Performance differences between local and Jupyter

Hi guys,

I am in the hw of “Transfer_learning_with_MobileNet_v1”, after the line

initial_epochs = 5
history =, validation_data=validation_dataset, epochs=initial_epochs)

I saw performance differences between my local PC and online Jupyter. I copied and pasted all the code to my local machine. After running the training many times, I found out that, for my PC, val_accuracy starts from about 0.5-0.7. but online Jupyter starts from 0.8 even higher. Is this telling me something is wrong for the local? If yes, what should I check ? Can someone give me a hint ?

Btw, Test failed for alpaca_model in my local (online Jupyter is passed). I think this is just a TF version problem, not the code (I guess…). It says

Test failed 
 Expected value 
 ['TensorFlowOpLayer', [(None, 160, 160, 3)], 0] 
 does not match the input value: 
 ['TFOpLambda', (None, 160, 160, 3), 0]

Most likely you’re using different versions of the tools and libraries.

Hi TMosh, so, if I use different version of TF, I will get a different performance? I think the performance should be the same, right?

Ideally, the version of TF you use would not matter.

Unfortunately, in practice TF doesn’t care very much about backward compatibility. They often change items like the order of the operands, the default values of arguments, or even the workflow inside a function.

Another consideration may be that if a pre-trained model is being used, it might not apply to a new version of TF because the training data may be obsolete or not work correctly with a different model.

1 Like