Can't import dataset into Pytorch (saved locally).

[ANTUMN0S]

I'm currently trying to import a small dataset I've created from CVAT into PyTorch, so that I can use it in combination with DeepLabV3+. However it always fails, giving me the AttributeError '_OfflineTaskModel' object has no attribute 'updated_date'.

My actions before raising this issue

I've searched for the error message and gone through the documentation. Neither has mentioned anything about it.

Steps to Reproduce (for bugs)

1. Have any small dataset that is not stored on a cloud service
2. Install cvat_sdk[pytorch]
3. Set either app.cvat.ai or localhost:8080 (if you have docker) as CVAT_HOST, and username and password as CVAT_USER and CVAT_PASS
4. make the client:
   client = make_client(CVAT_HOST, credentials=(CVAT_USER, CVAT_PASS)
5. train_set = ProjectVisionDataset(client, TASK_ID)
6. get said Attribute Error

Expected Behaviour

I'd expect the dataset to properly load.

Current Behaviour

It doesn't load, because it doesn't have the Attribute 'updated_date'

Context

My Master Thesis revolves around computer vision and I also need to annotate a lot of data, that needs to be segmented later on. I was advised to use CVAT for annotation and should try to do segmentation with the DeepLabV3+ Architecture. As my data is not yet generated, I've found a small dataset that's similar to what I'll be doing and I wanted to use it as a small proof of concept. But right now, I can't load the dataset, because of that attribute error. I don't know if there is a seperate command for loading datasets that are not stored on a cloud service, but my search so far does not support that claim.

Your Environment

• To code I use VSC and Python 3.10.6
• The Docker container runs CVAT 2.4.0, because that's the version the sdk is on
• The Docker works fine, I can access and use it with chrome.
• My OS is Ubuntu 22.04 LTS

[josiahls]

Same issue, same version. I'm copying the modules I need and making the fix myself on my end.

[josiahls]

Changing that line to saved_task.api_model['updated_date'] fixes it.

If the task is a OpenApiModel (downloaded from online), then doing something like saved_task.updated_date' will work, since OpenApiModel` provides machinery to grab nested attributes.

The problem is when a model is cached, and loaded back, it is loaded as a _OfflineTaskModel

    def _deserialize_model(self, obj: _CacheObject, model_type: _ModelType) -> _ModelType:
        if issubclass(model_type, OpenApiModel):
            return cast(OpenApiModel, model_type)._new_from_openapi_data(**obj)
        elif issubclass(model_type, _CacheObjectModel):
            return cast(_CacheObjectModel, model_type).load(obj)
        else:
            raise NotImplementedError("Unexpected model type")

I think the idea is that we are using the attrs package to grab the nested attributes:

@define
class _OfflineTaskModel(_CacheObjectModel):
    api_model: models.ITaskRead
    labels: List[models.ILabel]

However this doesnt actually work in 23.1.0 or 22.2.0 versions of the package. Looking at https://www.attrs.org/en/stable/examples.html it looks like you need things like attr.s decorator and attr.ib, but even then, searching for "nested" behaviors doesn't seem to be a major feature of this package (?)