TensorFlow模型从训练到部署：实战步骤+避坑全攻略

import tensorflow as tf
from tensorflow.keras.layers.experimental import preprocessing

# 加载CSV数据（自动处理缺失值）
dataset = tf.data.experimental.make_csv_dataset(
    'train_data.csv',
    batch_size=32,  # 根据显存调整，小显存选16/8
    label_name='target',  # 标签列名
    na_value='?',  # 缺失值标识
    num_epochs=1,
    ignore_errors=True  # 跳过错误行
)

# 特征标准化（避免不同特征尺度影响训练）
normalizer = preprocessing.Normalization()
# 适配数据集的特征分布
normalizer.adapt(dataset.map(lambda x, y: x))  # x是特征，y是标签

model = tf.keras.Sequential([
    normalizer,  # 前面定义的标准化层
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dropout(0.3),  # 加Dropout缓解过拟合
    tf.keras.layers.Dense(32, activation='relu'),
    tf.keras.layers.Dense(10, activation='softmax')  # 多分类输出
])

model.compile(
    optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]
)

# 开始训练（记得分训练集和验证集）
history = model.fit(
    train_dataset,
    validation_data=val_dataset,  # 验证集用来监控过拟合
    epochs=20,  # 先跑20轮看看趋势
    callbacks=[tf.keras.callbacks.EarlyStopping(patience=3)]  # 早停法：3轮没提升就停止
)

import matplotlib.pyplot as plt

plt.figure(figsize=(12, 4))
# 训练loss vs 验证loss
plt.subplot(1, 2, 1)
plt.plot(history.history['loss'], label='Train Loss')
plt.plot(history.history['val_loss'], label='Val Loss')
plt.legend()
plt.title('Loss Trend')

# 训练accuracy vs 验证accuracy
plt.subplot(1, 2, 2)
plt.plot(history.history['sparse_categorical_accuracy'], label='Train Acc')
plt.plot(history.history['val_sparse_categorical_accuracy'], label='Val Acc')
plt.legend()
plt.title('Accuracy Trend')
plt.show()

lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(
    monitor='val_loss',  # 监控验证集loss
    factor=0.5,  # 学习率乘以0.5
    patience=3,  # 3轮没提升就调整
    min_lr=0.00001  # 最小学习率
)
# 训练时加入回调
history = model.fit(train_dataset, validation_data=val_dataset, epochs=20, callbacks=[lr_scheduler])

model.save('saved_model/my_model', save_format='tf')

docker run -p 8501:8501 
--mount type=bind,source=/path/to/saved_model/my_model,target=/models/my_model 
-e MODEL_NAME=my_model 
-t tensorflow/serving:latest

import requests
import numpy as np

# 生成测试数据（比如1张MNIST图片）
test_data = np.random.rand(1, 784).tolist()
# 发POST请求
response = requests.post(
    'http://localhost:8501/v1/models/my_model:predict',
    json={'instances': test_data}
)
# 解析结果
predictions = response.json()['predictions']

converter = tf.lite.TFLiteConverter.from_saved_model('saved_model/my_model')
# 量化模型（减小体积，加快推理速度）
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_model = converter.convert()

# 保存TFLite模型
with open('model.tflite', 'wb') as f:
    f.write(tflite_model)

val interpreter = Interpreter(loadModelFile(context, "model.tflite"))
// 输入数据
val input = ByteBuffer.allocateDirect(4 * 784).order(ByteOrder.nativeOrder())
input.putFloatArray(testData)
// 输出数据
val output = ByteBuffer.allocateDirect(4 * 10)
// 推理
interpreter.run(input, output)