Skflow API

Question

无尽的谷歌搜索让我对Python和numpy有了更好的教育，但仍然无法解决我的任务。我想读取整数/浮点值的CSV并使用神经网络预测值。我找到了几个读Iris数据集并进行分类的例子，但我不明白如何使它们适用于回归。有人可以帮我连接点吗？

以下是输入的一行：

16804,0,1,0,1,1,0,1,0,1,0,1,0,0,1,1,0,0,1,0,1,0,1,0 ，1,0,1,0,1,0,1,0,1,0,1,0,1,1,0,0,1,1,0,0,1,0,1,0,1 ，0,1,0,1,0,1,0,1,1,0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0 ，1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,1,0,0,1,0,1,0,1,0,1 ，1,0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0 ，1,0,1,0,1,0,1,1,0,0,1,0,1,0,1,0,1,0,1,0,1,1,0,0,1 ，0,0,0,1,1,0,0,1,0,0,0,0,0,1,0,1,0,0,0,0,0,1,0,1,0 ，0,0,0,0,1,0,0,0,1,0,0,0,0,1,0,0,1,0,0,0,1,0,0,0,1 ，0,0,0,1,0,0,0,0,1,0,0,0,0,0,0,1,0,0,0,0,0,0,1,0,0 ，1,0,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,0,0,0,0,0,1,0 ，0,0,0,0,0,0,0,0,0,1,0,0,1,0,0,0,0,0,0,0,1,0,0,0,0 ，0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,1,0,0,0,0 ，0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,1,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ，0,0,0,0,0,0,0,0,0,0,0,0,0,0.490265,0.620805,0.54977,0.869299,0.422268,0.351223,0.33572,0.68308,0.40455,0.47779,0.307628,0.301921 ，0.318646,0.365993,6135.81

这应该是925个值。最后一列是输出。第一个是RowID。大多数是二进制值，因为我已经完成了单热编码。测试文件没有输出/最后一列。完整的培训文件大约有10M行。一般的MxN解决方案都可以。

编辑：由于Iris是一个分类问题，让我们使用这个样本数据，但请注意以上是我真正的目标。我删除了ID列。让我们预测给出其他6列的最后一列。这有45行。（src：http://www.stat.ufl.edu/~winner/data/civwar2.dat）

100,1861,5,2,3,5,38 112,1863,11,7,4,59.82,15.18 113,1862,34,32,1,79.65,2.65 90,1862,5,2,3,68.89,5.56 93,1862,14,10,4,61.29,17.2 179,1862,22,19,3,62.01,8.89 99,1861,22,16,6,67.68,27.27 111,1862,16,11,4,78.38,8.11 107,1863,17,11,5,60.75,5.61 156,1862,32,30,2,60.9,12.82 152,1862,23,21,2,73.55,6.41 72,1863,7,3,3,54.17,20.83 134,1862,22,21,1,67.91,9.7 180,1862,23,16,4,69.44,3.89 143,1863,23,19,4,81.12,8.39 110,1862,16,12,2,31.82,9.09 157,1862,15,10,5,52.23,24.84 101,1863,4,1,3,58.42,18.81 115,1862,14,11,3,86.96,5.22 103,1862,7,6,1,70.87,0 90,1862,11,11,0,70,4.44 105,1862,20,17,3,80,4.76 104,1862,11,9,1,29.81,9.62 102,1862,17,10,7,49.02,6.86 112,1862,19,14,5,26.79,14.29 87,1862,6,3,3,8.05,72.41 92,1862,4,3,0,11.96,86.96 108,1862,12,7,3,16.67,25 86,1864,0,0,0,2.33,11.63 82,1864,4,3,1,81.71,8.54 76,1864,1,0,1,48.68,6.58 79,1864,0,0,0,15.19,21.52 85,1864,1,1,0,89.41,3.53 85,1864,1,1,0,56.47,0 85,1864,0,0,0,31.76,15.29 87,1864,6,5,0,81.61,3.45 85,1864,5,5,0,72.94,0 83,1864,0,0,0,46.99,2.38 101,1864,5,5,0,1.98,95.05 99,1864,6,6,0,42.42,9.09 10,1864,0,0,0,50,9 98,1864,6,6,0,79.59,3.06 10,1864,0,0,0,71,9 78,1864,5,5,0,70.51,1.28 89,1864,4,4,0,59.55,13.48

让我补充一点，这是一项常见的任务，但我似乎没有通过任何论坛回答，因此我已经问过这个问题。我可以给你破碎的代码，但我不想浪费你的时间来使用功能不正确的代码。对不起，我已经这样问了。我只是不了解API而且文档没有告诉我数据类型。

