3.6.2 非线性可分学习过程

import warnings
warnings.simplefilter("ignore")

导入功能模块

import sys

load_dataset_module_path = '../../'
sys.path.append(load_dataset_module_path)

from load_hyperspectral_dataset import (load_hyperspectral_data, y_labels,
                                        extract_features,
                                        plot_selected_categories,
                                        plot_decision_function)

#%matplotlib inline
#%matplotlib notebook

#%config InlineBackend.figure_format = 'retina'
import matplotlib as mpl
import matplotlib.pyplot as plt
from matplotlib import cm
from sklearn.utils import shuffle
import seaborn as sns
from ipywidgets import interact_manual


mpl.style.use('ggplot')

加载数据集

img_file_path = '../../Hyperspectral_Project/dc.tif'
label_file_path = '../../Hyperspectral_Project/dctest.project'

raw_X, raw_y, pixel_position = load_hyperspectral_data(img_file_path,
                                                       label_file_path)

hyperspectral_df =  extract_features(raw_X, raw_y)

画出指定类别，训练模型分类

cmap = sns.diverging_palette(200,20,sep=20,as_cmap=True)

selected_categories = ['Roofs', 'Path']

from sklearn.neural_network import MLPClassifier

model = MLPClassifier(
    hidden_layer_sizes=(200, 200, 200),
    alpha=1e-6,
    tol=1e-8,
    learning_rate='adaptive',
    n_iter_no_change=100,
    learning_rate_init=1e-3,
    beta_1=0.99,
    max_iter=1000,
    batch_size=512,
    random_state=7,
    early_stopping=False,
    warm_start=True
)


selected_features, selected_target = plot_selected_categories(
hyperspectral_df,
seleted_features=['MNDWI', 'NDBI'],
selected_categories=selected_categories,
s=10)

plt.xlim(0.65, 0.80)
plt.ylim(-0.74, -0.66)
ax = plt.gca()
ax.set_aspect(1.5)

plt.grid(b=False)

model.fit(selected_features, selected_target)
plot_decision_function(model, cmap, normalization_range=[0, 1])


plt.savefig(f'mlp_{selected_categories[0]}_{selected_categories[1]}.pdf',
            bbox_inches='tight')

绘制训练损失曲线

plt.plot(range(len(model.loss_curve_)),model.loss_curve_,label='Training loss')
plt.legend()

plt.xlabel('Epoch')
plt.ylabel('Loss')

plt.title('Training Loss Curve')
plt.savefig(f'mlp_{selected_categories[0]}_{selected_categories[1]}_training_loss_curve.pdf',
            bbox_inches='tight')

MLP不同训练周期分类边界可视化，保存为GIF动图

from sklearn.neural_network import MLPClassifier

model = MLPClassifier(
    hidden_layer_sizes=(200, 200, 200),
    alpha=1e-6,
    tol=1e-8,
    learning_rate='adaptive',
    n_iter_no_change=20,
    learning_rate_init=1e-3,
    beta_1=0.99,
    max_iter=1,
    batch_size=512,
    random_state=7,
    early_stopping=False,
    warm_start=True
)

n_iter = 1000

n_iter = 1000

for i in range(n_iter):
    selected_features, selected_target = plot_selected_categories(
    hyperspectral_df,
    seleted_features=['MNDWI', 'NDBI'],
    selected_categories=selected_categories,
    s=10)

    plt.xlim(0.65, 0.80)
    plt.ylim(-0.74, -0.66)
    ax = plt.gca()
    ax.set_aspect(1.5)
    ax.set_title(f'Epoch {i:>4d}')

    plt.grid(b=False)

    model.partial_fit(selected_features, selected_target,classes=selected_categories)
    plot_decision_function(model, cmap, normalization_range=[0, 1])



    plt.savefig(f'n_iter/mlp_{selected_categories[0]}_{selected_categories[1]}_{i}.jpg',dpi=100,
            bbox_inches='tight',quality=90)

import imageio

file_names = [
    f'n_iter/mlp_{selected_categories[0]}_{selected_categories[1]}_{i}.jpg'
    for i in range(0,n_iter,2)
]

frames = []

for image_name in file_names:
    frames.append(imageio.imread(image_name))

imageio.mimsave(
    f'training_mlp_{selected_categories[0]}_{selected_categories[1]}_2.gif',
    frames,
    'GIF',
    duration=0.01)