Python 尝试使用简单的Keras神经网络示例

我一直在胡闹，试图让我创建的简单示例发挥作用，因为我发现给出的大型复杂数据集示例很难直观地理解。下面的程序获取权重列表

[x_0 x_1…x_n]

，并使用它们在平面上创建点的随机散射，同时添加一些随机噪声。然后，我根据这些数据训练简单的神经网络，并检查结果

当我使用图形模型进行此操作时，一切都很完美，当模型在给定的权重上收敛时，损失分数可预测地降到零。然而，当我尝试使用顺序模型时，什么也没有发生。代码如下

如果您愿意，我可以发布我的另一个脚本，它使用图形而不是顺序，并显示它可以完美地找到输入权重

#!/usr/bin/env python
from keras.models import Sequential, Graph
from keras.layers.core import Dense, Dropout, Activation
from keras.optimizers import SGD
import numpy as np
import theano, sys

NUM_TRAIN = 100000
NUM_TEST = 10000
INDIM = 3

mn = 1

def myrand(a, b) :
    return (b)*(np.random.random_sample()-0.5)+a

def get_data(count, ws, xno, bounds=100, rweight=0.0) :
    xt = np.random.rand(count, len(ws))
    xt = np.multiply(bounds, xt)
    yt = np.random.rand(count, 1)
    ws = np.array(ws, dtype=np.float)
    xno = np.array([float(xno) + rweight*myrand(-mn, mn) for x in xt], dtype=np.float)
    yt = np.dot(xt, ws)
    yt = np.add(yt, xno)

    return (xt, yt)


if __name__ == '__main__' :
    if len(sys.argv) > 1 :
       EPOCHS = int(sys.argv[1])
       XNO = float(sys.argv[2])
       WS = [float(x) for x in sys.argv[3:]]
       mx = max([abs(x) for x in (WS+[XNO])])
       mn = min([abs(x) for x in (WS+[XNO])])
       mn = min(1, mn)
       WS = [float(x)/mx for x in WS]
       XNO = float(XNO)/mx
       INDIM = len(WS)
    else :
        INDIM = 3
        WS = [2.0, 1.0, 0.5]
        XNO = 2.2

    X_test, y_test = get_data(10000, WS, XNO, 10000, rweight=0.4)
    X_train, y_train = get_data(100000, WS, XNO, 10000)

    model = Sequential()
    model.add(Dense(INDIM, input_dim=INDIM, init='uniform', activation='tanh'))
    model.add(Dropout(0.5))
    model.add(Dense(2, init='uniform', activation='tanh'))
    model.add(Dropout(0.5))
    model.add(Dense(1, init='uniform', activation='softmax'))

    sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
    model.compile(loss='mean_squared_error', optimizer=sgd)

    model.fit(X_train, y_train, shuffle="batch", show_accuracy=True, nb_epoch=EPOCHS)
    score, acc = model.evaluate(X_test, y_test, batch_size=16, show_accuracy=True)
    print score
    print acc

    predict_data = np.random.rand(100*100, INDIM)
    predictions = model.predict(predict_data)

    for x in range(len(predict_data)) :
        print "%s --> %s" % (str(predict_data[x]), str(predictions[x]))

结果如下

$ ./keras_hello.py 20 10 5 4 3 2 1
Using gpu device 0: GeForce GTX 970 (CNMeM is disabled)
Epoch 1/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 2/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 3/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 4/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 5/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 6/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 7/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 8/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 9/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 10/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 11/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 12/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 13/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 14/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 15/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 16/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 17/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 18/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 19/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
Epoch 20/20
100000/100000 [==============================] - 0s - loss: 60726734.3061 - acc: 1.0000     
10000/10000 [==============================] - 0s     
60247198.6661
1.0
[ 0.06698217  0.70033048  0.4317502   0.78504855  0.26173543] --> [ 1.]
[ 0.28940025  0.21746189  0.93097653  0.94885535  0.56790348] --> [ 1.]
[ 0.69430499  0.1622601   0.22802859  0.75709315  0.88948355] --> [ 1.]
[ 0.90714721  0.99918648  0.31404901  0.83920051  0.84081288] --> [ 1.]
[ 0.02214092  0.03132355  0.14417082  0.33901317  0.91491426] --> [ 1.]
[ 0.31426055  0.80830795  0.46686523  0.58353359  0.50000842] --> [ 1.]
[ 0.27649579  0.77914451  0.33572287  0.08703303  0.50865592] --> [ 1.]
[ 0.99280349  0.24028343  0.05556034  0.31411902  0.41912574] --> [ 1.]
[ 0.91897031  0.96840695  0.23561379  0.16005505  0.06567748] --> [ 1.]
[ 0.27392867  0.44021533  0.44129147  0.40658522  0.47582736] --> [ 1.]
[ 0.82063221  0.95182938  0.64210378  0.69578691  0.2946907 ] --> [ 1.]
[ 0.12672415  0.35700418  0.89303047  0.80726545  0.79870725] --> [ 1.]
[ 0.6662085   0.41358115  0.76637022  0.82093095  0.76973305] --> [ 1.]
[ 0.96201937  0.29706843  0.22856618  0.59924945  0.05653825] --> [ 1.]
[ 0.34120276  0.71866377  0.18758929  0.52424856  0.64061623] --> [ 1.]
[ 0.25471237  0.35001821  0.63248632  0.45442404  0.96967989] --> [ 1.]
[ 0.79390087  0.00100834  0.49645204  0.55574269  0.33487764] --> [ 1.]
[ 0.41330261  0.38061826  0.33766183  0.23133121  0.80999653] --> [ 1.]
[ 0.49603561  0.33414841  0.10180184  0.9227252   0.35073833] --> [ 1.]
[ 0.17960345  0.05259438  0.565135    0.40465603  0.91518233] --> [ 1.]
[ 0.36129943  0.903603    0.63047644  0.96553285  0.94006713] --> [ 1.]
[ 0.7150973   0.93945141  0.31802763  0.15849441  0.92902078] --> [ 1.]
[ 0.23730571  0.65360248  0.68776259  0.79697206  0.86814652] --> [ 1.]
[ 0.47414382  0.75421265  0.32531333  0.43218305  0.4680773 ] --> [ 1.]
[ 0.4887811   0.66130135  0.79913557  0.68948405  0.48376372] --> [ 1.]
....

---

产生巨大错误的原因是您的标签不是二进制的，而且非常大，但softmax的输出是二进制的。例如，如果标签是10000，但您只能预测0到1之间的某个值，那么无论您预测什么，都会出现巨大的错误。您的意思是在最后一层中进行回归的

activation='linear'

？或者您是否希望在

get_data（）

？

末尾将标签通过softmax，因为您的y_序列由5个元素组成，所以您的输出模型也应该是5个元素

[ 0.06698217  0.70033048  0.4317502   0.78504855  0.26173543] --> [ 1.]
[ 0.28940025  0.21746189  0.93097653  0.94885535  0.56790348] --> [ 1.]
[ 0.69430499  0.1622601   0.22802859  0.75709315  0.88948355] --> [ 1.]

例如，试试这个网络

model = Sequential()
model.add(Dense(INDIM, input_dim=INDIM, init='uniform', activation='tanh'))
model.add(Dropout(0.5))
model.add(Dense(10, init='uniform', activation='tanh'))
model.add(Dropout(0.5))
model.add(Dense(5, init='uniform', activation='softmax'))