I fully understand why you have scaled the training/validation in the information session video for demonstation purposes.
But can I please get clarification whether this scaling down will also take place prior for our submitted models when you apply them on the private image dataset?
(i.e. can I assume the original resolution of the image files will be used and if not what width/height will be used?)
From my understanding, when you are testing your entry in the notebook, it is called with the training set of images, and these are scaled to the size you specify in your code. When you submit your entry, it will be called in the same manner but with the validation set of images, and these will be similarly scaled as your exact same code is running. So you do not need to be concerned with the resolution of the validation images, as they will be treated exactly the same way as the training ones.
The main trade off with scaling is that some information will be lost so after a point your model might be less accurate as features become indistinguishable. But you will be able to train and iterate on your model faster if you use a smaller size, which might work out better overall. One major consideration might be if you want to use an existing model as a starting point. Five of the top six entries from last year's competition scaled the images to 224x224, as required for MobileNet, ResNet etc.
Posted by: redchillipadi @ April 2, 2022, 12:47 a.m.Hi fbekema,
The scalling used during training needs to be the same at test time. This is why in the code, the image_size variable is also stored with the model.
Hence, during test time, the image_size variable is used again to process images so that they match the input size of the model. Check line 52 the deakin_ai_challenge_submission.py.
I hope this helps you. Please let me know if you have further questions.
cheers,
reda