Automatic augmentations gallery #4648
Conversation
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
stiepan
force-pushed
the
auto_aug_gallery
branch
from
f3fa6f0 to
fc904f4
Compare
|
!build |
|
CI MESSAGE: [7295532]: BUILD STARTED |
|
CI MESSAGE: [7295532]: BUILD FAILED |
| mag_range : (int, int) | ||
| Specifies the range of applicable magnitudes for the operation. | ||
| randomly_negate: bool | ||
| If true, the magnitude from the mag_range will be randomly negated for every sample. |
There was a problem hiding this comment.
Just a note, but this may be an interesting distinction between the magnitude being negated or the parameter being negated, as it might not hold that as_param(-magnitude) = -as_param(magnitude). And this may be a source of potential errors.
|
|
||
| def __init__(self, op, mag_range=None, randomly_negate=None, as_param=None, param_device=None): | ||
| self._op = op | ||
| self._mag_range = mag_range |
There was a problem hiding this comment.
Just a note, but maybe it is good idea to include some strict checks of what is in the mag_range, so we end up with less cryptic error messages?
| config = {name: value for name, value in self._get_config()} | ||
| for name, value in ( | ||
| ('mag_range', mag_range), | ||
| ('as_param', as_param), | ||
| ('randomly_negate', randomly_negate), | ||
| ('param_device', param_device), | ||
| ): | ||
| if value is not _DummyParam: | ||
| config[name] = value | ||
| return cls(self._op, **config) | ||
|
|
||
| def _get_config(self): | ||
| return [(name, value) for name, value in ( | ||
| ('mag_range', self._mag_range), | ||
| ('as_param', self._as_param), | ||
| ('randomly_negate', self._randomly_negate), | ||
| ('param_device', self._param_device), | ||
| )] |
There was a problem hiding this comment.
Just a nitpick, this looks a bit repetititive.
| def __repr__(self): | ||
| aug_params_repr = [repr(self._op)] | ||
| config = [(name, val) for name, val in self._get_config() if val is not None] | ||
| config_reprs = [f"{name}={repr(param)}" for name, param in config] | ||
| aug_params_repr.extend(config_reprs) | ||
| return f"Augmentation({', '.join(aug_params_repr)})" |
There was a problem hiding this comment.
👍 I wanted to suggest that we should have a good repr that shows the argument <-> magnitude mappings, if we add the symmetry via negation + possibly the name of the operation somehow. This is a good start.
| def param_device(self): | ||
| return self._param_device or "cpu" | ||
|
|
||
| def augmentation(self, mag_range=_DummyParam, randomly_negate=_DummyParam, as_param=_DummyParam, |
There was a problem hiding this comment.
Hmm, neat. If we intend to use this as a function, maybe we should give it a bit different name though? Just a though.
Another thought I had was to just allow @augmentation decorator stacking, this is possibly a bit more tricky for us, but maybe it would be more consistent for the user?
Basically:
@augmentation(mag_range=(0, 100)
@augmentation(mag_range=(0, 10)
def foo(....)
passshould work, but the usage would be to decorate already existing one?
OTOH, it's less convenient as you don't get the result as a return value and are better off just calling the decorator directly to capture the new callable.
There was a problem hiding this comment.
It does not hurt to allow nesting, I can adjust that. As for other names - I don't have a good candidate. And the augmentation name has the benefit of being consitent with the decorator name.
| The `@augmentation` lets you to specify what should the range of magnitudes and if the magnitude | ||
| should be randomly negated. Additionally, the `as_param` helps to separate the computation of |
There was a problem hiding this comment.
And I think our wording should be a bit different. (this is a bit nitpicking, and doesn't take away from the implementation).
We have magnitude space: [0, 1, ..., n], mapped into parameter range [0.0, p). The range of magnitudes would be selected by the number of bins, the range of the applied parameter is customizable.
There was a problem hiding this comment.
The AA paper states
We discretize the range of magnitudes into 10 values (uniform spacing) so that we can use
a discrete search algorithm to find them.
so there, the [0.0, p] is the magnitude range. Overall it is a bit ambigious, but instead of renaming the magnitudes into parameters I'd rather put more emphasis on magnitude and magnitude bins. How about such distinction?
|
|
||
| The augmentations in this module are defined with some default ranges passed to `@augmentation`. | ||
| The parameters can be easily adjusted. For example, to increase the magnitudes range | ||
| of `shear_x`, you can create `my_shear_x = shear_x.augmentation(mag_range=(0, 0.5))`. |
There was a problem hiding this comment.
If there is randomly_negate=True, should we check if the user passed a range that makes sense - that is (0, something)?
There was a problem hiding this comment.
I don't think so. ((-10, 10), False) and ((-10, 10), True) for AA or RA result in different runtime magnitudes. One or another coud make sense. Or (-5, 15) etc.
| [('shear_x', 0.0, 0), ('solarize', 0.8, 4)], | ||
| [('shear_y', 0.8, 0), ('color', 0.6, 4)], | ||
| [('color', 1.0, 0), ('rotate', 0.6, 2)], | ||
| [('equalize', 0.8, 4)], |
There was a problem hiding this comment.
Originally it is equalize with 0 probability of being run followed by equalize with 0.8 probability.
|
|
||
| auto_augment_image_net_policy = Policy( | ||
| "ImageNet", 11, { | ||
| "shear_x": a.shear_x.augmentation((0, 0.3), True), |
There was a problem hiding this comment.
Should the @augmentation get optional name argument otherwise infer it from the function name, so we can get neat __repr__ for all that stuff?
