src/leaderboard/read_evals.py

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, ModelType, ModelArch, Precision, HarnessTasks, WeightType, ClinicalTypes
from src.submission.check_validity import is_model_on_hub


@dataclass
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_precision (uid)
    full_model: str  # org/model (path on hub)
    org: str
    model: str
    revision: str  # commit hash, "" if main
    dataset_results: dict
    is_domain_specific: bool
    use_chat_template: bool
    # clinical_type_results:dict
    precision: Precision = Precision.Unknown
    model_type: ModelType = ModelType.Unknown  # Pretrained, fine tuned, ...
    weight_type: WeightType = WeightType.Original  # Original or Adapter
    backbone:str = "Unknown"
    license: str = "?"
    likes: int = 0
    num_params: int = 0
    date: str = ""  # submission date of request file
    still_on_hub: bool = False
    display_result:bool = True

    @classmethod
    def init_from_json_file(self, json_filepath, evaluation_metric):
        """Inits the result from the specific model result file"""
        with open(json_filepath) as fp:
            data = json.load(fp)
 
        config = data.get("config")

        # Precision
        precision = Precision.from_str(config.get("model_dtype"))
        model_type = ModelType.from_str(config.get("model_type", ""))
        license = config.get("license", "?")
        num_params = config.get("num_params", "?")
        display_result = config.get("display_result", True)
        display_result = False if display_result=="False" else True

        # Get model and org
        org_and_model = config.get("model_name", config.get("model_args", None))
        org_and_model = org_and_model.split("/", 1)

        if len(org_and_model) == 1:
            org = None
            model = org_and_model[0]
            result_key = f"{model}_{precision.value.name}"
        else:
            org = org_and_model[0]
            model = org_and_model[1]
            result_key = f"{org}_{model}_{precision.value.name}"
        full_model = "/".join(org_and_model)

        still_on_hub, _, model_config = is_model_on_hub(
            full_model, config.get("revision", "main"), trust_remote_code=True, test_tokenizer=False
        )
        backbone = "?"
        if model_config is not None:
            backbones = getattr(model_config, "architectures", None)
            if backbones:
                backbone = ";".join(backbones)

        # Extract results available in this file (some results are split in several files)
        dataset_results = {}
        for task in HarnessTasks:
            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
            dataset_results[task.benchmark] = mean_acc
        print(dataset_results)
        # types_results = {}
        # for clinical_type in ClinicalTypes:
        #     clinical_type = clinical_type.value

        #     # We average all scores of a given metric (not all metrics are present in all files)
        #     accs = np.array([v.get(clinical_type.metric, None) for k, v in data[evaluation_metric]["clinical_type_results"].items() if clinical_type.benchmark == k])
        #     if accs.size == 0 or any([acc is None for acc in accs]):
        #         continue

        #     mean_acc = np.mean(accs)  # * 100.0
        #     types_results[clinical_type.benchmark] = mean_acc

        return self(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            revision=config.get("revision", ""),
            dataset_results=dataset_results,
            is_domain_specific=config.get("is_domain_specific", False),  # Assuming a default value
            use_chat_template=config.get("use_chat_template", False),  # Assuming a default value
            precision=precision,
            model_type=model_type,
            weight_type=WeightType.from_str(config.get("weight_type", "")),  # Assuming the default value
            backbone=backbone,
            license=license,
            likes=config.get("likes", 0),  # Assuming a default value
            num_params=num_params,
            still_on_hub=still_on_hub,
            display_result=display_result,
            date=config.get("submitted_time","")
        )

    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.precision.value.name)

        try:
            with open(request_file, "r") as f:
                request = json.load(f)
            self.model_type = ModelType.from_str(request.get("model_type", ""))
            self.weight_type = WeightType[request.get("weight_type", "Original")]
            self.license = request.get("license", "?")
            self.likes = request.get("likes", 0)
            self.num_params = request.get("params", 0)
            self.date = request.get("submitted_time", "")
            # self.precision = request.get("precision", "float32")
        except Exception:
            pass
            # print(
            #     f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}"
            # )
            # print(f" Args used were - {request_file=}, {requests_path=}, {self.full_model=},")

    def to_dict(self, subset):
        """Converts the Eval Result to a dict compatible with our dataframe display"""
        if subset == "datasets":
            average = sum([v for v in self.dataset_results.values() if v is not None]) / len(HarnessTasks)
            data_dict = {
                "eval_name": self.eval_name,  # not a column, just a save name,
                AutoEvalColumn.precision.name: self.precision.value.name,
                AutoEvalColumn.model_type.name: self.model_type.value.name,
                AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol + (" 🏥" if self.is_domain_specific else ""),
                AutoEvalColumn.weight_type.name: self.weight_type.value.name,
                # AutoEvalColumn.architecture.name: self.architecture.value.name,
                # AutoEvalColumn.backbone.name: self.backbone,
                AutoEvalColumn.model.name: make_clickable_model(self.full_model),
                AutoEvalColumn.is_domain_specific.name: self.is_domain_specific,
                AutoEvalColumn.use_chat_template.name: self.use_chat_template,
                AutoEvalColumn.revision.name: self.revision,
                AutoEvalColumn.average.name: average,
                AutoEvalColumn.license.name: self.license,
                AutoEvalColumn.likes.name: self.likes,
                AutoEvalColumn.params.name: self.num_params,
                AutoEvalColumn.still_on_hub.name: self.still_on_hub,
                AutoEvalColumn.date.name: self.date,
                "display_result" : self.display_result,
            }

            for task in HarnessTasks:
                data_dict[task.value.col_name] = self.dataset_results[task.value.benchmark]

            return data_dict
        
        if subset == "clinical_types":
            average = sum([v for v in self.clinical_type_results.values() if v is not None]) / len(ClinicalTypes)
            data_dict = {
                "eval_name": self.eval_name,  # not a column, just a save name,
                AutoEvalColumn.precision.name: self.precision.value.name,
                AutoEvalColumn.model_type.name: self.model_type.value.name,
                AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
                AutoEvalColumn.weight_type.name: self.weight_type.value.name,
                AutoEvalColumn.architecture.name: self.architecture.value.name,
                AutoEvalColumn.backbone.name: self.backbone,
                AutoEvalColumn.model.name: make_clickable_model(self.full_model),
                AutoEvalColumn.revision.name: self.revision,
                AutoEvalColumn.average.name: average,
                AutoEvalColumn.license.name: self.license,
                AutoEvalColumn.likes.name: self.likes,
                AutoEvalColumn.params.name: self.num_params,
                AutoEvalColumn.still_on_hub.name: self.still_on_hub,
                "display_result" : self.display_result,
            }

            for clinical_type in ClinicalTypes:
                data_dict[clinical_type.value.col_name] = self.clinical_type_results[clinical_type.value.benchmark]

            return data_dict
        


def get_request_file_for_model(requests_path, model_name, precision):
    """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 (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_file


def get_raw_eval_results(results_path: str, requests_path: str, evaluation_metric: 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, evaluation_metric)
        # 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 = []
    # clinical_type_results = []
    for v in eval_results.values():
        try:
            v.to_dict(subset="dataset")  # we test if the dict version is complete
            if not v.display_result:
                continue
            results.append(v)
        except KeyError:  # not all eval values present
            continue

    return results