Simplify AutoAugment graph #4751
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Kamil Tokarski <[email protected]>
Signed-off-by: Kamil Tokarski <[email protected]>
Signed-off-by: Kamil Tokarski <[email protected]>
stiepan
commented
Mar 28, 2023
| self._random_sign = random_sign | ||
| self._signed_magnitude_idx = signed_magnitude_idx | ||
|
|
||
| def __getitem__(self, idx: int): |
There was a problem hiding this comment.
This enables the signed_bin(mutli_stage_bin)[stage_idx].
jantonguirao
approved these changes
Mar 29, 2023
stiepan
added
the
automatic augmentations
|
!build |
|
CI MESSAGE: [7762008]: BUILD STARTED |
|
CI MESSAGE: [7762008]: BUILD PASSED |
klecki
approved these changes
Apr 4, 2023
Comment on lines
+227
to
+229
| The output is a tuple of matrix `m` and per stage operators augmentations list `augments`, | ||
| such that for policy `sub_policy_idx` as the `stage_idx`-ith operation in a sequence, the | ||
| `augments[stage_idx][m[sub_policy_idx][stage_idx]]` operator should be called. |
There was a problem hiding this comment.
If I may try to expand the comment:
Suggested change
| The output is a tuple of matrix `m` and per stage operators augmentations list `augments`, | |
| such that for policy `sub_policy_idx` as the `stage_idx`-ith operation in a sequence, the | |
| `augments[stage_idx][m[sub_policy_idx][stage_idx]]` operator should be called. | |
| The output is a tuple `(m, augments)`, where `augments` is a list of augmentations to be | |
| used for given stage - each entry contains the reduced list of unique augmentations to be | |
| used for that stage. | |
| The `m` matrix contains the mapping from the original sub_policy_id, to the index within the | |
| reduced list, for every stage.T | |
| That is for policy `sub_policy_idx` as the `stage_idx`-ith operation in a sequence, the | |
| `augments[stage_idx][m[sub_policy_idx][stage_idx]]` operator should be called. |
There was a problem hiding this comment.
Sure, I'll merge this one, revase the #4753 and expand the comment there.
Comment on lines
117
to
123
| for stage_id in range(max_policy_len): | ||
| magnitude_bin = magnitude_bins[stage_id] | ||
| if use_signed_magnitudes: | ||
| magnitude_bin = signed_bin(magnitude_bin) | ||
| if should_run[stage_id] < run_probabilities[stage_id]: | ||
| op_kwargs = dict(sample=sample, magnitude_bin=magnitude_bin, | ||
| op_kwargs = dict(sample=sample, magnitude_bin=magnitude_bins[stage_id], | ||
| num_magnitude_bins=policy.num_magnitude_bins, **kwargs) | ||
| sample = _pretty_select(augmentations, aug_ids[stage_id], op_kwargs, | ||
| sample = _pretty_select(augmentations[stage_id], aug_ids[stage_id], op_kwargs, | ||
| auto_aug_name='apply_auto_augment', | ||
| ref_suite_name='get_image_net_policy') |
There was a problem hiding this comment.
Honestly, this is so cool ;)
18 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Category:
Refactoring (Redesign of existing code that doesn't affect functionality
Description:
This PR combines a couple of improvements to AutoAugment processing graph.
selectoperation.For example, the default v0 policy uses 25 sub-policies, each sub-policy is a sequence of (at most) 2 augmentations. Even though there are 25 sub-policies, there are only 11 unique augmentations present in those sub-polcies. DALI already splits the computation only in 11 parts. However, it can be further improved if the operators used in given stages of sub-policies form even smaller groups. In case of the default v0 policy, it is 8 vs 11.
In other words, when the comutation is split to apply the "i-th" augmentation in a sequence according to selected sub-polciy, there is no need to take into account all augmentations present in all sub-polcies, but only the augmentations at the i-th position in all sub-policies.
signed_binhelper that handles random negation of magnitudes can now accept additional shape parameter. With that, if you have a multiple random augmentations to apply in a sequence, you can generate the random signs of magnitudes once for all stages, rather than do it for every stage separately.Additional information:
Affected modules and functionalities:
AutoAugment (points 1., 2.), RandAugment (point 2.).
Key points relevant for the review:
Tests:
Checklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: DALI-3357