Spaces:
Running
Running
| 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, AutoEvalColumnMultimodal, Tasks, TasksMultimodal | |
| from src.submission.check_validity import is_model_on_hub | |
| class EvalResult: | |
| """Represents one full evaluation. Built from a combination of the result and request file for a given run. | |
| """ | |
| eval_name: str # org_model_track (uid) | |
| full_model: str # org/model (name of model) | |
| repo_id: str # org/model (path to model on HF) | |
| track: str | |
| org: str | |
| model: str | |
| revision: str # commit hash, "" if main | |
| results: dict | |
| date: str = "" # submission date of request file | |
| still_on_hub: bool = False | |
| 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") | |
| track = data.get("track") | |
| # Get model and org | |
| org_and_model = config.get("model_name", config.get("model_args", None)) | |
| repo_id = config.get("hf_repo", config.get("hf_repo", None)) | |
| org_and_model = org_and_model.split("/", 1) | |
| if len(org_and_model) == 1: | |
| org = None | |
| model = org_and_model[0] | |
| else: | |
| org = org_and_model[0] | |
| model = org_and_model[1] | |
| full_model = "/".join(org_and_model) | |
| eval_name = "_".join(org_and_model) + f"_{track}" | |
| still_on_hub, _, model_config = is_model_on_hub( | |
| repo_id, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False | |
| ) | |
| def _get_task_results(task): | |
| # 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]): | |
| return None | |
| mean_acc = np.mean(accs) * 100.0 | |
| return mean_acc | |
| # Extract results available in this file (some results are split in several files) | |
| results = {} | |
| if track.lower() == "multimodal": | |
| for task in TasksMultimodal: | |
| task = task.value | |
| task_result = _get_task_results(task) | |
| if task_result is not None: | |
| results[task.benchmark] = task_result | |
| else: | |
| for task in Tasks: | |
| task = task.value | |
| task_result = _get_task_results(task) | |
| if task_result is not None: | |
| results[task.benchmark] = task_result | |
| return self( | |
| eval_name=eval_name, | |
| full_model=full_model, | |
| repo_id=repo_id, | |
| track=track, | |
| org=org, | |
| model=model, | |
| results=results, | |
| revision=config.get("model_sha", ""), | |
| still_on_hub=still_on_hub, | |
| ) | |
| 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, self.track) | |
| try: | |
| with open(request_file, "r") as f: | |
| request = json.load(f) | |
| self.date = request.get("submitted_time", "") | |
| except Exception: | |
| print(f"Could not find request file for {self.org}/{self.model}") | |
| def to_dict(self): | |
| """Converts the Eval Result to a dict compatible with our dataframe display""" | |
| eval_column = AutoEvalColumnMultimodal if self.track.lower() == "multimodal" else AutoEvalColumn | |
| vision_tasks = ("VQA", "Winoground", "DevBench", "vqa", "winoground", "devbench") | |
| num_text_tasks = len(Tasks) | |
| text_average = sum([v for k, v in self.results.items() if v is not None and k not in vision_tasks]) / num_text_tasks | |
| if self.still_on_hub: | |
| model_display_name = make_clickable_model(self.repo_id, self.full_model) | |
| else: | |
| model_display_name = self.full_model | |
| data_dict = { | |
| "eval_name": self.eval_name, # not a column, just a save name, | |
| eval_column.model.name: model_display_name, | |
| eval_column.hf_repo.name: self.repo_id, | |
| eval_column.revision.name: self.revision, | |
| eval_column.text_average.name: text_average, | |
| eval_column.still_on_hub.name: self.still_on_hub, | |
| } | |
| if self.track.lower() == "multimodal": | |
| taskset = TasksMultimodal | |
| num_vision_tasks = len(TasksMultimodal) - len(Tasks) | |
| vision_average = sum([v for k, v in self.results.items() if v is not None and k in vision_tasks]) / num_vision_tasks | |
| data_dict[eval_column.vision_average.name] = vision_average | |
| else: | |
| taskset = Tasks | |
| for task in taskset: | |
| data_dict[task.value.col_name] = self.results[task.value.benchmark] | |
| return data_dict | |
| def get_request_file_for_model(requests_path, model_name, track): | |
| """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 (track) | |
| 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 |