Add Custom Dataset Training Support #154
Conversation
…ib into feature/data/btad
samet-akcay
changed the title
README.md
Outdated
| task: segmentation # classification or segmentation | ||
| mask: <path/to/mask/annotations> #optional | ||
| extensions: null | ||
| split_ratio: 0.2 |
There was a problem hiding this comment.
Maybe add some comments here to the parameters that may be hard to understand. e.g.
split_ratio: 0.2 # ratio of the normal images that will be used to create a test split
README.md
Outdated
| It is also possible to train on a custom dataset. To do so, `data` section in `config.yaml` is to be modified as follows: | ||
| ```yaml | ||
| dataset: | ||
| name: custom |
There was a problem hiding this comment.
Maybe we should use format here instead of name. For MVTec we also have a format field in addition to name. The way I see it, format determines which dataset class is used under the hood, while name can be anything that identifies the specific dataset that is used.
anomalib/config/config.py
Outdated
| @@ -177,7 +177,8 @@ def get_configurable_parameters( | |||
| config = update_input_size_config(config) | |||
|
|
|||
| # Project Configs | |||
| project_path = Path(config.project.path) / config.model.name / config.dataset.name / config.dataset.category | |||
| category = config.dataset.category if "category" in config.dataset.keys() else "" | |||
There was a problem hiding this comment.
Maybe it would be a bit more clear if we check the dataset type here, and only add the category to the path if the type is MVTec.
anomalib/data/custom.py
Outdated
| return samples | ||
|
|
||
|
|
||
| class CustomDataset(Dataset): |
There was a problem hiding this comment.
I'm not sure about the naming. Maybe FolderDataset would be more appropriate? Custom sounds a bit like users can choose their own 'custom' format. But this class represents a dataset that follows a fixed format based on the folder structure of the data.
anomalib/data/custom.py
Outdated
| The dataset expects that mask annotation filenames must be same as the original filename. | ||
| To show an example, we therefore need to modify the mask filenames in MVTec dataset. | ||
|
|
||
| >>> # Rename MVTec mask annotations so that they are the same as image filanames |
There was a problem hiding this comment.
I'm afraid the example in the docstring might cause some confusion with the users (why use the custom dataset class for MVTec if there is a dataset class specific for mvtec). Maybe we could keep it simple and start the example with the assumption that the user has a folder of normal images and a folder of abnormal images, and explicitly state this at the beginning of the example.
…o feature/data/custom-dataset
…o feature/data/custom-dataset
Description
This PR adds custom dataset support.
Fixes Support training with custom MVTec like dataset but without masks (ground truths) #147
Changes
Checklist