Add stubs and fakes for Dataset

[caisq]

DEV

---

This change is 

[davidsoergel]

Reviewable status: 0 of 1 approvals obtained (waiting on @caisq and @davidsoergel)

---

src/engine/dataset_stub.ts, line 47 at r1 (raw file):

  /**
   * The size of each batch generated by the iteartor.

nit: iterator

---

src/engine/dataset_stub.ts, line 115 at r1 (raw file):

/**
 * A fake dataset with configurable feature and target shapes.

I don't know what tests you have in mind, but this all seems unnecessarily complex. I'd just hardcode 10 trivial examples, maybe in batches of 3 (to exercise 'small last batch' logic).

Now that you already wrote this, we could keep it-- but even so I'm not sure it's worth the comprehension & maintenance cost. Do you think a much simpler fake would be sufficient?

---

src/engine/dataset_stub.ts, line 127 at r1 (raw file):

    tfc.util.assert(
        config.batchSize > 0 && Number.isInteger(config.batchSize),
        `batchSize must be a positive ineger, but got ${config.batchSize}`);

nit: integer

---

src/engine/dataset_stub_test.ts, line 17 at r1 (raw file):

import {FakeNumericDataset} from './dataset_stub';

describeMathCPUAndGPU('FakeNumericDataset', () => {

Wow, tests for the fake? Very thorough :)

---

src/engine/dataset_stub_test.ts, line 22 at r1 (raw file):

        {xShape: [3], yShape: [1], batchSize: 8, numBatches: 5});
    for (let k = 0; k < 2; ++k) {
      // Run twice to make sure that calling iteartor() multiple times works.

nit: iterator

---

src/engine/training.ts, line 35 at r1 (raw file):

import {SymbolicTensor} from './topology';

I thought lint wants two blank lines here-- maybe this is a mistake?

[caisq]

Reviewable status: 0 of 1 approvals obtained (waiting on @caisq)

---

src/engine/dataset_stub.ts, line 47 at r1 (raw file):

Previously, davidsoergel (David Soergel) wrote…

nit: iterator

Done.

---

src/engine/dataset_stub.ts, line 115 at r1 (raw file):

Previously, davidsoergel (David Soergel) wrote…

I don't know what tests you have in mind, but this all seems unnecessarily complex. I'd just hardcode 10 trivial examples, maybe in batches of 3 (to exercise 'small last batch' logic).

Now that you already wrote this, we could keep it-- but even so I'm not sure it's worth the comprehension & maintenance cost. Do you think a much simpler fake would be sufficient?

If we hardcode 10 examples, there will be quite some code duplication, potentially leading to higher maintenance cost. The API of the fake class here is flexible enough to support

1. Different numbers of input tensors and output tensors
2. Different shapes of the tensors
   which should satisfy most test cases. If we find we need other fake datasets, we can extend this existing class later without unnecessary code duplication.

---

src/engine/dataset_stub.ts, line 127 at r1 (raw file):

Previously, davidsoergel (David Soergel) wrote…

nit: integer

Done.

---

src/engine/dataset_stub_test.ts, line 17 at r1 (raw file):

Previously, davidsoergel (David Soergel) wrote…

Wow, tests for the fake? Very thorough :)

Ack. :)

---

src/engine/training.ts, line 35 at r1 (raw file):

Previously, davidsoergel (David Soergel) wrote…

I thought lint wants two blank lines here-- maybe this is a mistake?

Hmm. It looks it doesn't matter whether it's 1 or 2 lines. Reverted this change.

[davidsoergel]

Reviewed 3 of 3 files at r2.
Reviewable status: 0 of 1 approvals obtained (waiting on @davidsoergel and @caisq)