Upload dataset images failed AutoML Google Vision

Question

Having issues uploading dataset images to AutoML. I am choosing to upload images from my pc. I have attempted upload multiple and single images, ensuring the correct format. I have tried creating new cloud storage containers. I receive the following error details:

Operation ID: projects/362703496160/locations/us-central1/operations/ICN8572734381850886144 Error Messages: Error: INVALID_ROW: Invalid input found at row 1 of gs://cloud-automl-tenant-gcs-868eb49b-dec0-4c63-a25c-659fa5ffdd76/create-csv-1065568604314927104-2020-06-02T21:32:31.819Z.csv: "Unsupported file extension." Error: CORRUPTED_FILE: CSV file is empty or contains no valid row.

I am not sure why csv files are coming into play as the upload option is images from my PC.

Finally, when I view the cloud storage container the images are there and a url is also listed for the images. I can access them and view them, but they do not upload and display in Auto ML Vision images under datasets.

Any ideas???!!

Answer 1

To use the importData method, both the CSV file and the images it points to must be in a Google Cloud Storage bucket.

Additionally, the CSV file must also fulfill the following requirements

https://cloud.google.com/vision/automl/object-detection/docs/csv-format

https://cloud.google.com/vision/automl/docs/create-datasets

Answer 2

After further review of the documentation it does state that the images must be in.zip format. I was able to upload to the bucket without problem once I zipped the image directory.

Something worth noting:

If using Google Vision and uploading images for AutoML datasets, I suggest first, uploading the images to a storage bucket via google storage. Then, exporting the xml of the bucket directory and creating a cvs file for Vision dataset upload. The cvs format allows for the ability to assign the Training, Validation and Test classifications for each object(label). The documentation says that a minimum requirement of 8-1-1 (Training-Validation-Test) is required for each object to be identified successfully. So, if you just upload the zipped images to the bucket via the Google Vision UI you cannot assign classifications to your images, only labels. One can upload more than 30 images on one object(label) and the system does not automatically assign the images to Training-Validation-Test classifications to achieve the minimum requirements of 8-1-1. If you do not achieve this minimum requirement of 8 images for Training, 1 image for Validation and 1 image for Test you cannot begin to training your model. To save a lot of time and/or if you want to get a simple model training quickly, the csv format is essential.