P7-3.3 실제 분류기로 입력 표현을 다시 비교하기¶

import csv
from pathlib import Path

import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score

data_path = Path("docs/assets/part-07/chapter-03/p7-3-surface-patches.csv")
rows = list(csv.DictReader(data_path.open(encoding="utf-8")))
pixel_columns = [name for name in rows[0] if name.startswith("pixel_")]

train_rows = [row for row in rows if row["split"] == "train"]
test_rows = [row for row in rows if row["split"] == "test"]

def pixel_matrix(selected_rows):
    return np.array(
        [[float(row[column]) for column in pixel_columns] for row in selected_rows],
        dtype=float,
    )

def column_profile(matrix):
    # 8x8 공간을 열 평균 8개로 줄이면 세로 결함의 큰 흐름만 남는다.
    images = matrix.reshape(len(matrix), 8, 8)
    return images.mean(axis=1)

def center_band_profile(matrix):
    # 가운데 결함 후보 영역과 양쪽 주변 밝기만 비교하는 강한 축약 표현이다.
    images = matrix.reshape(len(matrix), 8, 8)
    center = images[:, :, 3:5].mean(axis=(1, 2))
    left = images[:, :, :3].mean(axis=(1, 2))
    right = images[:, :, 5:].mean(axis=(1, 2))
    return np.column_stack([center, left, right])

raw_train = pixel_matrix(train_rows)
raw_test = pixel_matrix(test_rows)
y_train = np.array([int(row["label"]) for row in train_rows])
y_test = np.array([int(row["label"]) for row in test_rows])

representations = {
    "64개 픽셀 그대로": (raw_train, raw_test),
    "열 평균 8개": (column_profile(raw_train), column_profile(raw_test)),
    "중심 band 3개": (center_band_profile(raw_train), center_band_profile(raw_test)),
}

experiment_summaries = []
sample_records = []

for representation_name, (x_train, x_test) in representations.items():
    model = LogisticRegression(max_iter=1000, random_state=7)
    model.fit(x_train, y_train)

    predictions = model.predict(x_test)
    probabilities = model.predict_proba(x_test)
    margins = np.abs(probabilities[:, 1] - probabilities[:, 0])

    errors = []
    low_margin_samples = []

    for row, actual, predicted, probability, margin in zip(
        test_rows, y_test, predictions, probabilities, margins
    ):
        record = {
            "표현": representation_name,
            "샘플": row["sample"],
            "패턴": row["pattern_name"],
            "실제": int(actual),
            "예측": int(predicted),
            "클래스별 확률": [round(float(value), 3) for value in probability],
            "확신 차이": round(float(margin), 3),
        }
        sample_records.append(record)

        if actual != predicted:
            errors.append(row["sample"])
        if margin < 0.25:
            low_margin_samples.append(row["sample"])

    experiment_summaries.append(
        {
            "표현": representation_name,
            "학습 입력 shape": tuple(x_train.shape),
            "평가 정확도": round(float(accuracy_score(y_test, predictions)), 3),
            "오류 샘플": errors,
            "낮은 확신 샘플": low_margin_samples,
        }
    )

print("표현별 실행 요약 =")
for summary in experiment_summaries:
    print(summary)

print("평가 샘플별 기록 =")
for record in sample_records:
    print(record)

표현별 실행 요약 =
{'표현': '64개 픽셀 그대로', '학습 입력 shape': (12, 64), '평가 정확도': 0.75, '오류 샘플': ['평가-결함-약함'], '낮은 확신 샘플': ['평가-결함-약함']}
{'표현': '열 평균 8개', '학습 입력 shape': (12, 8), '평가 정확도': 0.75, '오류 샘플': ['평가-결함-약함'], '낮은 확신 샘플': ['평가-정상-안정', '평가-결함-명확', '평가-결함-약함', '평가-정상-그림자']}
{'표현': '중심 band 3개', '학습 입력 shape': (12, 3), '평가 정확도': 0.75, '오류 샘플': ['평가-결함-약함'], '낮은 확신 샘플': ['평가-정상-안정', '평가-결함-명확', '평가-결함-약함', '평가-정상-그림자']}