[question] dispaly of wrong vlaue at max_features

[prabhjot26-jangli]

Enter the chapter number

chapter 2 End to end project

Enter the page number

page 92

What is the cell's number in the notebook

In the output of grid_search.best_params_

Enter the environment you are using to run the notebook

Jupyter on Windows

Question

here the outpur says : geo__n__clusters = 15, max_features = 6 , but after this output it is written that n_clusters to 15 and setting mex_features to 8 . why ? wont we say 6 ?

can anyone please clarify this
Email: prabhjotsohal26@gmail.com

Thank you

[ageron]

Hi @prabhjot26-jangli ,

Thanks for your question. Reproducibility is hard in Machine Learning because:

• Tiny differences across platforms will compound during training, so results may vary across platforms.
• Libraries change regularly, and again, even tiny changes compound during training.
• Many algorithms rely on randomness, so we must set the random seed, however this is not always guaranteed to yield the exact same sequence of random numbers across platforms and library versions.
• When computations are parallelized (e.g., multiprocessing or multiple threads, or any operation on a GPU), the order of operations isn't always guaranteed (it depends on the implementation), so even things like adding all the numbers in an array will not give the exact same result every time, due to floating point errors. The difference is tiny, but it can compound during training. For example:

>>> 1/5 + 1/7 + 1/11
0.4337662337662338
>>> 1/11 + 1/5 + 1/7
0.43376623376623374

• The datasets are sometimes cleaned up or reordered, or it is ported to a new format and there are tiny differences due to floating point errors. Again, the errors compound during training.
• Other sources of non-deterministic behavior are possible, for example if your algorithm loads a list of files from a directory, the order is not guaranteed (but you can sort the files by name to get deterministic behavior).

So in short, when I ran the code, I got the results shown in the book, but since then the results have changed for one of the reasons listed above. I did my best to make the notebooks perfectly reproducible by setting the seed everywhere, but changes in libraries, datasets, and platforms break reproducibility every few months, it's tough.

That said, you will generally get results in the same ballpark, except perhaps with "unstable" algorithms like GANs, Reinforcement Learning, and RNNs, where even a tiny little change can sometimes break everything.

I hope this helps.