Fix failing test in travis

[dsmilkov]

#1379 added a unit test that failed in travis due to numerical precision: https://travis-ci.org/tensorflow/tfjs-core/jobs/451654174#L801

This PR:

• fixes that test
• Also fixes test-travis.sh to make sure we don't forward webgl --> cpu except for the last browser run -- probably caused during conflict resolving of WebGL backend hands off kernel execution to CPU if inputs are small and on the CPU #1371 and Add integration test with layers/node/converter when we modify version.ts #1372 (https://reviewable.io/reviews/tensorflow/tfjs-core/1371#-LQ_a2y92YYLRMiYWqhC)

---

For repository owners only:

Please remember to apply all applicable tags to your pull request.
Tags: FEATURE, BREAKING, BUG, PERF, DEV, DOC, SECURITY

For more info see: https://github.com/tensorflow/tfjs/blob/master/DEVELOPMENT.md

---

This change is 

[nsthorat]

Reviewed 2 of 2 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @nsthorat and @annxingyuan)

[nsthorat]

Looks like those bash comments are screwing things up, my guess is because of the \ happening before the comment.

[dsmilkov]

Yeah, removed them.

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan)

[DirkToewe]

Out of curiosity: Why is the Safari JSVM behaving differently? The results seem to be way of. Even if summed up naively:

Array.from({ length: 10*1000 }, () => Math.fround(0.1**2) )
     .reduce( (x,y) => Math.fround(x+y), /*initVal=*/0.0 )

The result on my machine is 100.0029525756836. But if I am not mistaken the CPU matrix multiplication in TFJS is blocked up. Inside the blocks, a float64 value is used to sum the values up. So the result should be something along the lines of:

const blockSize = 50;
Array.from(
  { length: 10*1000/blockSize },
  () => Math.fround(
    Array.from({ length: blockSize }, () => Math.fround(0.1**2) )
         .reduce( (x,y) => x+y, /*initVal=*/0.0 )
  )
).reduce( (x,y) => Math.fround(x+y), /*initVal=*/0.0 )

That result even perfectly rounds to 100 on my machine.

In the Travis log the result seems to be 99.29019927978516 which would be way off.

[dsmilkov]

I noticed the same - Didn't give much thought since that test was not failing on my local laptop (Macbook Pro). The only diff between my laptop and Travis was the Safari version: mine is 12.0.1, Travis used 11.1.2

[DirkToewe]

I too had some issued with the Safari tests on Travis. My guess was that it was a WebGL thing, but maybe there is an issue with Float32Array in that Webkit/Safari version?