代码之家 › 专栏 › 技术社区 › erip Jigar Trivedi

为什么我的Keras模型在Iris数据集上表现如此差?

keras scikit-learn machine-learning python-3.x

0

erip Jigar Trivedi · 技术社区 · 9 年前

我正在努力 this Keras tutorial 我发现了一些有趣的东西。

import seaborn as sns
import numpy as np
from sklearn.cross_validation import train_test_split
from sklearn.linear_model import LogisticRegressionCV
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.utils import np_utils

# Load the iris dataset from seaborn.
iris = sns.load_dataset("iris")

# Use the first 4 variables to predict the species.
X, y = iris.values[:, 0:4], iris.values[:, 4]

# Split both independent and dependent variables in half
# for cross-validation
train_X, test_X, train_y, test_y = train_test_split(X, y, train_size=0.5, random_state=0)

# Train a scikit-learn log-regression model
lr = LogisticRegressionCV()
lr.fit(train_X, train_y)

# Test the model. Print the accuracy on the test data
pred_y = lr.predict(test_X)
print("Accuracy is {:.2f}".format(lr.score(test_X, test_y))) # Accuracy is 0.83

83%很好,但通过深度学习,我们应该能够做得更好。我训练一个Keras模特。。。

# Define a one-hot encoding of variables in an array.
def one_hot_encode_object_array(arr):
    '''One hot encode a numpy array of objects (e.g. strings)'''
    uniques, ids = np.unique(arr, return_inverse=True)
    return np_utils.to_categorical(ids, len(uniques))

# One-hot encode the train and test y's
train_y_ohe = one_hot_encode_object_array(train_y)
test_y_ohe = one_hot_encode_object_array(test_y)

# Build the keras model

model = Sequential()
# 4 features in the input layer (the four flower measurements)
# 16 hidden units
model.add(Dense(16, input_shape=(4,)))
model.add(Activation('sigmoid'))
# 3 classes in the ouput layer (corresponding to the 3 species)
model.add(Dense(3))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

# Train the keras model
model.fit(train_X, train_y_ohe, verbose=0, batch_size=1)

# Test the model. Print the accuracy on the test data
loss, accuracy = model.evaluate(test_X, test_y_ohe, verbose=0)
print("Accuracy is {:.2f}".format(accuracy)) # Accuracy is 0.60????

虽然这对某些数据有意义,但对于Keras序列模型来说,难以置信的线性可分离数据(如iris)应该是非常容易学习的。我已经尝试将隐藏层的数量增加到32、64和128,但精度没有提高。

以下显示了作为物种(因变量)函数的Iris数据(特别是自变量):

为什么我的模型表现这么差?

1 回复 | 直到 9 年前

1

4

erip Jigar Trivedi 9 年前

我已经替换了one_hot_encode,只使用keras' sparse_categorical_crossentropy .

显而易见的是: 增加学习时间 (默认值为10,我们试试100)。

密码

from sklearn.datasets import load_iris
from sklearn.cross_validation import train_test_split
from sklearn.linear_model import LogisticRegressionCV
import numpy as np

from keras.models import Sequential
from keras.layers import Dense, Activation

# Load the iris dataset from seaborn.
iris = load_iris()

# Use the first 4 variables to predict the species.
X, y = iris.data[:, :4], iris.target

# Split both independent and dependent variables in half
# for cross-validation
train_X, test_X, train_y, test_y = train_test_split(X, y, train_size=0.5, random_state=0)

# Train a scikit-learn log-regression model
lr = LogisticRegressionCV()
lr.fit(train_X, train_y)

# Test the model. Print the accuracy on the test data
pred_y = lr.predict(test_X)
print("Accuracy is {:.2f}".format(lr.score(test_X, test_y))) # Accuracy is 0.83


# Build the keras model

model = Sequential()
# 4 features in the input layer (the four flower measurements)
# 16 hidden units
model.add(Dense(16, input_shape=(4,)))
model.add(Activation('sigmoid'))
# 3 classes in the ouput layer (corresponding to the 3 species)
model.add(Dense(3))
model.add(Activation('softmax'))
model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

# Train the keras model
model.fit(train_X, train_y, verbose=1, batch_size=1, nb_epoch=100)

# Test the model. Print the accuracy on the test data
loss, accuracy = model.evaluate(test_X, test_y, verbose=0)
print("Accuracy is {:.2f}".format(accuracy))

输出

Accuracy is 0.83
Accuracy is 0.99