import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping, LambdaCallback
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
matplotlib.rc("font", family="FangSong")
import seaborn as sns
from sklearn.metrics import confusion_matrix, classification_report
import os
# 确认 GPU 是否可用
print("GPU 可用数量:", len(tf.config.list_physical_devices('GPU')))
if tf.config.list_physical_devices('GPU'):
print("GPU 已启用！")
else:
print("⚠️ 未检测到 GPU，将使用 CPU 训练")
# 设置数据集缓存路径
os.environ['KERAS_HOME'] = os.path.join(os.getcwd(), 'datasets')
# 加载并预处理 CIFAR-10 数据
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
#LeNET模型
model = models.Sequential([
# C1: 卷积层 - 6 个 5x5 卷积核
layers.Conv2D(6, kernel_size=(5, 5), activation='tanh', input_shape=(32, 32, 3)),
# S2: 平均池化
layers.AveragePooling2D(pool_size=(2, 2), strides=2),
# C3: 卷积层 - 16 个 5x5 卷积核
layers.Conv2D(16, kernel_size=(5, 5), activation='tanh'),
# S4: 平均池化
layers.AveragePooling2D(pool_size=(2, 2), strides=2),
# C5: 全连接式卷积层
layers.Conv2D(120, kernel_size=(5, 5), activation='tanh'),
# 展平
layers.Flatten(),
# F6: 全连接层
layers.Dense(84, activation='tanh'),
# 输出层
layers.Dense(10, activation='softmax')
])
# 编译模型（只保留一次）
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
loss='categorical_crossentropy', # 因为 y 是 one-hot 编码
metrics=['accuracy'])
# 打印模型结构
model.summary()
# 定义回调函数
reduce_lr = ReduceLROnPlateau(
monitor='val_loss',
factor=0.5,
patience=3,
min_lr=1e-7,
verbose=1
)
early_stopping = EarlyStopping(
monitor='val_loss',
patience=5,
restore_best_weights=True,
verbose=1
)
# 记录学习率
lr_list = []
lr_callback = LambdaCallback(
on_epoch_end=lambda epoch, logs: lr_list.append(model.optimizer.learning_rate.numpy())
)
# 开始训练
print("开始训练...")
history = model.fit(
x_train, y_train,
epochs=50,
batch_size=64,
validation_data=(x_test, y_test),
callbacks=[reduce_lr, early_stopping, lr_callback],
verbose=1
)
# 评估模型
test_loss, test_acc = model.evaluate(x_test, y_test, verbose=0)
print(f"\n测试准确率: {test_acc:.4f}")
print(f"测试损失: {test_loss:.4f}")
# 绘制训练曲线
plt.figure(figsize=(15, 5))
# 准确率
plt.subplot(1, 3, 1)
plt.plot(history.history['accuracy'], label='训练准确率')
plt.plot(history.history['val_accuracy'], label='验证准确率')
plt.title('模型准确率')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()
# 损失
plt.subplot(1, 3, 2)
plt.plot(history.history['loss'], label='训练损失')
plt.plot(history.history['val_loss'], label='验证损失')
plt.title('模型损失')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
# 学习率变化
plt.subplot(1, 3, 3)
plt.plot(lr_list, label='学习率', color='purple')
plt.title('学习率变化')
plt.xlabel('Epoch')
plt.ylabel('Learning Rate')
plt.yscale('log')
plt.legend()
plt.tight_layout()
plt.show()
# 预测与混淆矩阵
y_pred_probs = model.predict(x_test)
y_pred_classes = np.argmax(y_pred_probs, axis=1)
y_true_classes = np.argmax(y_test, axis=1)
cm = confusion_matrix(y_true_classes, y_pred_classes)
plt.figure(figsize=(10, 8))
sns.heatmap(cm, annot=True, fmt='d', cmap='Blues',
xticklabels=[f'类{i}' for i in range(10)],
yticklabels=[f'类{i}' for i in range(10)])
plt.title('混淆矩阵')
plt.xlabel('预测标签')
plt.ylabel('真实标签')
plt.show()
# 错误分类样本可视化
error_indices = np.where(y_pred_classes != y_true_classes)[0]
print(f"共 {len(error_indices)} 个样本被错误分类")
num_show = min(12, len(error_indices))
plt.figure(figsize=(15, 6))
for i in range(num_show):
idx = error_indices[i]
plt.subplot(2, 6, i + 1)
plt.imshow(x_test[idx])
plt.title(f'真:{y_true_classes[idx]}, 预:{y_pred_classes[idx]}')
plt.axis('off')
plt.suptitle("错误分类的样本示例")
plt.tight_layout()
plt.show()
# 分类报告
print("分类报告:")
print(classification_report(y_true_classes, y_pred_classes, target_names=[f'类 {i}' for i in range(10)]))
改为Pytorch
最新发布
本文解析了BERT代码中用Tensorflow1.x实现的loss function，重点讲解了负对数似然(NLL)的概念，并结合实例说明了如何计算和应用在模型训练中。
扫一扫
1379
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
