[Enhance] Random generator refactor #1459
Conversation
shijianjian
requested review from
ducha-aiki,
edgarriba and
lferraz
as code owners
| if torch_version_geq(1, 10) and "cuda" in str(device): | ||
| pytest.skip("AssertionError: Tensor-likes are not close!") |
|
@edgarriba We need to take flake8 down to < 4.0, check this: PyCQA/flake8#1419 |
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
shijianjian
force-pushed
the
feat/rng_refactor
branch
from
6349a11
to
45111a4
Compare
for more information, see https://pre-commit.ci
|
@edgarriba Please approve and merge. Tests failed on flake8 bugs. |
This is the issue. It will probably be fixed in the next flake8 version. |
|
@shijianjian then downgrade flake version |
| To apply the exact augmenation again, you may take the advantage of the previous parameter state: | ||
| >>> input = torch.randn(1, 3, 32, 32) | ||
| >>> aug = RandomPerspective(0.5, p=1.) | ||
| >>> (aug(input) == aug(input, params=aug._params)).all() |
There was a problem hiding this comment.
How’s it working right now. Are the parameters generated before nowing the input shape?
There was a problem hiding this comment.
No. it has to be the same shaped inputs. But this code has decoupled that logic that allowed easier refactoring later.
| else: | ||
| raise TypeError(f"Unsupported type: {type(self.kernel_size)}") | ||
|
|
||
| self.angle_sampler = Uniform(angle[0], angle[1], validate_args=False) |
There was a problem hiding this comment.
Is it possible to customise the different samplers ? Not only here
There was a problem hiding this comment.
Yes. You can override the make_samplers function for doing so.
There was a problem hiding this comment.
worth to mention that in docs or add an example (or and later PR)
This PR only refactored the whole code base, and allows easier customization later. As a backlog, some ideas like:
|
Changes
This PR refactored the random generator module, mainly for dragging all param validation, distribution generation into
__init__which reduced the unnecessary object reinitialization.Benchmark
Here we benchmarked with Torchvision default CPU augmentation and Kornia GPU augmentation, on Google Colab K80 GPU with different batch sizes. unit=ms, and
bsmeans batch size.The results recorded mean per-sample augmentation speed. Kornia got a better results when larger batch size.