Add ten/five crop augmentation #110
Open
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
v-iashin
requested changes
Jan 23, 2024
There was a problem hiding this comment.
Thank you for your pull request. However, it should be significantly revised before merging.
I'd be more positive about it if:
- one pull request wouldn't implement two features: one for the ten crop, the other for the changes in disk writing
- I don't like that each crop is treated as a separate video (now each video creates 10/5 times more files. can't we implement crop as a batch dimension?
- the augs are applied only to RGB stream, not both.
- the augs are implemented for i3d but not other models.
- i am not happy that the classes/functions that are common for many models are being changed without reflecting on how logic for other features that depend on them changes
i appreciate the efforts but i am not convinced that it is enough
| self.i3d_transforms = { | ||
| 'rgb': torchvision.transforms.Compose([ | ||
| TensorCenterCrop(self.central_crop_size), | ||
| aug_transform, |
There was a problem hiding this comment.
any reason why we can't do it for the flow?
|
|
||
|
|
||
| if self.aug_type is not None: | ||
| feats_dict = {stream: [[] for _ in range(self.num_crop)] for stream in self.streams} |
There was a problem hiding this comment.
Why treat each crop as a separate tensor instead of a batch dimension: B, Crops, D --> B*Crops, D?
| @@ -145,13 +145,17 @@ def __call__(self, tensor: torch.FloatTensor) -> torch.FloatTensor: | |||
|
|
|||
| class ScaleTo1_1(object): | |||
|
|
|||
| def __call__(self, tensor: torch.FloatTensor) -> torch.FloatTensor: | |||
| def __call__(self, tensor): | |||
| if isinstance(tensor, tuple): | |||
| return (2 * tensor / 255) - 1 | ||
|
|
||
|
|
||
| class PermuteAndUnsqueeze(object): | ||
|
|
||
| def __call__(self, tensor: torch.FloatTensor) -> torch.FloatTensor: | ||
| def __call__(self, tensor): | ||
| if isinstance(tensor, tuple): |
| @@ -50,9 +50,18 @@ def show_predictions_on_dataset(logits: torch.FloatTensor, dataset: Union[str, L | |||
| print(f'{logit:8.3f} | {smax:.3f} | {cls}') | |||
| print() | |||
|
|
|||
| def make_path(output_root, video_path, output_key, ext): | |||
| def make_path(output_root, video_path, output_key, ext, idx=None): | |||
There was a problem hiding this comment.
we shouldn't resort to this. it became incredibly redundant. we need to save all features in one file
| @@ -76,6 +78,11 @@ def action_on_extraction( | |||
| return | |||
|
|
|||
| for key, value in feats_dict.items(): | |||
| if self.save_option == 'rgb_only': | |||
There was a problem hiding this comment.
what's wrong with the streams argument in i3d?
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.
Add ten/five crop augmentation when extract the I3D features. Solve issue #92 #72 . And add save_option to the i3d.yaml file to save only reg features. Because I think that fps and timestamp features are really redundant. The shape of the rgb features imply timestamp.
Thank you for your great work. 馃殌馃殌