Ersi lig 3910 update mae benchmark code #1468
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
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## ersi-lig-3912-refactor-mae-to-use-timm-vit #1468 +/- ##
==============================================================================
- Coverage 84.61% 84.37% -0.25%
==============================================================================
Files 136 134 -2
Lines 5799 5690 -109
==============================================================================
- Hits 4907 4801 -106
+ Misses 892 889 -3 ☔ View full report in Codecov by Sentry. |
…ig-3910-update-mae-benchmark-code
|
Imagenet |
ersi-lightly
changed the base branch from
master
to
ersi-lig-3912-refactor-mae-to-use-timm-vit
guarin
reviewed
Jan 17, 2024
benchmarks/imagenet/vitb16/mae.py
Outdated
| @@ -21,24 +22,25 @@ def __init__(self, batch_size_per_device: int, num_classes: int) -> None: | |||
| self.batch_size_per_device = batch_size_per_device | |||
|
|
|||
| decoder_dim = 512 | |||
| vit = vit_b_16() | |||
| vit = vit_base_patch16_224(dynamic_img_size=True) | |||
There was a problem hiding this comment.
Do we have to set dynamic_img_size=True? I think MAE uses 16x16 patches with 224px images everywhere.
benchmarks/imagenet/vitb16/mae.py
Outdated
| self.patch_size = vit.patch_size | ||
| self.sequence_length = vit.seq_length | ||
| self.patch_size = vit.patch_embed.patch_size[0] | ||
| self.sequence_length = vit.patch_embed.num_patches + 1 |
There was a problem hiding this comment.
Suggested change
| self.sequence_length = vit.patch_embed.num_patches + 1 | |
| self.sequence_length = vit.patch_embed.num_patches + vit.num_prefix_tokens |
benchmarks/imagenet/vitb16/mae.py
Outdated
| self.decoder = masked_autoencoder_timm.MAEDecoder( | ||
| num_patches=vit.patch_embed.num_patches, | ||
| patch_size=self.patch_size, | ||
| in_chans=3, |
There was a problem hiding this comment.
Let's not set in_chans here, it should always be 3 for RGB images.
* added hackathon * changed comments * formatted * addressed comments * fixed typing * addressed comments * added pre-norm and fixed arguments * added masked vision transformer with Torchvision * weight initialization * cleanup * modifies imagenette benchmark * made mask token optional and adapted benchmarks * removed unused import * adapted to dynamic image size * moved positional embed init to utils * updated benchmark * adapted benchmark * moved mask token to decoder * revert example * removed example * removed file * inheriting from Module * reverted dataset paths * use timm's drop_path_rate * removed unused import * removed private method * changed slicing * formatted * path dropout only for fine tune * formatted * account for mask token in backbone * mask token of decoder * removed appending of mask token in params
ersi-lightly
merged commit cc263fe
into
ersi-lig-3912-refactor-mae-to-use-timm-vit
4 of 8 checks passed
ersi-lightly
added a commit
that referenced
this pull request
Mar 6, 2024
* Add MAE evaluation * Add stochastic depth dropout * Add MAE * Drop assertion * Fix smooth cross entropy loss and mixup * Update comments * Add layer lr decay and weight decay * Update comment * Add test for MAE images_to_tokens * Disable BN update * Add BN before classification head * Format * Fix BN freezing * Cleanup * Use torch.no_grad instead of deactivating gradients manually * This is required as torch.no_grad doesn't change the model configuration while manual gradient deactivation/activation can have unintended consequences. For example, MAE ViT positional embeddings are parameters with requires_grad=False that should never receive an update. But if we use activate_requires_grad for finetuning we break those parameters. * Create new stochastic depth instances * Add mask token to learnable params * Add sine-cosine positional embedding * Initialize parameters as in paper * Fix types * Format * adjusted to existing interface * draft * remove * added modifications * added mae implementation with timm and example * formatted * fixed import * removed * fixed typing * addressed comments * fixed typing and formatted * addressed comments * added docstring and formatted * removed images to tokens method * Ersi lig 3910 update mae benchmark code (#1468) * modified imagenette benchmark * formatted * edited vitb16 benchmark * added the posibility to handle images of different sizes * formatted * removed comments * revert * changed import * initialize class token * specified that class token should be used * chabged architecture * addressed comments * formatted * Masked vision transformer (#1482) * added hackathon * changed comments * formatted * addressed comments * fixed typing * addressed comments * added pre-norm and fixed arguments * added masked vision transformer with Torchvision * weight initialization * cleanup * modifies imagenette benchmark * made mask token optional and adapted benchmarks * removed unused import * adapted to dynamic image size * moved positional embed init to utils * updated benchmark * adapted benchmark * moved mask token to decoder * revert example * removed example * removed file * inheriting from Module * reverted dataset paths * use timm's drop_path_rate * removed unused import * removed private method * changed slicing * formatted * path dropout only for fine tune * formatted * account for mask token in backbone * mask token of decoder * removed appending of mask token in params * resolved conflicts * formatted * adjusted examples * removed comment * added test * added message in case of ImportError * fixed skipping of test * removed example * handling the TIMM dependency * added note to docs for MAE installation * added unit tests for MAE with torchvision * removed unecessary maks token definition * addressed comments * moved test to separate file * added typing * fixed import * fixes typing * fixed typing * fixed typing * Ersi lig 4471 cleanup and merge mae branch (#1510) * renamed test class * fixed imports * ficed imports * fixed import * fixed imports and decreased batch size * format * removed comments * use function defined in utils * added docstrings * added doctrings * added docstring * formatted * formatted * import Tensor --------- Co-authored-by: guarin <[email protected]>
Closed
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.
Updates the MAE benchmark code to use the new TIMM ViT backbone.