我是机器学习和Python的新手。我已经成功地使用TensorFlow 2训练了一个线性模型。我使用了sklearn下的preprocessing.scale()函数标准化了输入数据。
问题在于我的测试数据的预测输出与实际目标值相差很大。我使用的是单层模型，优化器使用sgd，损失函数使用mean_squared_error，度量指标使用accuracy。
训练数据包含4个参数和一个目标值(273.39,420.12,91.84,57.41,1778.12)。最后一列是目标值。
我的模型预测输出与实际目标值的对比图完全不是线性的，如下图所示。
import numpy as npimport tensorflow as tfimport matplotlib.pyplot as pltfrom sklearn import preprocessing# Extract Data from CSVtrain_csv_data = np.loadtxt('train_data.csv',delimiter=',')test_csv_data = np.loadtxt('test_data.csv',delimiter=',')train_unscaled_input_all = train_csv_data[:,0:-1]train_target_all = train_csv_data[:,-1]# Shuffling the inputshuffled_indices = np.arange(train_unscaled_input_all.shape[0])np.random.shuffle(shuffled_indices)shuffled_inputs = train_unscaled_input_all[shuffled_indices]shuffled_targets = train_target_all[shuffled_indices]# Split date into train , validation and testtotal_count = shuffled_inputs.shape[0]train_data_count = int (0.8*total_count)validation_data_count = int (0.1*total_count)test_data_count = total_count - train_data_count - validation_data_counttrain_inputs = shuffled_inputs[:train_data_count]train_targets = shuffled_targets[:train_data_count]validation_inputs = shuffled_inputs[train_data_count:train_data_count+validation_data_count]validation_targets = shuffled_targets[train_data_count:train_data_count+validation_data_count]test_inputs = shuffled_inputs[train_data_count+validation_data_count:]test_targets = shuffled_targets[train_data_count+validation_data_count:]# Standardize the Inputstrain_inputs = preprocessing.scale(train_inputs)train_targets = preprocessing.scale(train_targets)validation_inputs = preprocessing.scale(validation_inputs)validation_targets = preprocessing.scale(validation_targets)# Save 3 Datasets in .npznp.savez('train',inputs = train_inputs, targets = train_targets)np.savez('validation',inputs = validation_inputs, targets = validation_targets)np.savez('test',inputs = test_inputs, targets = test_targets)train_npz = np.load('train.npz')train_inputs = train_npz['inputs']train_targets = train_npz['targets']validation_npz = np.load('validation.npz')validation_inputs = validation_npz['inputs']validation_targets = validation_npz['targets']test_npz = np.load('test.npz')test_inputs = test_npz['inputs']test_targets = test_npz['targets']# Model Creationoutput_size = 1model = tf.keras.Sequential([ tf.keras.layers.Dense(output_size) ])model.compile(optimizer='sgd', loss='mean_squared_error',metrics=['accuracy'])max_epochs = 50model.fit(train_inputs,train_targets, epochs = max_epochs, validation_data= . (validation_inputs,validation_targets), verbose=1 )predicted_outputs = (model.predict_on_batch(test_inputs))plt.plot(np.squeeze(predicted_outputs),np.squeeze(test_targets))plt.xlabel('Output')plt.ylabel('Target')plt.show()我的训练和验证准确率在整个训练过程中都为零。显然我在这里遗漏了某些东西。如果你能指出我在训练这个线性模型时遗漏了什么，那将是非常好的。
更新：训练模型
回答：
感谢你附上数据。现在我已经了解了你的问题。
我能够用你的代码重现这个问题。
你现有代码存在以下问题：
- 你对训练和验证数据进行了
preprocessing.scale处理，但没有对测试数据集进行处理。因此，测试数据的值与用于训练的值非常不同。这将无法有效工作。所以你也需要对测试数据集进行预处理。 - 你没有隐藏单元，正如@[隐藏人名]已经指出的那样，这就像一个感知器。你需要添加隐藏单元来显示一些改进。
- 在
plt中你使用了plt.plot。我建议你使用plt.scatter，因为图片会更有意义且不那么混乱。 - 你使用了
accuracy作为度量指标。由于你不是在做分类而是回归，我认为mae会是一个更好的度量指标。
当你加入这些更改后，你可以得到如下（最小的更改）的代码：
import numpy as npimport tensorflow as tfimport matplotlib.pyplot as pltfrom sklearn import preprocessing# Extract Data from CSVtrain_csv_data = np.loadtxt('train_data.csv',delimiter=',')test_csv_data = np.loadtxt('test_data.csv',delimiter=',')train_unscaled_input_all = train_csv_data[:,0:-1]train_target_all = train_csv_data[:,-1]# Shuffling the inputshuffled_indices = np.arange(train_unscaled_input_all.shape[0])np.random.shuffle(shuffled_indices)shuffled_inputs = train_unscaled_input_all[shuffled_indices]shuffled_targets = train_target_all[shuffled_indices]# Split date into train , validation and testtotal_count = shuffled_inputs.shape[0]train_data_count = int (0.8*total_count)validation_data_count = int (0.1*total_count)test_data_count = total_count - train_data_count - validation_data_counttrain_inputs = shuffled_inputs[:train_data_count]train_targets = shuffled_targets[:train_data_count]validation_inputs = shuffled_inputs[train_data_count:train_data_count+validation_data_count]validation_targets = shuffled_targets[train_data_count:train_data_count+validation_data_count]test_inputs = shuffled_inputs[train_data_count+validation_data_count:]test_targets = shuffled_targets[train_data_count+validation_data_count:]# Standardize the Inputstrain_inputs = preprocessing.scale(train_inputs)train_targets = preprocessing.scale(train_targets)validation_inputs = preprocessing.scale(validation_inputs)validation_targets = preprocessing.scale(validation_targets)test_inputs = preprocessing.scale(test_inputs)test_targets = preprocessing.scale(test_targets)# Save 3 Datasets in .npznp.savez('train',inputs = train_inputs, targets = train_targets)np.savez('validation',inputs = validation_inputs, targets = validation_targets)np.savez('test',inputs = test_inputs, targets = test_targets)train_npz = np.load('train.npz')train_inputs = train_npz['inputs']train_targets = train_npz['targets']validation_npz = np.load('validation.npz')validation_inputs = validation_npz['inputs']validation_targets = validation_npz['targets']test_npz = np.load('test.npz')test_inputs = test_npz['inputs']test_targets = test_npz['targets']# Model Creationoutput_size = 1model = tf.keras.Sequential()model.add(tf.keras.layers.Dense(1000, input_dim=4, activation='linear'))model.add(tf.keras.layers.Dense(1, activation='linear'))model.compile(optimizer='sgd', loss='mean_squared_error',metrics=['mae'])max_epochs = 50model.fit(train_inputs,train_targets, epochs = max_epochs, validation_data= (validation_inputs,validation_targets), verbose=1 )predicted_outputs = (model.predict_on_batch(test_inputs))plt.scatter(np.squeeze(predicted_outputs),np.squeeze(test_targets))plt.xlabel('Output')plt.ylabel('Target')plt.show()这会生成以下图表，图表非常线性，预测值与实际预处理值之间有很好的相关性。你需要进行某种逆变换将它们转换回原始值。
在训练过程中，损失和mae在一定程度上有所减少
