Fix: Adjusted PiecewiseConstantDecay schedule to match training steps on Fashion MNIST

[ShindeShivam]

Chapter: 11
Cell: Piecewise Constant Scheduling

Issue:

In the PiecewiseConstantDecay example, the original boundaries were set to [50000, 80000], but with 55,000 training images, batch size 32, and 10 epochs, total training steps are only about:

55000 / 32 × 10 ≈ 17,190 steps

The model never reaches the first boundary, and the learning rate never decays.

So,I changed:
boundaries=[50_000, 80_000],

to:
boundaries=[6000, 13_000],

Hope this PR might be meaningful.

I’ve also opened a few other PRs recently ->

#202
#205

[ageron]

Oh great catch, thanks! I'll look at the implications for the book, then I'll merge. 👍

[ageron]

It looks like there's no reference in the book, that's a relief. However, readers might not understand why we're doing this, so could you please use this code instead?

boundaries=[50_000 * n_epochs // batch_size, 80_000 * n_epochs // batch_size],

[ShindeShivam]

My Mac is currently in for repair, so I’m updating this directly from my phone. This is my first time pushing changes from mobile.
Hope this works.

[ageron]

Thanks again @ShindeShivam