[Feature] Add SRFolderMultipleGTDataset #355
Conversation
Codecov Report
@@ Coverage Diff @@
## master #355 +/- ##
==========================================
+ Coverage 80.11% 80.44% +0.32%
==========================================
Files 181 183 +2
Lines 9496 9760 +264
Branches 1385 1417 +32
==========================================
+ Hits 7608 7851 +243
- Misses 1698 1717 +19
- Partials 190 192 +2
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
|
Can the |
|
Not sure if |
|
|
How about After this PR, there will be another named |
|
|
||
| @DATASETS.register_module() | ||
| class SRCustomMultipleGTDataset(BaseSRDataset): | ||
| """Custom dataset for video super resolution for recurrent networks. |
There was a problem hiding this comment.
Docstring failed to explain Multiple.
Besides, This sentence is a bit difficult to understand (... for video super resolution for recurrent networks.)
There was a problem hiding this comment.
It is explained in line 15: The dataset loads several LQ (Low-Quality) frames and GT (Ground-Truth) frames.. There are multiple GT frames.
| pipeline (list[dict | callable]): A sequence of data transformations. | ||
| scale (int): Upsampling scale ratio. | ||
| num_input_frames (None | int): The number of frames per iteration. | ||
| If None, the whole clip are extracted. If it is an integer |
There was a problem hiding this comment.
If None, the whole clip will be extracted.
| pipeline (list[dict | callable]): A sequence of data transformations. | ||
| scale (int): Upsampling scale ratio. | ||
| num_input_frames (None | int): The number of frames per iteration. | ||
| If None, the whole clip are extracted. If it is an integer |
There was a problem hiding this comment.
please delete an integer.
| gt_path=self.gt_folder, | ||
| key=sequence.replace(f'{self.lq_folder}/', ''), | ||
| num_input_frames=num_input_frames, | ||
| sequence_length=int(sequence_length))) |
There was a problem hiding this comment.
sequence_length = len(glob.glob(osp.join(sequence, '*.png')))
len(__obj: Sized) -> int
No need to use int(sequence_length)
ckkelvinchan
changed the title
| @@ -27,6 +28,10 @@ class SRTestMultipleGTDataset(BaseSRDataset): | |||
| def __init__(self, lq_folder, gt_folder, pipeline, scale, test_mode=True): | |||
| super().__init__(pipeline, scale, test_mode) | |||
|
|
|||
| warnings.warn('"SRTestMultipleGTDataset" have been deprecated and ' | |||
| 'will be removed in future release. Please use ' | |||
| '"SRFolderMultipleGTDataset" instead.') | |||
There was a problem hiding this comment.
Details see https://github.com/open-mmlab/mmediting/pull/355
* Add SRCutomMultipleGTDataset * Rename and minor updates * Fix * Add dataset * Modify warnings for SRTestMultipleGTDataset
Motivation
Currently, MMEditing does not support training video SR models with custom datasets. Users can only trained on pre-determined datasets such as REDS and Vimeo-90K.
Modification
This PR adds support to training recurrent networks (e.g. BasicVSR, IconVSR) on custom datasets. Changes are as follows:
SRFolderMultipleGTDataset, which allows users to train and test on their own datasets.SRTestMultipleGTDatasetis deprecated. All existing uses ofSRTestMultipleGTDatasetare replaced withSRFolderMultipleGTDataset, but we still keepSRTestMultipleGTDatasetto avoid any unexpected compatibility issue. It can be removed in the future.