Add new differentiation method based on rewinding the tape [original PR, not to be merged] #1017
Conversation
|
Hello. You may have forgotten to update the changelog!
|
Codecov Report
@@ Coverage Diff @@
## master #1017 +/- ##
==========================================
- Coverage 97.94% 97.93% -0.02%
==========================================
Files 153 154 +1
Lines 11371 11452 +81
==========================================
+ Hits 11137 11215 +78
- Misses 234 237 +3
Continue to review full report at Codecov.
|
|
[ch4083] |
|
|
||
| # Perform the forward pass | ||
| # TODO: Could we use lower-level like device.apply, since we just need the state? | ||
| self.execute(device, params=params) |
There was a problem hiding this comment.
Would be wonderful if we could use a memoized/ cached circuit execution here. Though I guess execution memorization is on the to-do list for all the tapes.
There was a problem hiding this comment.
There is device-side caching - but there's also lots going in between tape.execute() and device.execute() which may be unnecessary. Also, we only need to update the state, so even dev.apply() might be enough.
|
|
||
| # Perform the forward pass | ||
| # TODO: Could we use lower-level like device.apply, since we just need the state? | ||
| self.execute(device, params=params) |
There was a problem hiding this comment.
There is device-side caching - but there's also lots going in between tape.execute() and device.execute() which may be unnecessary. Also, we only need to update the state, so even dev.apply() might be enough.
Co-authored-by: Christina Lee <chrissie.c.l@gmail.com>
trbromley
changed the title
* add device fixture * second class includes dev fixture
|
Closing this since it is superseded by #1032 |
WIP