Spaces:
Runtime error
Runtime error
| import glob | |
| import json | |
| import math | |
| import os | |
| from dataclasses import dataclass | |
| import dateutil | |
| import numpy as np | |
| from src.display.formatting import make_clickable_model | |
| from src.display.utils import AutoEvalColumn, Tasks | |
| class EvalResult: | |
| """Represents one full evaluation. Built from a combination of the result and request file for a given run.""" | |
| model_id: str | |
| results: dict | |
| date: str = "" # submission date of request file | |
| def init_from_json_file(self, json_filepath): | |
| """Inits the result from the specific model result file""" | |
| with open(json_filepath) as fp: | |
| data = json.load(fp) | |
| config = data.get("config") | |
| # Extract model info | |
| model = config.get("model_name", "") | |
| # Extract results available in this file (some results are split in several files) | |
| results = {} | |
| for task in Tasks: | |
| task = task.value | |
| # We average all scores of a given metric (not all metrics are present in all files) | |
| accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k]) | |
| if accs.size == 0 or any([acc is None for acc in accs]): | |
| continue | |
| mean_acc = np.mean(accs) * 100.0 | |
| results[task.benchmark] = mean_acc | |
| return self( | |
| model_id=model, | |
| results=results, | |
| ) | |
| def update_with_request_file(self, requests_path): | |
| """Finds the relevant request file for the current model and updates info with it""" | |
| request_file = get_request_file_for_model(requests_path, self.full_model) | |
| try: | |
| with open(request_file, "r") as f: | |
| request = json.load(f) | |
| self.model_id = request.get("model", self.model_id) | |
| self.results | |
| except Exception: | |
| print( | |
| f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}" | |
| ) | |
| def to_dict(self): | |
| """Converts the Eval Result to a dict compatible with our dataframe display""" | |
| data_dict = { | |
| def get_request_file_for_model(requests_path, model_name): | |
| """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED""" | |
| request_files = os.path.join( | |
| requests_path, | |
| f"{model_name}_eval_request_*.json", | |
| ) | |
| request_files = glob.glob(request_files) | |
| # Select correct request file | |
| request_file = "" | |
| request_files = sorted(request_files, reverse=True) | |
| for tmp_request_file in request_files: | |
| with open(tmp_request_file, "r") as f: | |
| req_content = json.load(f) | |
| if req_content["status"] in ["FINISHED"]: | |
| request_file = tmp_request_file | |
| return request_file | |
| def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]: | |
| """From the path of the results folder root, extract all needed info for results""" | |
| model_result_filepaths = [] | |
| for root, _, files in os.walk(results_path): | |
| # We should only have json files in model results | |
| if len(files) == 0 or any([not f.endswith(".json") for f in files]): | |
| continue | |
| # Sort the files by date | |
| try: | |
| files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7]) | |
| except dateutil.parser._parser.ParserError: | |
| files = [files[-1]] | |
| for file in files: | |
| model_result_filepaths.append(os.path.join(root, file)) | |
| eval_results = {} | |
| for model_result_filepath in model_result_filepaths: | |
| # Creation of result | |
| eval_result = EvalResult.init_from_json_file(model_result_filepath) | |
| eval_result.update_with_request_file(requests_path) | |
| # Store results of same eval together | |
| eval_name = eval_result.eval_name | |
| if eval_name in eval_results.keys(): | |
| eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None}) | |
| else: | |
| eval_results[eval_name] = eval_result | |
| results = [] | |
| for v in eval_results.values(): | |
| try: | |
| v.to_dict() # we test if the dict version is complete | |
| results.append(v) | |
| except KeyError: # not all eval values present | |
| continue | |
| return results | |