The purpose of the “Input features” checkboxes on the learning blocks isn’t clear to me, as I can try to manually de-select them, but they always get re-checked (i.e. every learning block depends on every available processing block) whenever I click on “Save Impulse.”
This is a problem for blocks such as Transfer Learning, which (understandably) expects to be fed from only a single Image block. When other kinds of processing blocks, or more than one Image block, are present in the project, the TL block reports an error and cannot be trained. I am working on a hybrid project where I’d like to use a TL block to learn some output features and a different pipeline to learn others, but this apparent bug in the Impulse Design interface means that I have to add and train the Image and TL blocks before any others; after I add any other blocks to the design, the TL block cannot be re-trained until I delete everything else.
Logically, it would make sense for the user to be able to use the Input features checkboxes to configure which processing stages feed which learning blocks. This would make it possible to use different Image blocks to feed different TL blocks, and to add components to a design without affecting the ability to re-train other blocks. As an alternative workaround, the TL block could be re-specified to use only the first Image block in the project (or, to be stricter, work as long as there is one and only one Image block present), rather than generating an error whenever processing blocks of other (non-Image) types are present, as it does now.