0

I have seen that Google's AutoML will use some pre-training image augmentation to increase the robustness of the model. (Adjacent discussion)

I have searched the documentation and forums for a way to limit these techniques. For instance it applies flips to the objects. However, in some cases flips hurt the predictions. For instance recognizing numbers in an image. For most fonts, 2's and 5's are different enough to have different features, even when flipped. However, a 7-segment display will have the same representation for 2's and 5's when they are flipped. 7-segment display example

I have labeled hundreds of images with many digits in each image. The model continues to confuse the 2's and 5's for the 7-segment displays. It has some success but not an acceptable amount.

Does anyone know if limiting the image augmentation with AutoML is possible?

Pit
  • 736
  • 3
  • 17
Taylor Hutyra
  • 1
  • 1
  • The best way to solve poor data augmentation issues would be to change and improve your training data, which will usually end up in a better model. If you can't modify your training data, then you could specify the aug_scale_min and aug_scale_max arguments when you submit the training job [1](https://cloud.google.com/ai-platform/training/docs/algorithms/object-detection-start#submit_a_training_job). – recodermoreno Jul 22 '21 at 13:00
  • How would someone go about doing this? As far as I've seen in the documentation, the trainingjob.definition.AutoMlImageObjectDetectionInputs function only takes 3 inputs which are which model type (high accuracy or low latency, or edge), budget, and early stopping. – Taylor Hutyra Jul 23 '21 at 14:30
  • You would have to modify the arguments specific to built-in image object detection algorithm when you [submit the training job](https://cloud.google.com/ai-platform/training/docs/algorithms/object-detection-start?hl=sr_Latn#submit_a_training_job). However, the best way to avoid automatic data augmentation is improving your training data, evading issues like insufficient data and/or incomplete or imperfect labels. – recodermoreno Jul 30 '21 at 12:14

0 Answers0