DRAG DROP -
You train a Custom Vision model used in a mobile app.
You receive 1,000 new images that do not have any associated data.
You need to use the images to retrain the model. The solution must minimize how long it takes to retrain the model.
Which three actions should you perform in the Custom Vision portal? To answer, move the appropriate actions from the list of actions to the answer area and arrange them in the correct order.
Select and Place:
Reference:
https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/getting-started-build-a-classifier
The given answer is incorrect. The question emphasizes two things - 1) the model has already been trained 2) the solution should be expedient. The given answer will be very slow to manually tag 1,000 images. instead: 1.) upload all the images 2.) Get suggested tags 3.) Review the suggestions and confirm the tags reference: https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/suggested-tags
Thank you.
When you tag images for a Custom Vision model, the service uses the latest trained iteration of the model to predict the labels of untagged images we need latest trained to predict the labels, but this isn NOT HAVE ANY ASSOCIATED DATA
Exactly. Here we need to use Smart Labeler instead. https://learn.microsoft.com/en-us/azure/ai-services/custom-vision-service/suggested-tags
Answer is correct. When uploading all images from a same folder, you can tag all of them with the same value at the same time. Then you wont tag all 1000 images one by one, but only once by category (which is time saving as the question ask for). Also, even if model is already trained, images are uploaded to workspace, and not to specific trained iteration. You then cannot get tag suggestion when importing an image. There is none, that feature simply does not exist. Try by yourself : https://learn.microsoft.com/en-us/training/modules/classify-images/5-exercise-custom-vision
my bad the feature is real : https://learn.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/suggested-tags so right answer is - Upload all - Get suggested tags - Review and confirm tags
I agree with SuperPetey. The answer should be 1.) upload all the images 2.) Get suggested tags 3.) Review the suggestions and confirm the tags Reason being that using the tools(suggested tags) would still applied to the new 1000 images item, even if those 1000 images doesn't associate with the original data pool. So, that means tagging even 1 less images using the suggested tags would still be faster than manually tagging them. https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/suggested-tags
To minimize the time required for retraining the model, the correct three steps are: Upload all images: First, you need to bulk upload the 1000 new images to the Custom Vision service. This is the foundational step for preparing the data. Get suggested tags: Utilize Custom Vision's functionality to automatically suggest tags for the uploaded images. This can significantly reduce the workload of manual tagging. Review and confirm suggested tags: Finally, manually review and confirm the tags suggested by the system to ensure their accuracy. Then, use these tagged images to retrain the model. This process leverages the automation tools provided by Custom Vision to streamline and expedite the data preparation process, particularly effective when dealing with a large number of untagged images.
@superpetey, kindly read through the article in the link you shared, I just did and confirmed from it that the provided answer by the platform is correct.
"You should only request suggested tags for images whose content has already been trained once. Don't get suggestions for a new tag that you're just beginning to train." And the question says RETRAINING of an existing model to which we are adding new images. So the response is actually wrong and @superpetey is correct
AHHHH but the key word is 'DO NOT HAVE ANY ASSOCIATED DATA'. So the content of images is brand new!!! Therefore we cant use suggester and the response is correct!
I support your highlighted point to the right point. So the given answer should be correct.
The point of machine learning is that a model eventually LEARNS how to do things independently. Even though there is no associated data, there is previous learning done and existing labels can be used. I am not sure why we would need ML if we still have to do things manually all the time?
I disagree, the images are unlabeled, but there is nothing in the text of the question mentioning that there are new tags. I agree with SuperPetey.
Answer given would be only option IF model had not already been trained with images, so... I agree with SuperPetey et al... Upload Get suggested tags Review and confirm tags
Was on exam 27/11/2021
The answer is correct - there is no magic here. You can’t suggest any new tags based on the model you currently have. Read the limitations of the smart labeler carefully: When to use Smart Labeler Keep the following limitations in mind: You should only request suggested tags for images whose tags have already been trained on once. Don't get suggestions for a new tag that you're just beginning to train.
"When you tag images for a Custom Vision model, the service uses the latest trained iteration of the model to predict the labels of new images" source: https://learn.microsoft.com/en-us/azure/ai-services/custom-vision-service/suggested-tags
Final Answer - Upload all - Get suggested tags - Review and confirm tags
1.) Upload all the images 2.) Get suggested tags 3.) Review the suggestions and confirm the tags If an image does not have any associated TAG, we can add a new one while reviewing https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/getting-started-improving-your-classifier
Well, it's a bit confusing. In both cases (ET answers and SuperPetey suggestion) - we will have to walk through the pictures manually if there is no info about them. IF they are stored in class folders - the ET answer is less time consuming, if not - it's not possible to tell if separating them manually or manual check of suggested tags will take less time.
This question was asked in the actual exam on April 30, 2024 (+9:00, Japan). I think SuperPetey's answer is CORRECT, because I passed the AI-102 exam with a score of 917/1000. Thank you very much.
So questions registered in 2021 will still be on the exam in April 2024? Japan is a scary country.
Group Upload category Tag
Can't be correct. You wanna tell me people should manually tag 1000 images? And how do you categorize them in folders when they have no association? Seems dumb It has to be 1. Upload all the Images 2. Get suggested tags 3. Review the suggested Tags and confirm
1.) upload all the images 2.) Get suggested tags 3.) Review the suggestions and confirm the tags