YOLOv8: Optimising for map with confidence and iou in prediction

Question

I'm trying to figure out what the best conf and iou is for the model.pred.

from ultralytics import YOLO
import pandas as pd
import numpy as np


df= pd.DataFrame()

# Load a model
for i in range(1,105):
  print('epoch: ',i)
  try:
    
    model = YOLO(f'/content/weights/epoch{i}.pt')


    for confidence in np.arange(0.1,0.4,0.02):
      for inter in np.arange(0.1,0.8,0.05):  
        # Customize validation settings
        validation_results = model.val(data='/content/myproject/data.yaml',
                                      imgsz=640,
                                      batch=16,
                                      conf=confidence,
                                      iou=inter,
                                      device='cpu')

        t = pd.DataFrame([{"epoch":i,
                          "conf":confidence,
                          "iou":inter,
                          "map50":validation_results.box.map50}])
        df = df.append(t, ignore_index = True)

    print(df.sort_values(by=['map50'],ascending=False).head(3))
  except:
    pass

above is my attempt at trying every combination and sorting this by the highest mean average precision (MAP).

This is slow because it's trying every combination. Maybe a package like Optuna or other Bayesian packages can be used? What would you do to optimise this scenario?

Answer 1

An answer using Optuna.

The following will setup an objective trialling various iou and conf until it finds the maximum MAP50 score. This will lead to a better understanding of the iou and conf values and how they affect the map score.

1. Install Optuna

!pip install optuna

2. Setup Objective

from ultralytics import YOLO
import pandas as pd
import numpy as np

class Objective:

    def __init__(self):
        self.best_map = 0

    def __call__(self, trial):

        i = trial.suggest_int("epoch", 1, 104)
        confidence = trial.suggest_float("confidence", 0.05, 0.5)
        inter = trial.suggest_float("iou", 0.1, 0.8)
        model = YOLO(f'/content/weights/epoch{i}.pt')
        validation_results = model.val(data='/content/myproject/data.yaml',
                              imgsz=640,
                              batch=16,
                              conf=confidence,
                              iou=inter,
                              device='cpu')

        print(validation_results.box.map50)
        self._map = float(validation_results.box.map50)
        map = float(validation_results.box.map50)
        return map

    def callback(self, study, trial):
        if study.best_trial == trial:
            self.best_map = self._map
            print('NEW BEST MAP: ', self._map)

3. start trail

import warnings
warnings.filterwarnings("ignore", category=RuntimeWarning) # for log error

import optuna
objective = Objective()

# Setting SEED 
from optuna.samplers import TPESampler
sampler = TPESampler(seed=10)

study = optuna.create_study(
    pruner=optuna.pruners.MedianPruner(n_warmup_steps=10), direction="maximize",
    sampler=sampler
)
study.optimize(objective, n_trials=1000, callbacks=[objective.callback])

print("Best trial:")
trial = study.best_trial

print("  Params: ")
for key, value in trial.params.items():
    print("    {}: {}".format(key, value))