import glob import json import os from dataclasses import dataclass import numpy as np import dateutil import src.display.formatting as formatting import src.display.utils as utils import src.submission.check_validity as check_validity @dataclass class EvalResult: eval_name: str # org_model_precision (uid) full_model: str # org/model (path on hub) org: str model: str revision: str # commit hash, "" if main results: dict precision: utils.Precision = utils.Precision.Unknown model_type: utils.ModelType = utils.ModelType.Unknown # Pretrained, fine tuned, ... weight_type: utils.WeightType = utils.WeightType.Original # Original or Adapter architecture: str = "Unknown" license: str = "?" likes: int = 0 num_params: int = 0 date: str = "" # submission date of request file still_on_hub: bool = False @classmethod 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) print('json_filepath',json_filepath) print(data) config = data.get("config") print(config) # Precision precision = utils.Precision.from_str(config.get("model_dtype")) # Get model and org full_model = config.get("model_name", config.get("model_args", None)) org, model = full_model.split("/", 1) if "/" in full_model else (None, full_model) if org: result_key = f"{org}_{model}_{precision.value.name}" else: result_key = f"{model}_{precision.value.name}" still_on_hub, _, model_config = check_validity.is_model_on_hub( full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False) if model_config: architecture = ";".join(getattr(model_config, "architectures", ["?"])) else: architecture = "?" # Extract results available in this file (some results are split in several files) results = {} for task in utils.Tasks: #print(task) task = task.value #print(task.benchmark) #print(task.metric) #print(task.col_name) #print(task.value) if isinstance(task.metric, str): # accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if # task.benchmark == k and isinstance(v, dict)]) # accs = np.array([np.around(v*100, decimals=0) for k, v in data["results"].items() if task.benchmark == k]) accs = [] import math for k, v in data["results"].items(): if task.benchmark == k: if isinstance(v, (int, float)) and not math.isnan(v): accs.append(np.around(v * 100, decimals=0)) elif isinstance(v, list): accs.extend([np.around(x * 100, decimals=0) for x in v if isinstance(x, (int, float)) and not math.isnan(x)]) else: # 跳过 NaN 或不符合条件的值 accs.append(None) accs = np.array([x for x in accs if x is not None]) accs = accs[accs != None] results[task.benchmark] = accs elif isinstance(task.metric, list): accs = np.array([str(v.get(task.metric, None)) for k, v in data["results"].items() if task.benchmark == k and isinstance(v, dict)]) accs = accs[accs != None] results[task.benchmark] = accs else: print(f"Skipping task with unhandled metric type: {type(task.metric)}") # # 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]) # # results[task.benchmark] = accs return self( eval_name=result_key, full_model=full_model, org=org, model=model, results=results, precision=precision, revision= config.get("model_sha", ""), still_on_hub=still_on_hub, architecture=architecture ) def update_with_request_file(self, requests_path): """Finds the relevant request file for the current model and updates info with it""" all_files_before = os.listdir(requests_path) print("test the variable:", all_files_before) print(self.full_model) #print(self.precision.value.name) request_file = get_request_file_for_model(requests_path, self.full_model) print("file name:",request_file) #all_files = os.listdir(request_file) #print("Files in the folder:", all_files) try: with open(request_file, "r") as f: request = json.load(f) print(request) self.model_type = utils.ModelType.from_str(request.get("model_type", "")) #self.weight_type = utils.WeightType[request.get("weight_type", "Original")] self.license = request.get("license", "?") self.likes = request.get("likes", 0) self.num_params = int(float(request.get("params", "0").replace('B', ''))) self.date = request.get("submitted_time", "") # print(self.license) print('updated:', self) except FileNotFoundError: print(f"Could not find request file for {self.org}/{self.model}") except json.JSONDecodeError: print(f"Error decoding JSON in request file for {self.org}/{self.model}") def to_dict(self): """Converts the Eval Result to a dict compatible with our dataframe display""" data_dict = { "eval_name": self.eval_name, # not a column, just a save name, utils.AutoEvalColumn.precision.name: self.precision.value.name, utils.AutoEvalColumn.model_type.name: self.model_type.value.name, #utils.AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol, utils.AutoEvalColumn.weight_type.name: self.weight_type.value.name, utils.AutoEvalColumn.architecture.name: self.architecture, utils.AutoEvalColumn.model.name: formatting.make_clickable_model(self.full_model), utils.AutoEvalColumn.dummy.name: self.full_model, # utils.AutoEvalColumn.revision.name: self.revision, utils.AutoEvalColumn.license.name: self.license, utils.AutoEvalColumn.likes.name: self.likes, utils.AutoEvalColumn.params.name: self.num_params, utils.AutoEvalColumn.still_on_hub.name: self.still_on_hub, } for task in utils.Tasks: data_dict[task.value.col_name] = self.results[task.value.benchmark] return 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}.json", ) #request_files = glob.glob(request_files) # Select correct request file (precision) # 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"] # # and req_content["precision"] == precision.split(".")[-1] # # ): # # request_file = tmp_request_file return request_files 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 = [] print("results_path", results_path) for root, _, files in os.walk(results_path): print("file",files) for f in files: if f.endswith(".json"): model_result_filepaths.extend([os.path.join(root, f)]) print("model_result_filepaths:", model_result_filepaths) # exit() eval_results = {} for model_result_filepath in model_result_filepaths: # Creation of result eval_result = EvalResult.init_from_json_file(model_result_filepath) print("request_path:",requests_path) eval_result.update_with_request_file(requests_path) print(eval_result) # 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