为什么 TensorFlow 估计器无法进行这种简单的线性回归预测

Question

我目前正在学习 tensorflow，无法理解为什么 tensorflow 不能对以下简单回归问题进行正确预测。

X 是从 1000 到 8000 的随机数 Y 为 X + 250

所以如果 X 是 2000，Y 是 2250。这对我来说似乎是一个线性回归问题。然而，当我尝试进行预测时，它与我的预期相差甚远，X of 1000 给我的预测是 1048 而不是 1250。

损失和平均损失也很大：

{'average_loss': 10269.81, 'loss': 82158.48, 'global_step': 1000}

完整代码如下：

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
from sklearn.model_selection import train_test_split

x_data = np.random.randint(1000, 8000, 1000000)
y_true = x_data + 250

feat_cols = [tf.feature_column.numeric_column('x', shape=[1])]
estimator = tf.estimator.LinearRegressor(feature_columns=feat_cols)

x_train, x_eval, y_train, y_eval = train_test_split(x_data, y_true, test_size=0.3, random_state=101)

input_func = tf.estimator.inputs.numpy_input_fn({'x': x_train}, y_train, batch_size=8, num_epochs=None, shuffle=True)
train_input_func = tf.estimator.inputs.numpy_input_fn({'x': x_train}, y_train, batch_size=8, num_epochs=1000, shuffle=False)
eval_input_func = tf.estimator.inputs.numpy_input_fn({'x': x_eval}, y_eval, batch_size=8, num_epochs=1000, shuffle=False)

estimator.train(input_fn=input_func, steps=1000)

train_metrics = estimator.evaluate(input_fn=train_input_func, steps=1000)
eval_metrics = estimator.evaluate(input_fn=eval_input_func, steps=1000)

print(train_metrics)
print(eval_metrics)

brand_new_data = np.array([1000, 2000, 7000])
input_fn_predict = tf.estimator.inputs.numpy_input_fn({'x': brand_new_data}, shuffle=False)

prediction_result = estimator.predict(input_fn=input_fn_predict)

print(list(prediction_result))

是我做错了什么还是我误解了 LinearRegression 的含义？

Answer 1

我认为当您调整一些超参数时会发生这种情况。我还将优化器更改为 AdamOptimizer。

主要batch size为1，epochs为None。

train_input_func = tf.estimator.inputs.numpy_input_fn({'x': x_train}, y_train, batch_size=1, num_epochs=None, 洗牌=真）

代码：

import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split

x_data = np.random.randint(1000, 8000, 10000)
y_true = x_data + 250


feat_cols = tf.feature_column.numeric_column('x')
optimizer = tf.train.AdamOptimizer(1e-3)

estimator = tf.estimator.LinearRegressor(feature_columns=[feat_cols],optimizer=optimizer)

x_train, x_eval, y_train, y_eval = train_test_split(x_data, y_true, test_size=0.3, random_state=101)


train_input_func = tf.estimator.inputs.numpy_input_fn({'x': x_train}, y_train, batch_size=1, num_epochs=None,
                                                      shuffle=True)

eval_input_func = tf.estimator.inputs.numpy_input_fn({'x': x_eval}, y_eval, batch_size=1, num_epochs=None,
                                                     shuffle=True)

estimator.train(input_fn=train_input_func, steps=1005555)

train_metrics = estimator.evaluate(input_fn=train_input_func, steps=10000)
eval_metrics = estimator.evaluate(input_fn=eval_input_func, steps=10000)

print(train_metrics)
print(eval_metrics)

brand_new_data = np.array([1000, 2000, 7000])
input_fn_predict = tf.estimator.inputs.numpy_input_fn({'x': brand_new_data}, num_epochs=1,shuffle=False)

prediction_result = estimator.predict(input_fn=input_fn_predict)

for prediction in prediction_result:
    print(prediction['predictions'])

指标：

{'average_loss': 3.9024353e-06, 'loss': 3.9024353e-06, 'global_step': 1005555}

{'average_loss': 3.9721594e-06, 'loss': 3.9721594e-06, 'global_step': 1005555}

[1250.003]

[2250.002]

[7249.997]

Answer 2

这种缓慢收敛的原因（你需要训练 100 万步，这对于这样一个看似微不足道的问题来说很奇怪）是数据没有标准化。

应用归一化，我可以训练模型以 420 步准确预测值（我选择 420，因为它是模因的数量）以获得这些预测：

[1250.387]
[2250.2046]
[7249.2915]

代码（在 TF 2.2 中完成）：

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
from sklearn.model_selection import train_test_split

x_data = np.random.randint(1000, 8000, 1000000)
y_true = x_data + 250

tf.get_logger().setLevel('ERROR')

feat_cols = [tf.feature_column.numeric_column('size', shape=[1])]
estimator = tf.estimator.LinearRegressor(feature_columns=feat_cols, optimizer=tf.keras.optimizers.Adam(learning_rate=0.01))

def normalize(arr):
    return (arr - arr.mean()) / arr.std()

x_data_norm = normalize(x_data)
y_true_norm = normalize(y_true)

x_train, x_eval, y_train, y_eval = train_test_split(x_data_norm, y_true_norm, test_size=0.3, random_state=101)

# numpy_input_fn is for when you have the full dataset available in an array already and want a quick way to do batching/shuffling/repeating
input_func = tf.compat.v1.estimator.inputs.numpy_input_fn({'size': x_train}, y_train, batch_size=1, num_epochs=None, shuffle=True)
train_input_func = tf.compat.v1.estimator.inputs.numpy_input_fn({'size': x_train}, y_train, batch_size=1, num_epochs=None, shuffle=True)
eval_input_func = tf.compat.v1.estimator.inputs.numpy_input_fn({'size': x_eval}, y_eval, batch_size=1, num_epochs=None, shuffle=True)

estimator.train(input_fn=input_func, steps=420)

train_metrics = estimator.evaluate(input_fn=train_input_func, steps=500)
eval_metrics = estimator.evaluate(input_fn=eval_input_func, steps=500)

print(train_metrics)
print(eval_metrics)

# brand_new_data = np.array([1000, 2000, 7000])
# input_fn_predict = tf.compat.v1.estimator.inputs.numpy_input_fn({'size': brand_new_data}, num_epochs=1, shuffle=False)
# prediction_result = estimator.predict(input_fn=input_fn_predict)
# print(list(prediction_result))

# predict w/ normalization
d1 = 1000 
d2 = 2000
d3 = 7000
brand_new_data = np.array([(d1 - x_data.mean()) / x_data.std(), (d2 - x_data.mean()) / x_data.std(), (d3 - x_data.mean()) / x_data.std()])

input_fn_predict = tf.compat.v1.estimator.inputs.numpy_input_fn({'size': brand_new_data}, num_epochs=1, shuffle=False)
prediction_result = estimator.predict(input_fn=input_fn_predict)

for res in prediction_result:
  print(res['predictions'] * y_true.std() + y_true.mean())