As discussed in the meeting, here are some issues identified while teaching the BMZ (please feel free to modify or add more things):
Searching engine: it works well suggesting existing tags but it doesn't work with capital letters or spaces. For example, bacillus-subtilis, bacteria, works but bacteria Bacillus doesn't. Can we get (more) fuzzy search?
When newcomers explore the model card, it's not clear what each of the icons represents. For example, notebooks are given as puzzle pieces and when you place the cursor on top of it, it doesn't say notebook but "ZeroCostDL4Mic XXXX". In general, is not very homogeneous which creates confusion. An idea could be to create a "card dictionary" for example that says what is each of the icons.
The most common question so far is "What do I have to do to upload a model?". What do I need? What is there that I can use to upload a model?
It feels that we need to fill the gap.
Proposal from Esti: Have three VERY BASIC notebooks (tensorflow, ONNX, Pytorch) that shows how to export a model that was previously stored and get the zip file. Then, we can have the video/doc for the upload that in theory can be changed quickly whenever the system changes.
The first entry to the zoo is not straightforward for newcomers and we should assume that they don't have general knowledge, nor they are good at using new technologies. For example, users, when they speak about DL they called everything in the zoo "models", i.e., for them a notebook is the same as a model.
In general, newcomers have the motivation to use DL and more or less they know the task to perform (e.g., segmentation of nuclei). They are also aware of the need for something called "training" that requires a lot of annotated data, but they are completely naive to any further details, including the idea of "fine-tuning" or "retraining".
Hi hi,
As discussed in the meeting, here are some issues identified while teaching the BMZ (please feel free to modify or add more things):
bacillus-subtilis
,bacteria
, works butbacteria Bacillus
doesn't. Can we get (more) fuzzy search?@oeway @fjug @akreshuk @constantinpape @FynnBe @cfusterbarcelo