logmetric.py

# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""TODO: Add a description here."""

import evaluate
import datasets
import re
import dateutil.parser
import numpy as np
from difflib import SequenceMatcher
import sacrebleu

import time


# TODO: Add BibTeX citation
_CITATION = """\
@InProceedings{huggingface:module,
title = {A great new module},
authors={huggingface, Inc.},
year={2020}
}
"""

# TODO: Add description of the module here
_DESCRIPTION = """\
This new module is designed to solve this great ML task and is crafted with a lot of care.
"""


# TODO: Add description of the arguments of the module here
_KWARGS_DESCRIPTION = """
Calculates how good are predictions given some references, using certain scores
Args:
    predictions: list of predictions to score. Each predictions
        should be a string with tokens separated by spaces.
    references: list of reference for each prediction. Each
        reference should be a string with tokens separated by spaces.
Returns:
    accuracy: description of the first score,
    another_score: description of the second score,
Examples:
    Examples should be written in doctest format, and should illustrate how
    to use the function.

    >>> my_new_module = evaluate.load("my_new_module")
    >>> results = my_new_module.compute(references=[0, 1], predictions=[0, 1])
    >>> print(results)
    {'accuracy': 1.0}
"""

# TODO: Define external resources urls if needed
BAD_WORDS_URL = "http://url/to/external/resource/bad_words.txt"


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class LogMetric(evaluate.Metric):
    """TODO: Short description of my evaluation module."""

    # Constant regex to get timestrings
    timestamp_regex = r'^\s*\[?\s*(\d{4}[-/.]\d{2}[-/.]\d{2}(?:[ T]\d{2}[:]\d{2}(?:[:]\d{2}(?:[.,]\d+)?)?(?:Z|[+-]\d{2}[:]\d{2})?)?)\s*\]?\s*'
    timestamp_pattern = re.compile(timestamp_regex, re.MULTILINE)
    sacrebleu_metric = evaluate.load("evaluate-metric/sacrebleu")


    def _info(self):
        # TODO: Specifies the evaluate.EvaluationModuleInfo object
        return evaluate.MetricInfo(
            # This is the description that will appear on the modules page.
            module_type="metric",
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            # This defines the format of each prediction and reference
            # Both prediction and reference are strings
            features=datasets.Features({
                "predictions": datasets.Value("string", id="sequence"),
                "references": datasets.Value("string", id="sequence"),
            }),
            # Homepage of the module for documentation
            homepage="http://module.homepage",
            # Additional links to the codebase or references
            codebase_urls=["http://github.com/path/to/codebase/of/new_module"],
            reference_urls=["http://path.to.reference.url/new_module"]
        )

    def _download_and_prepare(self, dl_manager):
        """Optional: download external resources useful to compute the scores"""
        # TODO: Download external resources if needed
        pass

    # Jaccard Similarity to measure closeness of two log-messages
    def get_jaccard_similarity(self, set1, set2):
        intersection = set1.intersection(set2)
        union = set1.union(set2)
        return len(intersection) / len(union)
    
    # A score depending on the difference in length of two sentences
    def get_length_score(self, preds_split, refs_split):

        pred_content_lengths = np.vectorize(len)(preds_split)
        ref_content_lengths = np.vectorize(len)(refs_split)

        return self.smapeScore(pred_content_lengths, ref_content_lengths)

    # helper function that computes the smape_score either between two numbers or two lists of numbers (must be the same length)
    def smapeScore(self, P, R):
        P_isnumber = isinstance(P, (int, float))
        R_isnumber = isinstance(R, (int, float))

        # either both must be numbers or both must be no number
        assert P_isnumber == R_isnumber

        if not P_isnumber:
            assert(len(P) == len(R))

        if P_isnumber and R_isnumber:
            if P == 0 and R == 0: return 1.0      # since this leads to (|R| + |P|) = 0
            n = 1
        else:
            if P == [] and R == []: return 1.0     # since this leads to n = 0
            n = len(P)

        return 1/n * np.sum(np.abs(R - P) / (np.abs(R) + np.abs(P)))

    # splits both strings at \n and then computes the smape_score of their lengths
    def getLineCountScore(self, pred, ref):
        pred_lines_amt = len(pred.splitlines())
        ref_lines_amt = len(ref.splitlines())

        # print("#pred_lines:", pred_lines_amt)
        # print("#ref_lines:", ref_lines_amt)

        return self.smapeScore(pred_lines_amt, ref_lines_amt)
    
    
    # Get differenct scores regarding the content of a log-message
    def getLineContentScore(self, pred_logMessages, ref_logMessages):
        if pred_logMessages == [] and ref_logMessages == []:
            pred_logMessages, ref_logMessages = [""]
        sacrebleu_score = self.sacrebleu_metric.compute(predictions=pred_logMessages, references=ref_logMessages)["score"] / 100.0
        
        smape_length_score = self.get_length_score(pred_logMessages, ref_logMessages)

        # Split the single log-messages (for jaccard)
        preds_split = [p.split() for p in pred_logMessages]
        refs_split = [r.split() for r in ref_logMessages]

        # compute jaccard pair-wise, then take mean of all results
        mean_jaccard_score = np.mean([self.get_jaccard_similarity(set(p_split), set(r_split)) for p_split, r_split in zip(preds_split,refs_split)])
        
        return sacrebleu_score, mean_jaccard_score, smape_length_score
    
    # get different scores regarding the timestamp
    def getTimestampsScore(self, pred_timestamps, ref_timestamps):
        timestamp_amt_score = self.smapeScore(len(pred_timestamps), len(ref_timestamps))
        
