安装TensorFlow，这里直接使用当前用户安装，不需要使用root用户安装。
$ pip install tensorflow
训练一个线性回归(Linear Regression)模型，计算华氏温度
F = 9/5 * C + 32
第一步，生成训练数据：
import numpy as np import tensorflow as tf # Independent variables temp_c = [] # Dependent variables temp_f = [] # Fill the arrays with calculated values for c in range(0, 16): f = 32 + 1.8 * c temp_c.append(float(c)) temp_f.append(float(f)) # Input array input_array = np.array(temp_c, dtype=float) # Correct values for the input array output_array = np.array(temp_f, dtype=float)
第二步，创建模型
model = tf.keras.Sequential() model.add(tf.keras.layers.Input(shape=[1])) model.add(tf.keras.layers.Dense(units=1)) model.compile(optimizer=tf.keras.optimizers.SGD(), loss=tf.keras.losses.MeanSquaredError())
第三步，训练
model.fit(input_array, output_array, epochs=1000)
第4步，使用模型预测
# Temperature in Celsius temp_c = [0,1,2,3,4,5] # Predicted temperature in Fahrenheit temp_f = model.predict(np.array(temp_c)) print(temp_f)
第5步，保存模型
model.save("c_to_f_conversion.keras")将上面的代码，组合成一个simple.py运行：
$ ./simple.py WARNING: All log messages before absl::InitializeLog() is called are written to STDERR I0000 00:00:1774608384.167101 402252 cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used. I0000 00:00:1774608384.209737 402252 cpu_feature_guard.cc:227] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. WARNING: All log messages before absl::InitializeLog() is called are written to STDERR I0000 00:00:1774608385.289595 402252 cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used. E0000 00:00:1774608385.466528 402252 cuda_platform.cc:52] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303) Epoch 1/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 118ms/step - loss: 1658.9235 Epoch 2/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - loss: 712.2487 Epoch 3/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - loss: 408.5208 。。。 Epoch 96/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - loss: 97.2175 Epoch 97/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 25ms/step - loss: 96.1640 Epoch 98/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - loss: 95.1219 Epoch 99/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - loss: 94.0912 Epoch 100/100 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 28ms/step - loss: 93.0716 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 35ms/step [[13.676559] [17.256033] [20.835506] [24.41498 ] [27.994453] [31.573927]]
预测结果和实际值偏差很大，请教有经验的同事，loss没有降下来，原因是训练次数不够，将
epochs=100改成1000，结果就准确了，
[[31.864792]
[33.67792 ]
[35.491055]
[37.304184]
[39.117313]
[40.930443]]
deepseek给出了优化的代码
#!/usr/bin/python3
import numpy as np
import tensorflow as tf
# 固定种子，保证可重复性
tf.random.set_seed(42)
np.random.seed(42)
# 数据
celsius = np.arange(0, 16, dtype=np.float32)
fahrenheit = 32 + 1.8 * celsius
# 模型
model = tf.keras.Sequential([
tf.keras.layers.Dense(1, input_shape=[1])
])
# 编译：使用 SGD，小学习率
model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=0.01),
loss='mse')
# 训练，打印损失
history = model.fit(celsius, fahrenheit, epochs=2000, verbose=1)
# 输出最终权重
w, b = model.layers[0].get_weights()
print(f"训练后权重: w={w[0][0]:.6f}, b={b[0]:.6f}")
# 预测
test_c = np.array([0, 1, 2, 3, 4, 5], dtype=np.float32)
pred = model.predict(test_c).flatten()
true = 32 + 1.8 * test_c
print("\n预测值:", pred)
print("真实值:", true)
print("最大误差:", np.max(np.abs(pred - true)))可以打印出w和b值。
训练后权重: w=1.800057, b=31.999413
预测值: [31.999413 33.79947 35.599525 37.399582 39.199642 40.9997 ]
真实值: [32. 33.8 35.6 37.4 39.2 41. ]
最大误差: 0.0005874634
w值和b值，就是线性回归的， y= w *x + b
可以看到w和b接近真实的9/5和32。误差也非常小。
这里有两个值需要调优：
learning_rate=0.01
epochs=2000
larning_rate太小，则学习的很慢，太大则误差大。
epochs太小，则训练不够。
参考：
https://vitalyzhukov.com/en/tensorflow-create-train-and-use-simple-model