Activate initialEpoch for model.fit

[hpssjellis]

See feature request to activate initialEpoch at

tensorflow/tfjs#497

---

This change is 

[bileschi]

Thanks for your PR! Do you think you could add a test for this behavior, possibly by cloning and modifying the test in 'training_test' named '1 input, 1 output, dense, 1 weight, string optimizer, 2 epochs'?

Reviewable status: 0 of 1 LGTMs obtained

[hpssjellis]

@bileschi

Tested on a ubuntu laptop. If you look at the training_test.ts code here

For a 2 epoch run by setting initialEpoch to 1 then the history.epoch variable should be [1] instead of [0,1]

[caisq]

Reviewable status: 0 of 1 LGTMs obtained

---

src/engine/training.ts, line 1311 at r2 (raw file):

?: number

Why is it necessary to change = 0 to ?: number here? I don't think this change is necessary.

---

src/engine/training_test.ts, line 521 at r2 (raw file):

 expect(history.epoch).toEqual([1]);

Can we test this more thoroughly by making an assertion on history.history.loss? Something like:

expect(history.history.loss.length).toEqual(1);

[hpssjellis]

@bileschi

Quote from @shanqing-cai

initialEpoch?: number
Why is it necessary to change = 0 to ?: number here? I don't think this change is necessary.

end of quote.

I can make the requested changes, I was just trying to make initialEpoch follow the same format as: epochs, batchSize and shuffle, since the following code already covers the default settings.

 if (batchSize == null) {
      batchSize = 32;
    }
    if (epochs == null) {
      epochs = 1;
    }
    if (shuffle == null) {
      shuffle = true;
    }
    if (initialEpoch == null) {
      initialEpoch = 0;
    }

P.S. I notice the above code does not detect incorrect negative inputs, would the following code be an improvement?

    if ((batchSize == null) || (batchSize < 0)) {
      batchSize = 32;
    }
    if ((epochs == null) || (epochs < 0)) {
      epochs = 1;
    }
    if (shuffle == null) {
      shuffle = true;
    }
    if ((initialEpoch == null) || (initialEpoch < 0)) {
      initialEpoch = 0;
    }

Either way I am fine with whatever you think is best, just thought I would explain my thinking with the changes.

[caisq]

Ah - that works. So if you can address my other comment about adding some more testing assertions, I will approve this PR.

Thanks.

Reviewable status: 0 of 1 LGTMs obtained

[caisq]

Thanks for this, @hpssjellis !

Reviewable status: complete! 1 of 1 LGTMs obtained

[bileschi]

Reviewable status: complete! 2 of 1 LGTMs obtained