-
|
Hello, love the library! I'm trying to train an image classifier with support for multilabel. I found this thread from 3+ years ago, and @rwightman alludes to maybe adding support.
|
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
|
@davidrs so yeah, that sorta stayed in limbo, biggest issue was a lack of standardization in terms of multi-label datasets to use as canonical examples / benchmarks. I've actually just started to address this on the data side
So, the plan is to now finally support other tasks, like multi-label, also multi-task (multiple multi-label and/or multi-class heads), AND finally support object detection, segmentation tasks. It's a fair bit to take on but I'm making headway, you'll probably see a bit more on the data handling side (annotation reading & transform to inputs & targets) before I get final train & modelling code sorted as I need to data part solid to refine & test the rest... |
Beta Was this translation helpful? Give feedback.
-
|
Oh yeah, and other tasks require more metrics, so will be work needed there too for meaningful multi-label support, and the other proposed tasks require more involved metrics too. |
Beta Was this translation helpful? Give feedback.
-
|
thank you, I appreciate the update and completely get what you mean about lack of standardization on multi-label datasets. for anyone else trying to roll-their-own in the meantime the previously linked repo and article is a helpful jumping off point: |
Beta Was this translation helpful? Give feedback.
@davidrs so yeah, that sorta stayed in limbo, biggest issue was a lack of standardization in terms of multi-label datasets to use as canonical examples / benchmarks. I've actually just started to address this on the data side
So, the plan is to now finally support other tasks, like multi-label, also multi-task (multiple multi-label and/or multi-class heads), AND finally support object detection, segmentation tasks. It's a fair bit to take on but I'm making headway, you'll probably see a bit more on the data handling side (annotation reading & transform to inputs & targets) be…