PBP dataset performance enhancements #361
Merged
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
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
M-R-Schaefer
requested review from
PythonFZ
and removed request for
PythonFZ
PythonFZ
reviewed
Nov 19, 2024
apax/data/input_pipeline.py
Outdated
Comment on lines
419
to
421
| labels["energy"][i] = lab["energy"] | ||
| if self.forces: | ||
| labels["forces"][i, : inp["n_atoms"]] = lab["forces"] |
There was a problem hiding this comment.
How are other labels treated, e.g. stress?
There was a problem hiding this comment.
oh, I removed the inclusion of stress... Good catch
M-R-Schaefer
commented
Nov 19, 2024
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
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Outdated suggestions cannot be applied.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.
I have reworked the PBP dataset to enhance performance.
Previously, the buffer would be filled with batches using multiprocessing, run dry and then be refilled.
Now, there is a dedicated processing thread which uses MP internally.
That means the buffer is now continuously filled without blocking train step compute.
On my home PC I have achieved a 20 % epoch time reduction compared to the cached dataset.
On my work laptop, it's roughly 30 % slower, but still much faster than the old pbp (old: 6.3 s, new: 1.3 s)
Apart from this dataset, there is also the CSVLogger class which uses tensorflow.
I think if we write our own logger, we could turn tensorflow into an optional dependency.
TODO:
Update:
After rerunning the benchmark on my workstation, the cached dataset is still faster than the PBB, especially in worst case scenarios (same system size, low batch size, lightweight model).
300 atom system, batch size 4, rmax=6
cached: 1.5 s/epoch
pbp new : 2.5 s/epoch
pbp old: 4.7 s/epoch
With BS 8, n_ens = 8, rmax=5
cached: 5.6 s/epoch
pbp: 6.7 s/epoch