This will hopefully be the only big breaking change we need to do. When I originally came up with this format I thought it would be a good idea to separate taskData and taskOutput to make it clear what work was performed and what the input data was. In practice, I don't see a need for this distinction and it makes indexing a pain. This change makes the JSON more consistent with the CSV representation and the representation of other libraries by removing taskData and taskOutput and creating a samples array which contains both input data and output data.
"output" is not quite a perfect name, but neither is "label". "annotations" is slightly more fitting but confusing in the case where image classifications have string as an output. So in short, I think "output" is general enough to cover all the cases.
This will hopefully be the only big breaking change we need to do. When I originally came up with this format I thought it would be a good idea to separate
taskData
andtaskOutput
to make it clear what work was performed and what the input data was. In practice, I don't see a need for this distinction and it makes indexing a pain. This change makes the JSON more consistent with the CSV representation and the representation of other libraries by removingtaskData
andtaskOutput
and creating asamples
array which contains both input data and output data."output" is not quite a perfect name, but neither is "label". "annotations" is slightly more fitting but confusing in the case where image classifications have string as an output. So in short, I think "output" is general enough to cover all the cases.