We’re fixing the upload feature so that it allows nested folder, it’s on our roadmap.
In the meantime here is a workaround, to make the directory flat. This will add the directory name into the filenames and flatten it out.
Make a copy of your directory if you want to keep the original organization (this deletes the folders)
pax -Xrwls '|/|_|g' / "$PWD"That is a terrifying command. Isn’t that removing the entire structure from the root directory? Maybe I am totally misunderstanding what this whole work around would do.
Any chance you could add an option to toggle seeing images already labeled? I know right now they dim, but it’d be great when iterating to make sure I don’t keep grabbing the same images.
If this already exists and I’m too dense to have seen it, I apologize in advance 
Hi Kyle, that option already exists. If you select the class, you see a color added to the border of the class and the corresponding images.
The fading is related to the certainty that the model is assigning to the prediction (not the label).
Understood - I’m talking about a toggle option to remove the image from the view rather than highlighting. Maybe I’m just an edge case, though.
I see what you mean. That’s a good idea. We’re working on a “push” based version of the app that will make the projections more dynamic. So perhaps we can remove labeled images and push other images into the projection that might interest the user (and help the model).
Been taking platform.ai for a spin. Works quickly and very useful. Thanks!
+1 on @knesgood’s suggestion for an option to hide the labeled images. Currently, selecting the class(es) to display colored borders around corresponding images doesn’t really help distinguish labeled from unlabeled images because those colored borders disappear as soon as you select an image or click on the projection area. So I still have to rely on visual memory of which images or general areas of the projection are unlabeled at the moment of selecting new images to label.
Alternatively to hiding labeled images, you could have it so the class-colored borders stay displayed while selecting new images?
Dynamically removing already labeled images and pushing new batches of unlabeled images into the projection would also be a great option.
You got it! I think we can include this in our Dec 18 release.
@arshak do you run this on a gpu by default, or do you spin up a GPU instance specifically for training?
We run a K8S cluster that manages the GPU/CPU workload depending on what the user needs.
We’ve done a lot of work to make the GPU processing efficient so that the public version of platform.ai can be free for the community.
Congrats @arshak !!! for starting up this wonderful platform with Jeremy. I am however facing issues while signing up. It says " This browser is not supported or 3rd party cookies and data may be disabled " . But I enabled cookies and my chrome is up-to-date. Please help me with the sign-up.
That’s cool. Did you guys end up using the fast.ai library or did you need to switch out to tensorflow?
Hi Vishal, thank you! We’re delighted to see the amount of interest from the community.
I believe the issue you’re experiencing is due to 3rd party cookies being disabled. We use Firebase for authentication which requires this option to be turned on:
Please let me know if that doesn’t solve it.
We use fast.ai library because it’s a good match for platform’s needs and we want to help make the library better through close collaboration with their team.
For example we’ve been researching / prototyping some additional automation for the learning rate finder, determining which augmentations work best for a given task. I hope some of these additional capabilities will end up in the library too.
Yay …!!! Thanks for the advice. I have successfully signed up. So excited to try it out now…!!!
Nice product. I think another opportunity is to build a human-machine interface for chatbots. The use case is very similar to image labelling. You give the human a head start curating dialogue from unsupervised learning and a rich user experience.
Interesting idea but we are focused on vision for the foreseeable future, so many interesting problems to solve in this area alone.
Curious what integrations folks think would be useful for platform.ai
How would you prioritize these and am I missing any other important integrations?
Im curious as to how the clustering works initially - do you guys normalize the images to imagenet and then cluster based upon pixel values?
Do you just remove the final layer of label classification and cluster upon the resultant weights?
Do you not do any clustering at all and just do a production onto a 2d space? Im curious because I uploaded like 6 images, and it auto clustered them according to the label I was going to assign.
We can’t fully disclose the projection approach (secret sauce) but suffice it to say that we leverage transfer learning and various dimensionality reduction techniques.
Glad to hear that your small test matched expectations! We’re doing a lot of research behind-the-scenes to make this kind of delightful experience possible for all users.