There was a problem hiding this comment.
The name idea is nice, we could accept here just a list and make a dict of it.
There was a problem hiding this comment.
On the other hand, it kind of compilcates things I guess. I just take the name (__name__) of the decorated function.
| "invert": a.invert, | ||
| "equalize": a.equalize, | ||
| "auto_contrast": a.auto_contrast, | ||
| }, [ |
There was a problem hiding this comment.
I guess there is no escape from the dict that has all the usable operations grouped up.
It's both nice for us (we can for example print them), and a bit of a shame (it was a bit shorter and more direct).
stiepan
force-pushed
the
auto_aug_gallery
branch
from
93154ab to
95af95f
Compare
|
!build |
|
CI MESSAGE: [7309341]: BUILD STARTED |
|
CI MESSAGE: [7310181]: BUILD STARTED |
|
CI MESSAGE: [7309341]: BUILD FAILED |
|
CI MESSAGE: [7310181]: BUILD PASSED |
|
!build |
|
CI MESSAGE: [7377149]: BUILD STARTED |
|
CI MESSAGE: [7377149]: BUILD FAILED |
|
!build |
|
CI MESSAGE: [7389682]: BUILD STARTED |
|
CI MESSAGE: [7389682]: BUILD PASSED |
|
!build |
|
CI MESSAGE: [7391042]: BUILD STARTED |
|
CI MESSAGE: [7391042]: BUILD PASSED |
stiepan
force-pushed
the
auto_aug_gallery
branch
from
382a760 to
5960f74
Compare
klecki
added
the
automatic augmentations
stiepan
force-pushed
the
auto_aug_gallery
branch
from
b466118 to
24b0ac6
Compare
Signed-off-by: Kamil Tokarski <[email protected]> Rename the selection operator Signed-off-by: Kamil Tokarski <[email protected]> Use random.uniform discrete mode Signed-off-by: Kamil Tokarski <[email protected]> Pass kwargs directly Signed-off-by: Kamil Tokarski <[email protected]> Assure uniform choice of manitudes in TA Signed-off-by: Kamil Tokarski <[email protected]> AutoAugment: coallsce pollicies Signed-off-by: Kamil Tokarski <[email protected]> Avoid unnecessary casts in color operations expressed in terms of arithm ops Signed-off-by: Kamil Tokarski <[email protected]> AutoContrast, one less float32 intermediate result Signed-off-by: Kamil Tokarski <[email protected]> Expose select operator Signed-off-by: Kamil Tokarski <[email protected]> Review rework, usage lessons learnt from the expiriments Signed-off-by: Kamil Tokarski <[email protected]> Adjust the docs Signed-off-by: Kamil Tokarski <[email protected]> Better handling of non-unique names, better custom kwargs handling, improved error messaging Signed-off-by: Kamil Tokarski <[email protected]> Rework pretty_select and utils Signed-off-by: Kamil Tokarski <[email protected]> Linter complaints Signed-off-by: Kamil Tokarski <[email protected]> Private modules Signed-off-by: Kamil Tokarski <[email protected]> Simplify docs, simplify code Signed-off-by: Kamil Tokarski <[email protected]> Move policy to separate module Signed-off-by: Kamil Tokarski <[email protected]> More (and more intelegible) docs Signed-off-by: Kamil Tokarski <[email protected]> Type annotations Signed-off-by: Kamil Tokarski <[email protected]> Improved error reporting for as_param Signed-off-by: Kamil Tokarski <[email protected]> More specific callable annotations Signed-off-by: Kamil Tokarski <[email protected]> Polish names Signed-off-by: Kamil Tokarski <[email protected]> Better error messaging, polishing Signed-off-by: Kamil Tokarski <[email protected]> Better error-reporting, warnings, default random_sign Signed-off-by: Kamil Tokarski <[email protected]> Make Aug private Signed-off-by: Kamil Tokarski <[email protected]> Wrapper tests Signed-off-by: Kamil Tokarski <[email protected]> More tests Signed-off-by: Kamil Tokarski <[email protected]> More tests Signed-off-by: Kamil Tokarski <[email protected]> Test select Signed-off-by: Kamil Tokarski <[email protected]> More tests Signed-off-by: Kamil Tokarski <[email protected]> Shape passing fix Signed-off-by: Kamil Tokarski <[email protected]> More tests Signed-off-by: Kamil Tokarski <[email protected]> More tests Signed-off-by: Kamil Tokarski <[email protected]> Even more tests Signed-off-by: Kamil Tokarski <[email protected]> Even even more Signed-off-by: Kamil Tokarski <[email protected]> Fix a comment Signed-off-by: Kamil Tokarski <[email protected]> Fix linting issues Signed-off-by: Kamil Tokarski <[email protected]> Replace contrast with arithm ops Signed-off-by: Kamil Tokarski <[email protected]> Signed magnitude wrapper Signed-off-by: Kamil Tokarski <[email protected]>
Signed-off-by: Kamil Tokarski <[email protected]>
Signed-off-by: Kamil Tokarski <[email protected]>
Signed-off-by: Kamil Tokarski <[email protected]>
stiepan
force-pushed
the
auto_aug_gallery
branch
from
da572bc to
d4f4bb5
Compare
Signed-off-by: Kamil Tokarski <[email protected]>
Signed-off-by: Kamil Tokarski [email protected]
Category:
Description:
Additional information:
Affected modules and functionalities:
Key points relevant for the review:
Tests:
Checklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: N/A