TF C2W2 assignment pitfall for M1 Macbook user!

I am sharing some personal experience here hopefully others won’t run into
similar experience as myself.

Due to the assignment needs to be completed on Colab but free T4 access
is limited so I thought why not testing it out locally on my M1 Macbook until
I get good results then upload to Colab to complete testing etc.

Boy oh boy was I so wrong!!! I struggled over few days but failed to get
80% training result let alone the validation result which needs to exceed
training result by certain anount.

My “local” M1 training result will never reach 80%, led me to do all sort of
silly things like setting augmentation options to some very small value, ie
0.05 and my dense layer to 2048 etc etc just to get over the line.

Low and behold I was all over the moon and went ahead with submission
and get a reality check of those “hacks” do not meet the passing criteria.

Out of frustration and I took it to Kaggle and started all over again and
after a few tweaks the settings that never yields 80% on M1 Mac came to
a pleasant surprise to me! That’s after wasted ffew days and missed the
submission deadline!

Finally today after waiting patiently enough for Colab free T4 availability,
copy the same notebook that worked on Kaggle and viola got the similar
results needed to pass the assignment!!!

Man o man … don’t ever make the same mistake as me to try to do as
much locally on the M1 Mac to make it work!! Just didnt work out for me.

However, I am keen to hear if other M1 Mac users have any success or not completing the assignment locally on the Mac!!!


1 Like

Sorry to hear about your experience on M1. If the performance varies to the extent you’ve shared, please create an issue on the tensorflow repository as there might be a bug in M1 specific code.