Draft document classifier #8
Conversation
vitawasalreadytaken
changed the base branch from
main
to
feature/data-pipeline
vitawasalreadytaken
force-pushed
the
feature/draft-classifier
branch
from
fef60cd to
3127e5a
Compare
|
This is now rebased on main where #6 has been merged. |
vitawasalreadytaken
force-pushed
the
feature/draft-classifier
branch
from
6ecde1a to
a34fe53
Compare
|
All dependencies such as the research libraries and code are now merged in |
| print(f"Number of dropped empty texts: {empty_count} ({100 * empty_count / len(df_input):.1f}%)") | ||
| return df.loc[~empty_index] | ||
|
|
||
| def create_embeddings(column_to_embed: str, cache_directory: pathlib.Path = REPOSITORY_ROOT / "data" / "embeddings-cache"): |
There was a problem hiding this comment.
This function relies on global variables (df_input and embedding_model). Instead of using global state, let's pass the necessary data in as arguments. Something like this:
| def create_embeddings(column_to_embed: str, cache_directory: pathlib.Path = REPOSITORY_ROOT / "data" / "embeddings-cache"): | |
| def create_embeddings(strings_to_embed: pd.Series, embedding_model: embeddings.EmbeddingModel, cache_directory: pathlib.Path = REPOSITORY_ROOT / "data" / "embeddings-cache") -> np.ndarray: |
| print(embeddings_doc_content_plain.shape) | ||
| return embeddings_doc_content_plain | ||
|
|
||
| def save_model(model_file_name): |
There was a problem hiding this comment.
This is similar - let's pass classifier in as an argument.
| from research.lib import data_access, embeddings | ||
|
|
||
| REPOSITORY_ROOT = (pathlib.Path().cwd() / ".." / "..").resolve() | ||
| sys.path.append(str(REPOSITORY_ROOT)) |
There was a problem hiding this comment.
Is this necessary? Since the imports above (from research.lib ...) apparently work?
There was a problem hiding this comment.
Note: I had to run the script like this:
cd research/document_types
PYTHONPATH=../../ uv run draft_classification.py
so that a) the research.lib imports on line 22 work and b) the relative path to draft_classification.toml works.
| REPOSITORY_ROOT = (pathlib.Path().cwd() / ".." / "..").resolve() | ||
| sys.path.append(str(REPOSITORY_ROOT)) | ||
|
|
||
| dotenv.load_dotenv() |
There was a problem hiding this comment.
Side effects such as this should be in the if __name__ == "__main__": block, otherwise they're triggered even when this module is imported by something else.
| with open('draft_classification.toml', 'r') as f: | ||
| config = toml.load(f) | ||
|
|
||
| os.environ["MLFLOW_TRACKING_URI"] = config["tracking"]["tracking_uri"] |
There was a problem hiding this comment.
This is also a side effect that should not be triggered at import time. In addition, we can use https://mlflow.org/docs/1.22.0/python_api/mlflow.html#mlflow.set_tracking_uri instead of manipulating environment variables. (I know it was like that in the original notebook but we can improve things in this new script 🙂)
| with open('draft_classification.toml', 'r') as f: | ||
| config = toml.load(f) |
There was a problem hiding this comment.
| with open('draft_classification.toml', 'r') as f: | |
| config = toml.load(f) | |
| config = toml.load("draft_classification.toml") |
However, it's also a side effect that would preferably happen in if __name__ == "__main__"...
| ### Preprocessing ### | ||
| df_input = remove_rows_with_missing_text(df_input) | ||
| # set target variable | ||
| df_input.loc[:, "is_draft"] = (df_input.loc[:, "document_type"] == "DRAFT").astype(bool) |
There was a problem hiding this comment.
Oddly, I'm getting this warning on this line. I don't know why.
SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
… topic classification
vitawasalreadytaken
changed the title
I've placed two somewhat unrelated things into this branch because they both depend on #6 and on some research-y code from our old, exploratory repository.
research/consultation_topics/label_exploration.ipynbtakes a look at the consultation topic labels we have.research/document_types/VM_draft_classifier.ipynbis a bare-bone document type classifier, on Fedlex data only, making "is DRAFT/is not a DRAFT" predictions only.We can split these up again if this branch stops making sense.
Handing this over to you now, @orieger 🙂