如何为线性回归和训练模型创建 tf.data.Datasets

Question

我可以使用 tf.data.Datasets 训练线性回归模型吗？ 如果我运行以下代码

import tensorflow as tf
import numpy as np

x = np.linspace(1, 10, num=10**2)
y = 54*x + 33

ds = tf.data.Dataset.from_tensor_slices(list(zip(x, y)))

model = tf.keras.models.Sequential([
        tf.keras.layers.Dense(1, input_shape = [1,]),
        tf.keras.layers.Dense(10, activation="sigmoid"),
        tf.keras.layers.Dense(1)
    ])

model.compile(loss="mean_absolute_error", optimizer="adam")
model.fit(ds, epochs=5)

我得到错误

ValueError: Target data is missing. Your model was compiled with loss=mean_absolute_error, and therefore expects target data to be provided in `fit()`.

可以这样训练吗？

Answer 1

你需要考虑：

1. 创建dataset ，如from_tensor_slices((x,y))
2. 使用批处理定义和获取dataset ，例如： ds = ds.batch(32)

import tensorflow as tf
import numpy as np

x = np.linspace(1, 10, num=10**2)
y = 54*x + 33

ds = tf.data.Dataset.from_tensor_slices((x,y))
ds = ds.batch(32)
model = tf.keras.models.Sequential([
        tf.keras.layers.Dense(1, input_shape = [1,]),
        tf.keras.layers.Dense(10, activation="sigmoid"),
        tf.keras.layers.Dense(1)
    ])

model.compile(loss="mean_absolute_error", optimizer="adam")
model.fit(ds, epochs=5)

输出：

Epoch 1/5
4/4 [==============================] - 0s 5ms/step - loss: 329.4714
Epoch 2/5
4/4 [==============================] - 0s 8ms/step - loss: 329.4355
Epoch 3/5
4/4 [==============================] - 0s 11ms/step - loss: 329.3994
Epoch 4/5
4/4 [==============================] - 0s 6ms/step - loss: 329.3628
Epoch 5/5
4/4 [==============================] - 0s 9ms/step - loss: 329.3259

更新：如何创建模型并训练线性回归？ 您不需要一个复杂而大型的网络，只有一个具有activation='linear'的Dense(1)就可以了。

import tensorflow as tf
import numpy as np

x = np.random.rand(10000)
y = 54*x + 33

ds = tf.data.Dataset.from_tensor_slices((x,y))
ds = ds.batch(64)
model = tf.keras.models.Sequential([
        tf.keras.layers.Dense(1, input_shape = [1,]),
        tf.keras.layers.Dense(1, activation='linear')
    ])

model.compile(loss="mean_absolute_error", optimizer="adam")
model.fit(ds, epochs=50)

---

Epoch 1/50
157/157 [==============================] - 1s 2ms/step - loss: 60.0440
Epoch 2/50
157/157 [==============================] - 0s 2ms/step - loss: 59.6723
Epoch 3/50
157/157 [==============================] - 0s 2ms/step - loss: 59.1068
...
Epoch 48/50
157/157 [==============================] - 0s 2ms/step - loss: 0.1588
Epoch 49/50
157/157 [==============================] - 0s 2ms/step - loss: 0.0053
Epoch 50/50
157/157 [==============================] - 0s 3ms/step - loss: 0.0039