Course 4 Week 2 assignment 2: Fine tuning not improving accuracy

Autograder gives 100% but I think something must be wrong with my code because the accuracy doesn’t increase with fine tuning:


Any help would be much appreciated!

Many factors can affect this. First you should make sure the pipeline is connected properly, by that I mean all layers are fitted including the extension layers. Assuming all that is good and no architecture problems are present, I would say that maybe you need to fine tune - train longer so the network has more time to learn features about the data.

Fine tune longer and with a bigger fine tune dataset.

Hopefully there is no over fitting overall.

But I see here the train and validation accuracy are too low, where those the same with the original network, or only in the fine tuned? If architecture is set up properly then I would think to fine tune longer, otherwise you have an architecture problem i.e. the way you set up the new model!

Hi, I had the same problem, but noticed that I didn’t set the weights in the Exercise 2. This should help.

    base_model = ... weights='imagenet')# From imageNet

Probably, it is nice to have additional test for this, cause students could be confused with the results.