File size: 5,515 Bytes
aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 eb8e45b aa37927 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 |
import glob
import json
import os
from dataclasses import dataclass
import numpy as np
from src.display.formatting import make_clickable_model
from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
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
results: dict
average_accuracy: float
precision: Precision = Precision.Unknown
model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
weight_type: WeightType = 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(cls, 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", {})
# Precision
precision = Precision.from_str(config.get("model_dtype", "Unknown"))
# Get model and org
org_and_model = config.get("model_name", "").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)
results_data = data.get("results", {})
# Extract per-subject accuracies
per_subject_results = {}
for task in Tasks:
subject = task.value.benchmark
accuracy = results_data.get(subject, None)
if accuracy is not None:
per_subject_results[subject] = accuracy
average_accuracy = results_data.get('average', None)
# Set other fields from config
model_type = ModelType.from_str(config.get("model_type", ""))
weight_type = WeightType[config.get("weight_type", "Original")]
license = config.get("license", "?")
likes = config.get("likes", 0)
num_params = config.get("params", 0)
date = config.get("submitted_time", "")
still_on_hub = config.get("still_on_hub", True)
architecture = config.get("architecture", "Unknown")
# Create EvalResult instance
return cls(
eval_name=result_key,
full_model=full_model,
org=org,
model=model,
results=per_subject_results,
average_accuracy=average_accuracy,
precision=precision,
revision=config.get("model_sha", ""),
still_on_hub=still_on_hub,
architecture=architecture,
model_type=model_type,
weight_type=weight_type,
license=license,
likes=likes,
num_params=num_params,
date=date,
)
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,
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,
AutoEvalColumn.model.name: make_clickable_model(self.full_model),
AutoEvalColumn.revision.name: self.revision,
AutoEvalColumn.average.name: self.average_accuracy,
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,
}
for task in Tasks:
subject = task.value.benchmark
data_dict[task.value.col_name] = self.results.get(subject, None)
return data_dict
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
for file in files:
if file.endswith(".json"):
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)
# Store results
eval_name = eval_result.eval_name
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
|