In case you don’t settle for the default scikit-learn metrics, you can define your own metric.
Here, we expect the organizer to know python.
To define a custom metric, change EVAL_METRIC in conf.json to custom. You must also make sure that EVAL_HIGHER_IS_BETTER is set to 1 or 0 depending on whether a higher value of the metric is better or not.
The second step is to create a file metric.py in the private competition repo.
The file should contain a compute function that takes competition params as input.
Here is the part where we check if metric is custom and calculate the metric value:
def compute_metrics(params):
if params.metric == "custom":
metric_file = hf_hub_download(
repo_id=params.competition_id,
filename="metric.py",
token=params.token,
repo_type="dataset",
)
sys.path.append(os.path.dirname(metric_file))
metric = importlib.import_module("metric")
evaluation = metric.compute(params)
.
.
.You can find the above part in competitions github repo compute_metrics.py
params is defined as:
class EvalParams(BaseModel):
competition_id: str
competition_type: str
metric: str
token: str
team_id: str
submission_id: str
submission_id_col: str
submission_cols: List[str]
submission_rows: int
output_path: str
submission_repo: str
time_limit: int
dataset: str # private test dataset, used only for script competitionsYou are free to do whatever you want to in the compute function.
In the end it must return a dictionary with the following keys:
{
"public_score": {
"metric1": metric_value,
},,
"private_score": {
"metric1": metric_value,
},,
}public and private scores must be dictionaries! You can also use multiple metrics. Example for multiple metrics:
{
"public_score": {
"metric1": metric_value,
"metric2": metric_value,
},
"private_score": {
"metric1": metric_value,
"metric2": metric_value,
},
}Note: When using multiple metrics, conf.json must have SCORING_METRIC specified to rank the participants in the competition.
For example, if I want to use metric2 to rank the participants, I will set SCORING_METRIC to metric2 in conf.json.
import pandas as pd
from huggingface_hub import hf_hub_download
def compute(params):
solution_file = hf_hub_download(
repo_id=params.competition_id,
filename="solution.csv",
token=params.token,
repo_type="dataset",
)
solution_df = pd.read_csv(solution_file)
submission_filename = f"submissions/{params.team_id}-{params.submission_id}.csv"
submission_file = hf_hub_download(
repo_id=params.competition_id,
filename=submission_filename,
token=params.token,
repo_type="dataset",
)
submission_df = pd.read_csv(submission_file)
public_ids = solution_df[solution_df.split == "public"][params.submission_id_col].values
private_ids = solution_df[solution_df.split == "private"][params.submission_id_col].values
public_solution_df = solution_df[solution_df[params.submission_id_col].isin(public_ids)]
public_submission_df = submission_df[submission_df[params.submission_id_col].isin(public_ids)]
private_solution_df = solution_df[solution_df[params.submission_id_col].isin(private_ids)]
private_submission_df = submission_df[submission_df[params.submission_id_col].isin(private_ids)]
public_solution_df = public_solution_df.sort_values(params.submission_id_col).reset_index(drop=True)
public_submission_df = public_submission_df.sort_values(params.submission_id_col).reset_index(drop=True)
private_solution_df = private_solution_df.sort_values(params.submission_id_col).reset_index(drop=True)
private_submission_df = private_submission_df.sort_values(params.submission_id_col).reset_index(drop=True)
# CALCULATE METRICS HERE.......
# _metric = SOME METRIC FUNCTION
target_cols = [col for col in solution_df.columns if col not in [params.submission_id_col, "split"]]
public_score = _metric(public_solution_df[target_cols], public_submission_df[target_cols])
private_score = _metric(private_solution_df[target_cols], private_submission_df[target_cols])
evaluation = {
"public_score": {
"metric1": public_score,
},
"private_score": {
"metric1": public_score,
}
}
return evaluationTake a careful look at the above code. You can see that we are downloading the solution file and the submission file from the dataset repo. We are then calculating the metric on the public and private splits of the solution and submission files. Finally, we are returning the metric values in a dictionary.
< > Update on GitHub