mlassitdatalabel.md

Machine learning assisted data labelling

Image multi-class and bounding box labelling

Architecture

Steps

• Copy Images to Blob location
• Resize the images - https://github.com/balakreshnan/customvisionai2020/blob/master/ReSize/resizeimages.md
• Create a GPU cluster
• Create a Data Set inside Azure machine learning service
• Create a Data label project
• Do labeling - serve as input to ML labelling
• Run labelling
• Validate output
• Conclustion

Step 1

First lets copy the data to blob storage for further processing

Step 2

Apply resize of images using the above URL https://github.com/balakreshnan/customvisionai2020/blob/master/ReSize/resizeimages.md

Step 3

Now create a new cluster for training. Select GPU since ML Assit works with GPU only as of writing to this document.

Step 4

Now time to creaet a dataset First click on Dataset and then click Create new data set

Now create data store where the file are saved:

Select the Path

Confirm the details

Step 5

Create a Data label project

Select the existing data set created above

Enable Incremental refresh

Add labels

Provide label instruction if available for labellers

Select ML Assit labelling

Click Create project

Now go to project and you will land in home page

Experiments section should show the details

• First step go to Label data
• Go through pictures and create bounding boxes
• Validate the bounding boxes.
• The Tool will show how many pictures to do bounding boxes
• Once bounding box is done then we back to home screen and wait.
• ML will automatically start to run based on what was labelled. Usualy takes time to do this task.

The tool will guide and lets us know how much more it needed to label. Usually it needs per label 25 images for 100 images.

Once the experiment is completed then we can back and validate what ML as done and then draw boxes where is it missing.

Also check the training cluster details

Training:

Validations:

inference:

Conclusion

More to come.