        # if there are no predicted timestamps, return early. It is still consistent and monotonic.
        if (len(pred_timestamps) == 0):
            return timestamp_amt_score, 1.0, 1.0


        # replace all digits in the reference timestamp (first timestamp) with '/d' to get
        # a regex that describes the format  
        pred_timestring_pattern = re.sub(r'\d', r'\\d', re.escape(pred_timestamps[0]))
        
        matchesPatternScore = 1.0
        monotonicallyIncreasingScore = 1.0
        
        # A variable to save the previous timestamp (as datetime obj) to check monotonicity
        prev_datetime = None
        # Convert matches to datetime objects
        for i in range(len(pred_timestamps)):
            ts = pred_timestamps[i]
            try:
                # Check if the format matches with the format of the first timestamp
                # TODO!! Check this later, maybe it is too restricting for training a llm
                matchesPattern = re.fullmatch(pred_timestring_pattern, ts) is not None
                # Check if the timestamps are monotonically increasing
                cur_datetime = dateutil.parser.parse(ts)
                monotonicallyIncreasing = True if prev_datetime == None else prev_datetime <= cur_datetime
                prev_datetime = cur_datetime

                # If one entry doesn't fulfill the matching pattern property or the monotinicity property, set to 0 for whole log
                matchesPatternScore = 0.0 if (not matchesPattern) else matchesPatternScore
                monotonicallyIncreasingScore = 0.0 if (not monotonicallyIncreasing) else monotonicallyIncreasingScore
                

            except Exception as e:
                # e.g. date format not parsable by dateutil.parser
                matchesPatternScore = 0.0
                monotonicallyIncreasingScore = 0.0
        
        # matchesPatternScore and monotonicallyIncreasingScore are in {0,1}
        return timestamp_amt_score, matchesPatternScore, monotonicallyIncreasingScore



    def getLogMetric(self, pred : str, ref : str):
        ref = ref.strip(' \t\n\r')
        pred = pred.strip(' \t\n\r')

        linecount_difference_SMAPE = self.getLineCountScore(pred, ref)

        
        # Split log on timestamps
        pred_split_log = self.timestamp_pattern.split(pred)
        ref_split_log = self.timestamp_pattern.split(ref)

        # One logentry always consists of timestamp + log-message
        # pred_logentries = []
        # ref_logentries = []

        pred_timestamps = []
        pred_logMessages = []

        ref_timestamps = []
        ref_logMessages = []
        # reorganize log into logentry-tuples, consisting of timestamp + log-message
        for i in range(1, len(pred_split_log), 2):
            # pred_logentries.append((pred_split_log[i],pred_split_log[i+1]))
            pred_timestamps.append(pred_split_log[i])
            pred_logMessages.append(pred_split_log[i+1])

        
        for i in range(1, len(ref_split_log), 2):
            # ref_logentries.append((ref_split_log[i],ref_split_log[i+1]))
            ref_timestamps.append(ref_split_log[i])
            ref_logMessages.append(ref_split_log[i+1])

        # We cut off the longer list, since we only want to compute the content-scores for actually predicted lines
        # -> hence, the content score can be good, even if the amount of predicted lines are wrong
        min_logentries = min(len(pred_logMessages), len(ref_logMessages))
        # print("min_logentries:", min_logentries)
        pred_logMessages = pred_logMessages[:min_logentries]
        ref_logMessages = ref_logMessages[:min_logentries]


        linecontent_sacrebleu, linecontent_jaccard, linecontentlength_difference_SMAPE = self.getLineContentScore(pred_logMessages, ref_logMessages)
        
        timestamps_difference_SMAPE, timestamps_formatConsistency_absolute, timestamps_monotinicity_absolute = self.getTimestampsScore(pred_timestamps, ref_timestamps)

        #TODO:
        # linecontentordering_permutations = getLineContentOrderingScore(pred_logMessages, ref_logMessages)
        

        # return weighted overall score of all the different scores
        return {"linecount_difference_SMAPE_score": linecount_difference_SMAPE,
                "linecontentlength_difference_SMAPE_score": linecontentlength_difference_SMAPE,
                "linecontent_sacrebleu_score": linecontent_sacrebleu,
                "linecontent_jaccard_score": linecontent_jaccard,
                "timestamps_SMAPE_difference_score": timestamps_difference_SMAPE,
                "timestamps_formatConsistency_score": timestamps_formatConsistency_absolute,
                "timestamps_monotinicity_score": timestamps_monotinicity_absolute
                }

    def _compute(self, predictions, references):
        """Returns the scores"""

        # TODO: get separate log entries (split before timestamps), replace timestamps with token and compare the log entry with BLEU

        t_before_logmetric = time.perf_counter()
        metric_dicts = [self.getLogMetric(p,r) for p,r in zip(predictions,references)]
        # Extract keys (assuming all dictionaries have the same keys)
        keys = metric_dicts[0].keys()
        
        # Convert list of dictionaries into a 2D numpy array
        values = np.array([list(d.values()) for d in metric_dicts])
        
        # Calculate the mean along the vertical axis (axis=0)
        mean_values = np.mean(values, axis=0)
        
        # a dictionary, matching the keys with their corresponding mean values
        metric_result = dict(zip(keys, mean_values))

        t_after_logmetric = time.perf_counter()
        logmetric_duration = f"{t_after_logmetric - t_before_logmetric:0.10f}"

        metric_result["duration"] = logmetric_duration

        return metric_result