问题
I have built an Azure Custom Vision model using ~ 5000 of my own domain-specific images and a set of ~ 30 hierarchical and non-hierarchical labels.
I am not sure how best to organize my label zoo in this particular multi-label classification problem. The best approach (see e.g. https://www.researchgate.net/publication/225379571_A_Tutorial_on_Multi-label_Classification_Techniques and https://towardsdatascience.com/journey-to-the-center-of-multi-label-classification-384c40229bff) must depend on the inner workings of Custom Vision, alas undocumented*. Consider for example
Image Document_Description
1 Barclays Bank Statement
2 HSBC Bank Statement
3 Joe Bloggs' Curriculum Vitae
Given the (perhaps) unknown modelling scheme(s) used by Custom Vision, and its support for arbitrary tags, which labelling taxonomy will be most efficient (in terms of training compute and model performance)?
1. Hierarchical (choose one from each level):
IsCV | IsBankStatement | IsOther | ...
|
Barclays | HSBC | ...
2. Non-hierarchical:
IsCV, IsBankStatementBarclays, IsBankStatementHSBC, IsOther, ...
3. Both
4. Some other scheme perhaps informed by insider information?
Bonus: How would you use the available performance indicators - or the V3.0 API - to measure the performance of two competing taxonomies (with minimal training compute/cost)?
*I apologise for the desperate question. Before voting to close it, please allow Azure Cognitive Services time to comment, since this seems to be about the only forum in which they might be able to give input, and they do ask for queries via SO. Thanks.
回答1:
I think custom vision only support non-hierarchical tags for now, but you can submit suggestions here https://cognitive.uservoice.com/forums/598141-custom-vision-service
来源：https://stackoverflow.com/questions/56426097/azure-cognitive-services-custom-vision-how-do-i-design-an-appropriate-multi-lab