I’ve noticed that SameDiff.output
/SameDiff.fit
(training phase) and SameDiff.evaluate
modify internal SameDiff state quite a bit. JavaDoc does not specify is it safe to call SameDiff.evaluate
when performing SameDiff.fit
(I need something like data preprocessor that runs directly within SameDiff evaluation rather than as preprocessor inside BaseDatasetIterator).
As I’ve found several implementation options, here are the questions I have regarding each of option:
- Is it safe to execute
SameDiff.evaluate
inside customDifferentialFunction
while executingSameDiff.fit
? Or it is safer to useSameDiff.invokeGraphOn
on some copy instead - Is it safe to execute
SameDiff.evaluate
inside custom training listener that will modify network inputs while executingSameDiff.fit
? Or it is safer to useSameDiff.invokeGraphOn
on some copy instead