The custom vision documentation contains an example of how to create and train an image classification model then use it to make predictions, all in Python. It would be useful to be able to retrain already existing models using additional images in Python, as this can currently be done when viewing projects on the website but when attempting to retrain via Python just creates a new project with only the new images.
There is also the issue that when training, training can only be completed when there are images present for all tags created, otherwise the error "CustomVisionErrorException: Not enough images per tag for training" is returned.

Hi ,

We have tried the other options of exporting the Custom Vision Model

It appears that , when models are exported from Custom Vision Projects , those may be built in the older versions

For Azure ML :

We are unable to find the steps shared in the zip file
I am able to create the image with latest version of Azure ML but while publishing as a service model crashes as this could be due to models created in older version.

For Tensorflow:

Though I have updated it to the latest version Tensorflow 2.0 , the documentation shared for both Classification & Object detection projects does not work.
This has been posted in another thread.

38490

For ONNX:

Though the image is created , it crashes while publishing as the webservices, This could be due to similar reasons ,

Best Regards,
Swati

I'd like to use the exported model w/OpenCV DNN. On edge devices, the OpenCV DNN is much faster and less processor intensive than the tensorflow lib. The OpenCV DNN works with TensorFlow models from other trainers, and I think it'll work with the Azure Custom Vision models, but they need some tweaks. This repo: https://github.com/opencv/opencv/wiki/TensorFlow-Object-Detection-API has a converter tool but requires the config parameters used for training. The repo has a bunch of sample config files (https://github.com/tensorflow/models/tree/master/research/object_detection/samples/configs), but it's not clear which config settings the Custom Vision trainer uses. Could you provide the config file/settings in your github repo or as part of the ZIP file generated by the Export process? Thanks.

When using Advanced Training, do iterations build on each other or does each iteration get trained starting from scratch? E.g. if I train Iteration 1 for 10 hours, and then train Iteration 2 for 1 hour with exactly the same training data, will Iteration 2 have better or worse results than Iteration 1? Assuming the iterations do build on each other, will Iteration 2 be equally good as if I had just done a single 11-hour iteration, or is there a fixed overhead that happens in order to start an iteration? (E.g. if the iterations build on each other but there is a 0.5 hour overhead to start each iteration, then Iteration 2 would be the same as a single iteration trained for 10.5 hours).

UI Wishlist:

1. After tagging, the user is returned to the first page, every time. Complete nightmare if you are working through a batch of 1000s.




2. Permissions: Users should be restricted in their ability to download / train / add / delete images - for audit purposes it needs to be possible to restrict to tagging only.




3. Images should be tagged automatically with user name. At the moment we waste one tag per username by way of a nuisance parameter (with associated training expense too).




4. Cross-checking: If one user tags an image, they have no handle by which to refer to it when corroborating with other users. Trying to find the exact image for a second pair of eyes is pot luck.




5. File metadata alongside each image e.g. name, id, anything would be extremely useful.




6. How about a quick "Test this image I am trying to tag" button?