使用 MLPCLassifier，多次使用 partial_fit 比使用 fit() 产生最差的准确度，尽管数据已打乱

我正在使用 sklearn 的 MLPClassifier 在 Python 中为分类任务构建神经网络。我想绘制一条精度与时代数的关系曲线，看看我需要多少个时代才能达到某种程度的精度。我能够做到这一点的唯一方法是partial_fit()在循环中使用。这是执行此操作的代码：

from sklearn.preprocessing   import StandardScaler
from sklearn.decomposition   import PCA
from sklearn.neural_network  import MLPClassifier
import pandas as pd
import numpy  as np
import matplotlib.pyplot as plt

scaler = StandardScaler()
scaler.fit(df_train_sample)
X_train = scaler.transform(df_train_sample)
scaler.fit(df_val)
X_val = scaler.transform(df_val)

pca = PCA(pca_frac)
pca.fit(X_train)
X_train = pca.transform(X_train)
X_val = pca.transform(X_val)

n_classes = np.unique(labels_train_sample)
n_train_sample = len(df_train_sample)

scores_train = []
scores_val = []

epoch = 0
while epoch < max_iter:
   
    random_perm = np.random.permutation(n_train_sample)
    mini_batch_index = 0

    while True:
        indices = random_perm[mini_batch_index:mini_batch_index + batch_size]
        mlpc.partial_fit(X_train[indices], labels_train_sample[indices], classes=n_classes)
        mini_batch_index += batch_size

        if mini_batch_index >= n_train_sample:
            break
    
    scores_train.append(mlpc.score(X_train, labels_train_sample))
    scores_val.append(mlpc.score(X_val, labels_val))

    epoch += 1

fig, ax = plt.subplots()

ax.plot(np.arange(1, max_iter + 1), scores_train, label = "Train")
ax.plot(np.arange(1, max_iter + 1), scores_val, label = "Validation")

这里，max_iter是时代数，mlpc是分类器，定义如下：

seed          = 123
hidden_layers = [30, 15]
activation    = "relu"
learning_rate = 5e-4
beta_1        = 0.99
epsilon       = 1e-4

batch_size    = 200 
max_iter      = 200 
tol           = 1e-4

warm_start    = True
shuffle       = True

mlpc = MLPClassifier(
    hidden_layer_sizes = hidden_layers,
    activation         = activation,
    batch_size         = batch_size,
    learning_rate_init = learning_rate,
    beta_1             = beta_1,
    epsilon            = epsilon,
    warm_start         = warm_start,
    shuffle            = shuffle,
    max_iter           = max_iter,
    tol                = tol,
    random_state       = seed
)

只是可以肯定，这里是如何df_train_sample和labels_train_sample从原始数据帧结构：

df_train_sample = df_train.sample(N, replace = False).reset_index(drop = True)
labels_train_sample = labels_train[df_train_sample.index].reset_index(drop = True)

其中N是要采样的行数。df_val和labels_val是验证数据，直接从.csv文件中读取而无需修改。请注意，标签是布尔值。

问题在于，如果使用 调用该算法，则mlpc.fit()在采样数据集上的准确度约为 82%，而我发布的这段代码的准确度为 65%。这是情节：

在线搜索我发现改组数据会有所帮助，但正如您所看到的，数据已经在每个时期都进行了改组。为什么会这样？是否有另一种方式以另一种更直接的方式构建所述情节？