Conversation
|
Seems like there is an error with It is weird as I can see tensorflow 2.18.0 in their pypi page https://pypi.org/project/tensorflow/#history |
| Args: | ||
| predictions: A floating point 2D vector representing the prediction | ||
| generated from the model. The shape should be (batch_size, seq_len). | ||
| labels: True value. The shape should be (batch_size, seq_len). |
There was a problem hiding this comment.
does this mean that this can be multi label? for instance, predictions may be [5 0 2] and the label is [1 0 1]?
There was a problem hiding this comment.
actually, getting the probability of each element inside predictions in line 583 suggests that we have a single label scenario
There was a problem hiding this comment.
Yeah I don't think this works for multi-label currently (it also converts the labels to one-hot on L586). How common are labels with >1 truth values in NLP eval?
There was a problem hiding this comment.
then the shape of labels is (batch_size, ) then?
There was a problem hiding this comment.
Discussed offline - will update predictions shape to add vocab_size
jeffcarp
force-pushed
the
perplexity
branch
7 times, most recently
from
fcc5933 to
f7c7534
Compare
Specifying Python 3.12 (one major version behind the latest) fixed the problem. |
2 participants
No description provided.