以下是我将CSV读入两个ndarray的最新代码：

#!/usr/bin/env python
import tensorflow as tf
import csv
import numpy as np
from numpy import genfromtxt

# Build Example Data is CSV format, but use Iris data
from sklearn import datasets
from sklearn.cross_validation import train_test_split
import sklearn
def buildDataFromIris():
    iris = datasets.load_iris()
    data = np.loadtxt(open("t100.csv.out","rb"),delimiter=",",skiprows=0)
    labels = np.copy(data)
    labels = labels[:,924]
    print "labels: ", type (labels), labels.shape, labels.ndim
    data = np.delete(data, [924], axis=1)
    print "data: ", type (data), data.shape, data.ndim

这是我想要使用的基本代码。这个来自的例子也没有完成。以下链接中的API含糊不清。如果我至少可以找出输入到DNNRegressor中的数据类型以及文档中的其他数据类型，我可能会编写一些自定义代码。

estimator = DNNRegressor(
    feature_columns=[education_emb, occupation_emb],
    hidden_units=[1024, 512, 256])

# Or estimator using the ProximalAdagradOptimizer optimizer with
# regularization.
estimator = DNNRegressor(
    feature_columns=[education_emb, occupation_emb],
    hidden_units=[1024, 512, 256],
    optimizer=tf.train.ProximalAdagradOptimizer(
      learning_rate=0.1,
      l1_regularization_strength=0.001
    ))

# Input builders
def input_fn_train: # returns x, Y
  pass
estimator.fit(input_fn=input_fn_train)

def input_fn_eval: # returns x, Y
  pass
estimator.evaluate(input_fn=input_fn_eval)
estimator.predict(x=x)

然后最重要的问题是如何让这些一起工作。

以下是我一直在关注的几页。

读取CSV并工作的基本代码（分类器）： https://www.tensorflow.org/versions/r0.11/tutorials/tflearn/index.html
回归： https://www.tensorflow.org/versions/r0.11/api_docs/python/contrib.learn.html#DNNRegressor -CSV阅读： https://www.tensorflow.org/versions/master/how_tos/reading_data/index.html#csv-files
列嵌入： https://www.tensorflow.org/versions/r0.11/tutorials/wide_and_deep/index.html
API列表（DNNRegressor，TensorFlowDNNRegressor，LinearRegressor， TensorFlowLinearRegressor，TensorFlowRNNRegressor， TensorFlowRegressor）： https://www.tensorflow.org/versions/r0.11/api_docs/python/contrib.learn.html

Answer 1

我发现较低级别的Tensorflow在过去也很难弄清楚。文档并不令人惊讶。如果您专注于抓住sklearn，您应该会发现使用skflow相对容易。 skflow的级别比tensorflow高得多，且{ap}与sklearn几乎相同。

现在回答：

作为回归示例，我们只对虹膜数据集执行回归。现在这是一个愚蠢的想法，但它只是为了演示如何使用DNNRegressor。

Skflow API

首次使用新API时，请尝试使用尽可能少的参数。你只想得到一些有用的东西。所以，我建议您可以像这样设置DNNRegressor：

estimator = skflow.DNNRegressor(hidden_units=[16, 16])

我保持#hidden单位小，因为我现在没有太多的计算能力。

然后，您可以为其提供培训数据train_X和培训标签train_y，并按照以下方式进行操作：

estimator.fit(train_X, train_y)

这是所有sklearn分类器和回归量的标准程序，skflow只是将tensorflow扩展为与sklearn类似。我还设置了参数steps = 10，以便在仅运行10次迭代时训练完成得更快。

现在，如果您希望它预测某些新数据test_X，请执行以下操作：

pred = estimator.predict(test_X)

同样，这是所有sklearn代码的标准程序。这就是它 - skflow如此简化，你只需要这三行！

train_X和train_y的格式是什么？

如果您对机器学习不太熟悉，那么您的训练数据通常是ndarray（矩阵），大小为M x d，其中您有M个训练样本和d个特征。您的标签为M x 1（形状ndarray的{{1}}）。

所以你拥有的是这样的：

(M,)

（注意我刚刚提出了所有这些数字）。

测试数据只是一个N x d矩阵，其中有N个测试示例。测试示例都需要具有d功能。预测函数将接收测试数据并返回形状为N x 1（Features: Sepal Width Sepal Length ... Labels [ 5.1 2.5 ] [0 (setosa) ] X = [ 2.3 2.4 ] y = [1 (virginica) ] [ ... ... ] [ .... ] [ 1.3 4.5 ] [2 (Versicolour)]形状ndarray）的测试标签

您没有提供.csv文件，所以我会让您将数据解析为该格式。不过很方便，我们可以使用(N,)来获取我们想要的sklearn.datsets.load_iris()和X。这只是

使用回归器作为分类器

iris = datasets.load_iris() X = iris.data y = iris.target的输出将是一堆实数（如1.6789）。但是虹膜数据集有标签0,1和2 - Setosa，Versicolour和Virginia的整数ID。要使用此回归量进行分类，我们只需将其舍入到最近的标签（0,1,2）。例如，1.6789的预测将四舍五入为2.

