Hparams: Treat Nan/Infinity numeric hparam values as unset values. #6496
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
arcra
approved these changes
Jul 17, 2023
| @@ -304,6 +300,11 @@ def _compute_hparam_info_from_values(self, name, values): | |||
| return None | |||
|
|
|||
| if result.type == api_pb2.DATA_TYPE_STRING: | |||
| distinct_values = set() | |||
| for value in values: | |||
| if _protobuf_value_type(value): | |||
There was a problem hiding this comment.
nit: Consider extracting into a separate named function, where you can use early returns, to reduce nesting, and so you can use a set comprehension here, like it was before.
Optional. This is fine too.
| False, otherwise. | ||
| """ | ||
| return not value.HasField("number_value") or ( | ||
| not math.isnan(value.number_value) |
There was a problem hiding this comment.
nit: Consider using early returns, so the "short circuit" evaluation is more evident:
if not value.HasField("..."):
return True
val = value.number_value
return not math.isnan(val) and not math.isinf(val)
Again, optional.
| request = """ | ||
| start_index: 0 | ||
| slice_size: 10 | ||
| allowed_statuses: [STATUS_UNKNOWN] |
There was a problem hiding this comment.
Is this just an arbitrary value that is irrelevant for the test? Can it be omitted, or is it a required field?
There was a problem hiding this comment.
Ya unfortunately it is necessary for the query to work. (allowed_status: [] will return no results).
bmd3k
added a commit
that referenced
this pull request
Jul 18, 2023
#6496 broke the internal import due to missing BUILD dependency for json_format_compat_test.py.
rileyajones
pushed a commit
to rileyajones/tensorboard
that referenced
this pull request
Jul 18, 2023
…ensorflow#6496) ## Motivation for features / changes There are cases where users might log hparam numeric values that are NaN or Infinity. Recall, we use proto's Value message for logging and storing hparam values. While processing these values in the python layer of the Hparams plugin, we may run these Value message through json_format.MessageToJson, at which point the json_format library raises an error. The limitation is documented at https://protobuf.dev/reference/protobuf/google.protobuf/#value. "[A]ttempting to serialize NaN or Infinity results in error." ## Technical description of changes We now will try to avoid calling MessageToJson on these problematic hparam values. We instead chose to treat these as "unset" values -- equivalent to having not set an hparam value at all. It will mean the values will not contribute to the calculation of discrete domains. And it means that the values will appear blank in the hparams UI rather than appearing as "NaN". ## Alternate designs / implementations considered (or N/A) Given infinite amount of time: * Fix json_format to handle NaN and Infinity :O. * Change the hparams api.proto to use something other than proto's Value for representing hparam values. Other reasonable options: * Change the NaN numeric value to "NaN" string. This was my first choice but it unfortunately leads to other complications - for instance the plugin will raise errors when it encounters a "NaN" string for an hparam that is supposed to have a numeric value.
rileyajones
pushed a commit
to rileyajones/tensorboard
that referenced
this pull request
Jul 18, 2023
tensorflow#6496 broke the internal import due to missing BUILD dependency for json_format_compat_test.py.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation for features / changes
There are cases where users might log hparam numeric values that are NaN or Infinity. Recall, we use proto's Value message for logging and storing hparam values. While processing these values in the python layer of the Hparams plugin, we may run these Value message through json_format.MessageToJson, at which point the json_format library raises an error.
The limitation is documented at https://protobuf.dev/reference/protobuf/google.protobuf/#value. "[A]ttempting to serialize NaN or Infinity results in error."
Technical description of changes
We now will try to avoid calling MessageToJson on these problematic hparam values. We instead chose to treat these as "unset" values -- equivalent to having not set an hparam value at all. It will mean the values will not contribute to the calculation of discrete domains. And it means that the values will appear blank in the hparams UI rather than appearing as "NaN".
Alternate designs / implementations considered (or N/A)
Given infinite amount of time:
Other reasonable options: