Return type of tf.oneHot depends on backend

[bileschi]

To get help from the community, check out our Google group.

TensorFlow.js version

0.11.8

Browser version

irrelevant

Describe the problem or feature request

tf.oneHot returns a tensor of dtype 'float32' with some backends, and with dtype 'int32' in other backends.

Code to reproduce the bug / link to feature request

describeWithFlags('oneHotType', ALL_ENVS, () => {
  expectedType = 'int32'
  fit('Output type is what', () => {
    const indices = tf.tensor1d([0, 1], 'int32');
    const res = tf.oneHot(indices, 2);
    console.log("XXXXXXXXXXXXXXXX");
    console.log(res.dtype);
    expect(res.dtype).toEqual(expectedType);
  });
});

outputs

LOG: 'XXXXXXXXXXXXXXXX'
LOG: 'int32'
LOG: 'XXXXXXXXXXXXXXXX'
LOG: 'int32'
LOG: 'XXXXXXXXXXXXXXXX'
LOG: 'float32'
Chrome 67.0.3396 (Linux 0.0.0) oneHot {"BACKEND":"test-cpu"} Output type is int32 FAILED
	Expected 'float32' to equal 'int32'.
	    at <Jasmine>
	    at UserContext.<anonymous> (src/ops/array_ops_test.ts:2171:22 <- src/ops/array_ops_test.js:7084:39)
	    at <Jasmine>

[manrajgrover]

@bileschi Looks like CPU backend was using default float32 as dtype. Providing int32 as dtype to output fixes it.