[GCP] Train an Image AutoML Model

For the past one and half months, I was a member of an AI development team. Our team collaborated and trained ML models on GCP, and finally won a prize in an AI competition held by ITRI (Industrial Technology Research Institute) and MOL (Ministry of Labor) last week. Therefore, I will share a series of GCP tips that we found very useful with you!

Here with steps that we trained image AutoML models:

1. Go to GCP AI Platform (Unified) or Cloud Vision

2. Click on Dataset > “Create”

3. Import your training data from your laptop or from CSV

   1. From CSV: You need to upload images to Cloud Storage beforehand so that files can be found through assigned paths in the CSV file.

   2. From laptop: It's the easiest way! You can first group different images in separate folders named by their own labels. Then, zip all the folders and upload it to Cloud Vision. After all the files have been imported, they will be well placed under their labels (their folder names). P.S. I failed to use this way on AI Platform (Unified), which may be fixed in future. Based on my last try, you can upload images from only one laptop folder at a time.

4. After you check the status in the Overview page, then click on “Training New Model”. You can select whether the model will be cloud-based or edge-used, and start training!

5. After training is done, you will receive an email notice. Then, just to check all the details in the Evaluation page, including accuracy. If everything goes well, you can deploy the model and start further testing.

Message or email me if you have any further questions. I will be happy to share my experience with you!

Thanks and have a lovely day!