Geti classification labels messed up by OTX

[ssainis-0131]

The labels from the OTX are completely different order than what is coming in from Geti Deployment

Steps to Reproduce

1. Create a geti classification project with multiple labels (non hierarchical) and train a model
2. Take an image from the training data and pass it to the model and look a the prediction
3. The prediction will be wrong.
4. Walking through the trace of the load_inference method, you will see it is creating an _inference_conveter in the deployment object using the get_labels method on line 340 of otx/api/entities/label_schema.py

line 340 currently reads

labels = {label for group in self._groups for label in group.labels if include_empty or not `label.is_empty}

As the "set" function destroys the local order of the list it is asked to convert into a set, we get label order that is completey different from the input label list.

My proposed fix to his is

labels = [label for group in self._groups for label in group.labels if include_empty or not label.is_empty]
# remove duplicates
labels = [label for i, label in labels if label not in labels[:i]]

This should preserve the order.

I have tested this at my end. I recommend implementing it in the main code.

Environment:
Bug is independent of the environment

• OS: Windows (but irrelevant to the bug
• Framework version: none
• Python version: 3.8
• OpenVINO version: 1.2.4
• CUDA/cuDNN version: none
• GPU model and memory: none

[goodsong81]

Thank you for the report! This PR fixes the issue: #2369.
(also applied to develop branch).