hello, can someone tell me what’s wrong? I used the exact same model (as confirmed by model.summary()) and I can reach an accuracy higher than 34.5%.
moreover, all the bboxes seem to start at the top left corner I don’t know why this is happening.
I am using BATCH_SIZE 32 if that helps. (Even in an earlier cell it’s defined to be 64)
Things to look out at this point woulb be:
-
Make sure you ate using mobilnet with imagenet weights.
-
Make sure your bounding_box_regression outputs numbers.
-
Make sure you are using the suggested optimizer and loss for the overall model.
Try the batch size of 64.
I have the same issue. It’s stubbornly stuck at 44.2% and I can’t see why. Batch is 64 and I’ve used a learning rate scheduler to get that. It’s not clear from the instructions what other things we can alter without messing up the grader unfortunately. I was tempted to go for image augmentation for example.
Finally got it but it’s barely over the threshold. For anyone else looking, the hint I’ll give - assuming no other mistakes are present - is consider a learning rate schedule of some kind with SGD.
1 Like
Are you using any kind of activation in the bounding_box_regression layer?
since I remember this assignment es really well guided you don’t need to do a lot of things outside the instructions.
Maybe the problem is in the steps_per_epoch and validation_steps calculation, just be sure that you are using all the data in every step, taking account your batch size.