Does platform.ai work with unlabelled text data? Or just images as of now?
It’s only images currently.
Is Platform.ai open source?
No but it is free for public projects, like Github.
Nicely done @arshak. Looks very cool.
it’s not open source but it is free to use for public projects (similar to Github)
thanks @gamino! we have a lot of work to do to improve usability / stability but eager to have folks use it and give us feedback.
Is it comparable to Prodi.gy by the spacy people?
Cool to see it live in today’s class!
From the website, it says:
No need for … crowd-workers to label your data.
All the best.
Our focus is on batch labeling dozens maybe hundreds of examples at a time and learning from every aspect of the human interaction (not just the explicit label).
We’re trying to make it easy for domain experts to do their own labeling / model training, without the need to outsource to crowd workers. From my past experience, outsourcing can be very expensive and time consuming and crowd workers seldom have the necessary expertise to provide accurate labels (particularly in technical domains).
Thanks for this, I have the perfect example to try it out.
I have a few thousand of Architectural buildings, do you think Imagenet would be able to ‘understand’ this dataset? Is the platform learning itself from the input data we use in it? Is it also going to use DL models trained on it to label images in the future?
Nice. Will definitively check it out…
Congrats on the launch!! This is definitely a very interesting product and I see many applications to use by domain experts, but therefore the question:
Do you plan on making this available as on-site solution for customers as well or will it stay an online solution?
The reason I am asking is that the data itself is a big asset for some companies, they will probably not want to upload their data to “some server somewhere”, then there are also issues of data privacy and also areas in which contracts/laws explicitly prohibit sharing the data…
Under pricing, it indicates that the free tier makes data public, but there is a paid tier @$100/mo that keeps data and models private. There is also a ‘let’s talk’ option for on-prem use.
We’d love to work with you on this project (and if necessary make modifications to Platform to meet your needs). What aspect of the buildings are you trying to capture? Are these aerial pictures or building fronts?
Platform starts with an ImageNet trained network when you load your first batch of images, but after you’ve labeled a few dozen and have trained a model, this custom model becomes the basis for future projections (and labeling activity). For most tasks this makes the labeling easier as you go.
Hi Marc, great question! We initially expect the model training to occur through our SaaS app. Enterprise customers can get inference code from us and do predictions on-premise. This allows enterprise customers to only upload a few hundred images (per class) to our app. Since the vast majority of the data will be unlabeled and therefore, can remain on-premise.
Sometime in the future, we may consider running the entire application on-premise, if enough enterprise customers express an interest.
I was wondering if you had considered allowing folders to be added so you could drag the images into a folder and they would disappear from the images that need sorted. Maybe that’s already how it works but I couldn’t quite figure that piece out. I think click and dragging into the folders and those images disappearing would be cool to see. Great work so far though. I could see this making image labeling much less painful!