generation_steps/metrics_analysis.py

import functools
import operator

import Levenshtein
import evaluate
import pandas as pd
from tqdm import tqdm

import config
from api_wrappers import hf_data_loader
from custom_metrics import gpt_eval

BLEU = evaluate.load('bleu', cache_dir=config.CACHE_DIR)


def bleu_fn(pred, ref, **kwargs):
    return BLEU.compute(predictions=[pred], references=[ref])["bleu"]


METEOR = evaluate.load('meteor', cache_dir=config.CACHE_DIR)


def meteor_fn(pred, ref, **kwargs):
    return METEOR.compute(predictions=[pred], references=[ref])["meteor"]


ROUGE = evaluate.load('rouge', cache_dir=config.CACHE_DIR)


def rouge1_fn(pred, ref, **kwargs):
    return ROUGE.compute(predictions=[pred], references=[ref])["rouge1"]


def rouge2_fn(pred, ref, **kwargs):
    return ROUGE.compute(predictions=[pred], references=[ref])["rouge2"]


def rougeL_fn(pred, ref, **kwargs):
    return ROUGE.compute(predictions=[pred], references=[ref])["rougeL"]


BERTSCORE = evaluate.load('bertscore', cache_dir=config.CACHE_DIR)


def bertscore_fn(pred, ref, **kwargs):
    return BERTSCORE.compute(predictions=[pred], references=[ref], model_type="distilbert-base-uncased")["f1"][0]


CHRF = evaluate.load("chrf")


def chrf_fn(pred, ref, **kwargs):
    return CHRF.compute(predictions=[pred], references=[[ref]])["score"]


TER = evaluate.load("ter")


def ter_fn(pred, ref, **kwargs):
    return TER.compute(predictions=[pred], references=[[ref]])["score"]


def edit_distance_fn(pred, ref, **kwargs):
    return Levenshtein.distance(pred, ref)


def edit_time_fn(pred, ref, **kwargs):
    return kwargs["edittime"]


def gptscore_ref_1_fn(pred, ref, **kwargs):
    return gpt_eval.compute_ref(prediction=pred, reference=ref, n_requests=1)


def gptscore_ref_3_fn(pred, ref, **kwargs):
    return gpt_eval.compute_ref(prediction=pred, reference=ref, n_requests=3)


def gptscore_ref_5_fn(pred, ref, **kwargs):
    return gpt_eval.compute_ref(prediction=pred, reference=ref, n_requests=5)


IND_METRICS = {
    "gptscore-ref-1-req": gptscore_ref_1_fn,
    "gptscore-ref-3-req": gptscore_ref_3_fn,
    # "gptscore-ref-5-req": gptscore_ref_5_fn,
    "editdist": edit_distance_fn,
    "bleu": bleu_fn,
    "meteor": meteor_fn,
    "rouge1": rouge1_fn,
    "rouge2": rouge2_fn,
    "rougeL": rougeL_fn,
    "bertscore": bertscore_fn,
    "chrF": chrf_fn,
    "ter": ter_fn,
}

REL_METRICS = {
    "editdist": edit_distance_fn,
    "edittime": edit_time_fn,
}


def attach_references(df):
    reference_df = hf_data_loader.load_full_commit_as_pandas().set_index(["hash", "repo"])[["reference"]]
    df = df.set_index(["hash", "repo"])
    return df.join(other=reference_df, how="left").reset_index()


def compute_metrics(df):
    tqdm.pandas()

    def apply_metric_fn_to_row(row, fn, col_pred, col_ref):
        return fn(row[col_pred], row[col_ref], edittime=row['edit_time'])

    for metric in REL_METRICS:
        print(f"Computing {metric} for the related pairs")
        metric_fn = REL_METRICS[metric]
        df[f"{metric}_related"] = df.progress_apply(
            lambda row: apply_metric_fn_to_row(row=row,
                                               fn=metric_fn,
                                               col_pred="commit_msg_start",
                                               col_ref="commit_msg_end"),
            axis=1
        )

    for metric in IND_METRICS:
        print(f"Computing {metric} for the independent pairs")
        metric_fn = IND_METRICS[metric]
        df[f"{metric}_independent"] = df.progress_apply(
            lambda row: apply_metric_fn_to_row(row=row,
                                               fn=metric_fn,
                                               col_pred="commit_msg_start",
                                               col_ref="reference"),
            axis=1
        )

    for rel_metric in REL_METRICS:
        for ind_metric in IND_METRICS:
            df[f"rel_{rel_metric}_ind_{ind_metric}_pearson"] = (
                df[f"{rel_metric}_related"].corr(df[f"{ind_metric}_independent"], method="pearson"))

            df[f"rel_{rel_metric}_ind_{ind_metric}_spearman"] = (
                df[f"{rel_metric}_related"].corr(df[f"{ind_metric}_independent"], method="spearman"))

    return df


def correlations_for_group(group):
    correlations = []
    for rel_metric in REL_METRICS:
        # correlations.append({
        #     f"{metric}_pearson": group[f"{metric}_related"].corr(group[f"{metric}_independent"], method="pearson"),
        #     f"{metric}_spearman": group[f"{metric}_related"].corr(group[f"{metric}_independent"], method="spearman")
        # })
        for ind_metric in IND_METRICS:
            correlations.append({
                f"rel_{rel_metric}_ind_{ind_metric}_pearson": group[f"{rel_metric}_related"].corr(
                    group[f"{ind_metric}_independent"], method="pearson"),
                f"rel_{rel_metric}_ind_{ind_metric}_spearman": group[f"{rel_metric}_related"].corr(
                    group[f"{ind_metric}_independent"], method="spearman"),
            })
    return pd.Series(functools.reduce(operator.ior, correlations, {}))


def compute_correlations(df: pd.DataFrame):
    grouped_df = df.groupby(by=["end_to_start", "start_to_end"])
    correlations = grouped_df.apply(correlations_for_group, include_groups=False)
    return correlations


def transform(df):
    print("Computing metrics")

    df = attach_references(df)
    df = compute_metrics(df)

    correlations_for_groups = compute_correlations(df)
    correlations_for_groups.to_csv(config.METRICS_CORRELATIONS_ARTIFACT)

    df.to_csv(config.SYNTHETIC_DATASET_ARTIFACT)

    print("Done")
    return df


def main():
    df = pd.read_csv(config.START_TO_END_ARTIFACT, index_col=[0])
    transform(df)


if __name__ == '__main__':
    main()