Scientific notation on y axis appears when unhelpful

[iansf]

TensorBoard plots switch to scientific notation for 5 digit numbers, which requires more characters to represent, and actually hides relevant precision, as demonstrated in the attached screenshot.

-1.329e+4 -- 9 characters for 4 digits of precision and sign.
-13290 -- 6 characters for 5 digits of precision and sign.

It seems clear that scientific notation should only be enabled when the range of the plots contains numbers with more digits than would be used with the (less precise) scientific notation representation. In other words, scientific notation should only be used when one or both ends of the range has a magnitude >= 10^(5+desired digits of precision), where the 5 comes from the first digit, the decimal point, and the minimum number of exponent characters, which is 3.

The same basic reasoning also applies to numbers close to 0.

• TensorBoard version (from pip package, also printed out when running tensorboard)
  TensorBoard 67
• Python version (e.g. 2.7, 3.5)
  Python 2.7

[nfelt]

Agreed that the scientific notation activates too readily, it should be more conservative.

There's another issue (which I thought had been reported somewhere but couldn't find after a quick search), in that if the absolute scale warrants scientific notation but the actual range of values is very small, then the result is that the y axis labels all are the same (e.g. "1.131e+4" when the actual tick mark values are 11310, 11311, 11312, ...). In that case as well we should perhaps avoid the scientific notation and just make the axis labels take up more room.