다층신경망(2) 미니배치

(2) 미니배치

# 필요한 모듈 및 임포트 

import numpy as np import matplotlib.pyplot as plt from sklearn.datasets import load_breast_cancer from sklearn.model_selection import train_test_split

 

 

# breast cancer 데이터 활용

cancer = load_breast_cancer() x = cancer.data y = cancer.target x_train_all, x_test, y_train_all, y_test = train_test_split(x, y, stratify=y, test_size=0.2, random_state=42) x_train, x_val, y_train, y_val = train_test_split(x_train_all, y_train_all, stratify=y_train_all,                                                   test_size=0.2, random_state=42)

 

 

# 신경망

class MinibatchNetwork(RandomInitNetwork):       def __init__(self, units=10, batch_size=32, learning_rate=0.1, l1=0, l2=0):         super().__init__(units, learning_rate, l1, l2)         self.batch_size = batch_size  # 배치 크기       def fit(self, x, y, epochs=100, x_val=None, y_val=None):         y_val = y_val.reshape(-1, 1)  # 타깃을 열 벡터로 바꿉니다.         self.init_weights(x.shape[1])  # 은닉층과 출력층의 가중치를 초기화합니다.         np.random.seed(42)         # epochs만큼 반복합니다.         for i in range(epochs):             loss = 0             # 제너레이터 함수에서 반환한 미니배치를 순환합니다.             for x_batch, y_batch in self.gen_batch(x, y):                 y_batch = y_batch.reshape(-1, 1)  # 타깃을 열 벡터로 바꿉니다.                 m = len(x_batch)  # 샘플 개수를 저장합니다.                 a = self.training(x_batch, y_batch, m)                 # 안전한 로그 계산을 위해 클리핑합니다.                 a = np.clip(a, 1e-10, 1 - 1e-10)                 # 로그 손실과 규제 손실을 더하여 리스트에 추가합니다.                 loss += np.sum(-(y_batch * np.log(a) + (1 - y_batch) * np.log(1 - a)))             self.losses.append((loss + self.reg_loss()) / len(x))             # 검증 세트에 대한 손실을 계산합니다.             self.update_val_loss(x_val, y_val)       # 미니배치 제너레이터 함수     def gen_batch(self, x, y):         length = len(x)         bins = length // self.batch_size  # 미니배치 횟수         if length % self.batch_size:             bins += 1  # 나누어 떨어지지 않을 때         indexes = np.random.permutation(np.arange(len(x)))  # 인덱스를 섞습니다.         x = x[indexes]         y = y[indexes]         for i in range(bins):             start = self.batch_size * i             end = self.batch_size * (i + 1)             yield x[start:end], y[start:end]  # batch_size만큼 슬라이싱하여 반환합니다.

 

 

# 훈련하기

minibatch_net = MinibatchNetwork(l2=0.01, batch_size=32) minibatch_net.fit(x_train_scaled, y_train, x_val=x_val_scaled, y_val=y_val, epochs=500) minibatch_net.score(x_val_scaled, y_val)

# 0.978021978021978

 

 

# 시각화

plt.plot(minibatch_net.losses) plt.plot(minibatch_net.val_losses) plt.ylabel('loss') plt.xlabel('iteration') plt.legend(['train_loss', 'val_loss']) plt.show()