target=np.zeros((0, 4, 2), dtype=np.float32)))
# ValueError: zero-size array to reduction operation maximum which has no identity
RandomCrop()(Sample(image=img, target=np.zeros((0, 4), dtype=np.float32)))
# same error

mask = torch.ones(40, 100, dtype=torch.bool)
out = RandomRotate(30, expand=True)(Sample(image=img, mask=mask, target={"words": np.zeros((0, 4, 2), dtype=np.float32)}))
out.image.shape, out.mask.shape  # (3, 40, 100) vs (1, 66, 108)

The dict-target paths already skip empty classes, but plain array targets went straight into the min/max reductions. And when every class of a dict target was empty, RandomRotate returned the original image while the mask had been rotated, so the image and mask no longer matched.

Changes:

• RandomRotate: an empty array target, or a dict target with only empty classes, now rotates the image (like the target=None path) and keeps the empty target. A small _rotate_image_only helper, mirroring RandomCrop._crop_image_only, is shared with the target=None path.
• RandomCrop: an empty array target keeps the sample as is, the same as a dict target without any box does today.

Tests: added test_random_rotate_empty_target (box, polygon and dict targets, with a mask) and test_random_crop_empty_target. They fail on main and pass here; pytest tests/pytorch/test_transforms_pt.py passes (45 tests). ruff check, ruff format --check and mypy on the changed module are clean.

[felixdittrich92]

Hi @raashish1601 👋,

Thanks for the PR only a small change and we are good to merge 👍

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.55%. Comparing base (b72ca69) to head (39c603d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2163      +/-   ##
==========================================
+ Coverage   97.54%   97.55%   +0.01%     
==========================================
  Files         170      170              
  Lines       10658    10665       +7     
==========================================
+ Hits        10396    10404       +8     
+ Misses        262      261       -1     

Flag	Coverage Δ
unittests	97.55% <100.00%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[felixdittrich92]

Thanks @raashish1601 LGTM 👍