Fix training resumption #312
Draft
+962
−615
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
sichu2023
force-pushed
the
sichu/fix-pplcallback-sanity
branch
from
c8a7d1a
to
4575a61
Compare
sichu2023
requested review from
jstjohn,
malcolmgreaves,
skothenhill-nv,
farhadrgh,
dorotat-nv,
gwarmstrong,
jomitchellnv and
pstjohn
as code owners
sub-packages/bionemo-geneformer/tests/bionemo/geneformer/test_stop_and_go.py
Outdated
Show resolved
Hide resolved
sub-packages/bionemo-esm2/tests/bionemo/esm2/model/test_stop_and_go.py
Outdated
Show resolved
Hide resolved
sub-packages/bionemo-testing/src/bionemo/testing/harnesses/stop_and_go.py
Outdated
Show resolved
Hide resolved
jstjohn
reviewed
Oct 23, 2024
sub-packages/bionemo-testing/src/bionemo/testing/testing_callbacks.py
Outdated
Show resolved
Hide resolved
sub-packages/bionemo-testing/src/bionemo/testing/testing_callbacks.py
Outdated
Show resolved
Hide resolved
skothenhill-nv
approved these changes
Oct 23, 2024
jstjohn
approved these changes
Oct 24, 2024
sichu2023
force-pushed
the
sichu/fix-pplcallback-sanity
branch
from
68fa7cb
to
332a7b9
Compare
|
/build-ci |
|
Confirm that we can reproduce loss curve upon resumption despite a small jump on validation loss at the beginning. |
pstjohn
reviewed
Oct 25, 2024
Comment on lines
+35
to
+85
| def create_dummy_parquet_train_val_inputs(tmp_path: Path) -> Tuple[Path, Path]: | ||
| """Create a mock protein train and val cluster parquet.""" | ||
| train_cluster_path = tmp_path / "train_clusters.parquet" | ||
| train_clusters = pd.DataFrame( | ||
| { | ||
| "ur90_id": [["UniRef90_A"], ["UniRef90_B", "UniRef90_C"]], | ||
| } | ||
| ) | ||
| train_clusters.to_parquet(train_cluster_path) | ||
|
|
||
| valid_cluster_path = tmp_path / "valid_clusters.parquet" | ||
| valid_clusters = pd.DataFrame( | ||
| { | ||
| "ur50_id": ["UniRef50_A", "UniRef50_B", "UniRef90_A", "UniRef90_B"], | ||
| } | ||
| ) | ||
| valid_clusters.to_parquet(valid_cluster_path) | ||
| return train_cluster_path, valid_cluster_path | ||
|
|
||
|
|
||
| def create_dummy_protein_dataset(tmp_path) -> Path: | ||
| """Create a mock protein dataset.""" | ||
| if not isinstance(tmp_path, Path): | ||
| tmp_path = Path(str(tmp_path)) | ||
|
|
||
| db_file = tmp_path / "protein_dataset.db" | ||
| conn = sqlite3.connect(str(db_file)) | ||
| cursor = conn.cursor() | ||
|
|
||
| cursor.execute( | ||
| """ | ||
| CREATE TABLE protein ( | ||
| id TEXT PRIMARY KEY, | ||
| sequence TEXT | ||
| ) | ||
| """ | ||
| ) | ||
|
|
||
| proteins = [ | ||
| ("UniRef90_A", "ACDEFGHIKLMNPQRSTVWY"), | ||
| ("UniRef90_B", "DEFGHIKLMNPQRSTVWYAC"), | ||
| ("UniRef90_C", "MGHIKLMNPQRSTVWYACDE"), | ||
| ("UniRef50_A", "MKTVRQERLKSIVRI"), | ||
| ("UniRef50_B", "MRILERSKEPVSGAQLA"), | ||
| ] | ||
| cursor.executemany("INSERT INTO protein VALUES (?, ?)", proteins) | ||
|
|
||
| conn.commit() | ||
| conn.close() | ||
|
|
||
| return db_file |
There was a problem hiding this comment.
These are testing-specific functions, let's not have them in the library itself
|
Broke down the PR into smaller pieces. Now we have a PR specifically tackles inconsistency stop-and-go v.s. continuous training curve. |
pstjohn
force-pushed
the
sichu/fix-pplcallback-sanity
branch
from
631a51e
to
4994705
Compare
jstjohn
changed the base branch from
main
to
sichu/match-resumption-loss-curve
|
/build-ci |
|
/build-ci |
|
/build-ci |
|
/build-ci |
|
/build-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.
Summary
Loss curve from training resumption is inconsistent with a single uninterrupted loss curve.
Details
It has now been identified partly from incorrect datamodule behavior and is fixed with implementing
state_dictfor datamodule. However, there is still a very minor dip in validation loss curve that does not affect subsequent training curve.