工作示例

我发现我用一个有效的例子来学习最多。所以这是一个非常简化的工作示例：

随意发表任何进一步的问题作为评论。

Answer 2

我最终得到了一些选择。我不知道为什么起床和跑步这么困难。首先，这是基于@ user2570465的代码。

import tensorflow as tf
import numpy as np
from sklearn import datasets
from sklearn.model_selection import train_test_split
import tensorflow.contrib.learn as skflow

def buildDataFromIris():
    iris = datasets.load_iris()
    return iris.data, iris.target

X, y = buildDataFromIris()
feature_cols = tf.contrib.learn.infer_real_valued_columns_from_input(X)
estimator = skflow.DNNRegressor( feature_columns=feature_cols, hidden_units=[10, 10])
train_X, test_X, train_y, test_y = train_test_split(X, y)
estimator.fit(X, y, steps=10)

test_preds = estimator.predict(test_X)

def CalculateAccuracy(X, y):
    continuous_predictions = estimator.predict(X)
    closest_class = []
    for pred in continuous_predictions:
        differences = np.array([abs(pred-1), abs(pred-1), abs(pred-1)])
        closest_class.append(np.argmin(differences))

    num_correct = np.sum(closest_class == y)
    accuracy = float(num_correct)/len(y)
    return accuracy

train_accuracy = CalculateAccuracy(train_X, train_y)
test_accuracy = CalculateAccuracy(test_X, test_y)

print("Train accuracy: %f" % train_accuracy)
print("Test accuracy: %f" % test_accuracy)

其他解决方案使用较小的组件构建模型。这是一个计算Sig（X * W1 + b1）* W2 + b2 = Y的片段。优化器=亚当，损失= L2，eval = L2和MSE。

x_train = X[:train_size]
y_train = Y[:train_size]
x_val = X[train_size:]
y_val = Y[train_size:]
print("x_train: {}".format(x_train.shape))

x_train = all_x[:train_size]
print("x_train: {}".format(x_train.shape))
# y_train = func(x_train)
# x_val = all_x[train_size:]
# y_val = func(x_val)

# plt.figure(1)
# plt.scatter(x_train, y_train, c='blue', label='train')
# plt.scatter(x_val, y_val, c='red', label='validation')
# plt.legend()
# plt.savefig("../img/nn_mlp1.png")


#build the model
"""
X = [
"""
X = tf.placeholder(tf.float32, [None, n_input], name = 'X')
Y = tf.placeholder(tf.float32, [None, n_output], name = 'Y')

w_h = tf.Variable(tf.random_uniform([n_input, layer1_neurons], minval=-1, maxval=1, dtype=tf.float32))
b_h = tf.Variable(tf.zeros([1, layer1_neurons], dtype=tf.float32))
h = tf.nn.sigmoid(tf.matmul(X, w_h) + b_h)

w_o = tf.Variable(tf.random_uniform([layer1_neurons, 1], minval=-1, maxval=1, dtype=tf.float32))
b_o = tf.Variable(tf.zeros([1, 1], dtype=tf.float32))
model = tf.matmul(h, w_o) + b_o

train_op = tf.train.AdamOptimizer().minimize(tf.nn.l2_loss(model - Y))
tf.nn.l2_loss(model - Y)

#output = sum((model - Y) ** 2)/2
output = tf.reduce_sum(tf.square(model - Y))/2

#launch the session
sess = tf.Session()
sess.run(tf.initialize_all_variables())

errors = []
for i in range(numEpochs):
    for start, end in zip(range(0, len(x_train), batchSize), range(batchSize, len(x_train), batchSize)):
        sess.run(train_op, feed_dict={X: x_train[start:end], Y: y_train[start:end]})
    cost = sess.run(tf.nn.l2_loss(model - y_val), feed_dict={X: x_val})
    errors.append(cost)
    if i%100 == 0: print("epoch %d, cost = %g" % (i,cost))

CSV＆gt;＆gt; Tensorflow＆gt;＆gt;回归（通过神经网络）模型

2 个答案:

Skflow API

train_X和train_y的格式是什么？

使用回归器作为分类器

工作示例