Correct numpy tolerence #167
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #167 +/- ##
==========================================
+ Coverage 85.87% 87.23% +1.35%
==========================================
Files 117 118 +1
Lines 12342 13710 +1368
==========================================
+ Hits 10599 11960 +1361
- Misses 1743 1750 +7 ☔ View full report in Codecov by Sentry. |
|
In general the PR is OK, but I'm curious if this setting has a practical effect on the calculations you've carried out? |
Nope. Code from the lastest master branch still messes up our Holstein test after I set Anyway, proposing a fix on atol is the least thing I can do at this stage : ( |
|
What do you mean by messing up?
…On Thu, May 9, 2024 at 6:03 AM Yu Xiong ***@***.***> wrote:
if this setting has a practical effect on the calculations you've carried
out
Nope. Code from the lastest master branch still messes up our Holstein
test after I set $RENO_NUM_THREADS to 1 and appied the new atol. I am
still investigating on why it happens on my setup.
Anyway, proposing a fix on atol is the least thing I can do at this stage
: (
—
Reply to this email directly, view it on GitHub
<#167 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AIZLWMJW2PXQ7FUFOAVL7ZLZBNCXBAVCNFSM6AAAAABHONXYD2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMBSGM2TENBYGQ>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
|
@jiangtong1000 I intend to open a new issue for discussion of the current problem. Please check #168 : D |
|
@atxy-blip ready to merge if you fix my review comment. It'll be better if you can solve the codecov issue. |
atxy-blip
force-pushed
the
fix-atol
branch
2 times, most recently
from
6eb51b2
to
2c37a67
Compare
Fixed. |
|
Thanks for the contribution! Looking forward to more of your PRs 😄 |
- originally introduced in commit 254b981
|
Note that this branch has been rebased and its commits are different from your local version @atxy-blip |
In commit 254b981,
atolfor numpy backend was wrongly set to1e-5inmps/backend.py, instead of the original value1e-8. Meanwhile, in this version onlyatolwas set up.Renormalizer/renormalizer/mps/backend.py
Lines 136 to 140 in 254b981
As for
numpy, the basic logic for considering two numbersaandbclose in is:The existence of
rtolallows for the comparison of large numbers, e.g., 10001 and 10000, which may emerge in imaginary time-evolution. Therefore, I propose adding back the rtol parameter.This commit does the following things:
mps.backend.canonical_rtolandmps.backend.canonical_atolmps.matrix.check_lortho,mps.mp.check_left_canonical,mps.mp.ensure_left_canonicaland their right counterpartsI am not sure whether
rtolset to 1e-2 is appropriate for 32-bits. Discussions are welcome for any issue concerned : )