【发布时间】：2017-12-21 07:30:02
【问题描述】：
我是 tflearn/tensorflow 的初学者。 我正在制作一个 DNN 以将心跳分类为 Normal (0) 或 心律失常 (1)。我的数据集是ECG by MIT Arrhytmia Dataset ..
我建立了以下网络：
## Build neural network
net = tflearn.input_data(shape=[None, 200])
net = tflearn.fully_connected(net, 400)
net = tflearn.fully_connected(net, 400)
net = tflearn.fully_connected(net, 100)
net = tflearn.fully_connected(net, 50)
net = tflearn.fully_connected(net, 1, activation='linear')
net = tflearn.regression(net, optimizer='adam', learning_rate=0.01,loss='mean_square', name='target')
##
## Define model
model = tflearn.DNN(net)
print('fiting')
model.fit(TrainSet, labelSet, n_epoch=5000, batch_size=len(TrainSet), show_metric=True)
我的输出是：
y pred y target
predict:[0.01360663] [0]
predict:[0.00861748] [0]
predict:[-0.00685573] [0]
predict:[-0.20846206] [1]
我的网络分离良好，正常样本 (predict:[0.00861748] [0]) 的输出与心律失常样本 (predict:[-0.20846206] [1]) 的输出完全不同
那么，我怎样才能使输出只有0 或1 否则这个浮点值？
也许改变我的激活功能？或者用我的实际输出做点什么？
这是我的输出：model.predict(TrainingSet)
y predicted y target
predict:[-0.138634] [0]
predict:[-0.13436639] [0]
predict:[-0.12879151] [0]
predict:[-0.12057236] [0]
predict:[-0.13836551] [0]
predict:[-0.08525576] [0]
predict:[ 1.01741135] [1]
predict:[-0.11624834] [0]
predict:[-0.12631142] [0]
predict:[-0.11693959] [0]
predict:[-0.10779606] [0]
predict:[-0.11510199] [0]
predict:[-0.12450527] [0]
predict:[-0.12869376] [0]
predict:[-0.15167347] [0]
predict:[-0.14081171] [0]
predict:[-0.14235598] [0]
predict:[-0.13095573] [0]
predict:[-0.12757528] [0]
predict:[-0.14675851] [0]
predict:[-0.12311366] [0]
predict:[-0.15386838] [0]
predict:[-0.17505151] [0]
predict:[-0.13848163] [0]
predict:[-0.11671469] [0]
predict:[-0.13247125] [0]
predict:[-0.13718334] [0]
predict:[-0.12702732] [0]
predict:[-0.12665084] [0]
predict:[-0.1367469] [0]
predict:[-0.15925398] [0]
predict:[-0.13639028] [0]
predict:[-0.11569472] [0]
predict:[-0.14167] [0]
predict:[-0.12262306] [0]
predict:[-0.10863069] [0]
predict:[-0.14324963] [0]
predict:[-0.14792402] [0]
predict:[-0.14929616] [0]
predict:[-0.15551159] [0]
predict:[-0.11816701] [0]
predict:[-0.11785387] [0]
predict:[-0.15215725] [0]
predict:[-0.11279716] [0]
predict:[-0.1469961] [0]
predict:[-0.14991215] [0]
predict:[-0.11661309] [0]
predict:[-0.09011015] [0]
predict:[-0.09775476] [0]
predict:[-0.1065342] [0]
predict:[-0.11091903] [0]
predict:[-0.10344772] [0]
predict:[-0.12412915] [0]
predict:[-0.13605709] [0]
predict:[-0.12797417] [0]
predict:[-0.1076207] [0]
predict:[-0.12150024] [0]
predict:[-0.13840012] [0]
predict:[-0.13084875] [0]
predict:[-0.11066008] [0]
predict:[-0.12374203] [0]
predict:[-0.13341869] [0]
predict:[-0.12912038] [0]
predict:[-0.13748281] [0]
predict:[-0.13966258] [0]
predict:[-0.13894111] [0]
predict:[-0.10213074] [0]
predict:[-0.15602994] [0]
predict:[-0.12982219] [0]
predict:[-0.09376201] [0]
predict:[-0.08830833] [0]
predict:[-0.12029025] [0]
predict:[-0.09362413] [0]
predict:[ 1.09521723] [1]
predict:[-0.13147078] [0]
predict:[-0.1182971] [0]
predict:[-0.12983324] [0]
predict:[-0.18321729] [0]
predict:[-0.18334746] [0]
predict:[-0.2399022] [0]
【问题讨论】：
-
我不熟悉数据集，但在我看来，分离并不是那么好......我原以为输出中的第三个结果是
[1]，因为它是负面的。但是，仅从您提供的 4 个示例中不可能发现任何模式。 -
您可以将激活更改为 sigmoid。
-
您似乎在滥用网络的激活功能。奇怪的是，您使用 MSE 作为损失函数，而网络在最后一层具有线性激活。不幸的是，在 SO 中解释这个问题是题外话。考虑对神经网络进行进一步研究（包括关于典型的激活函数和损失函数），并最终调整您的问题并将其移至Cross Validated。
-
感谢回复，我改网为：
net = flearn.fully_connected(net, 1, activation='sigmoid')net = tflearn.regression(net,optimizer='adam',loss='binary_crossentropy') -
这也是不合适的：二元交叉熵期望一个 logit 作为网络的输入，而你给它一个具有 sigmoid 激活的神经元。您可能会发现使用“binary_crossentropy”对最后一个神经元使用“线性”效果更好。另请参阅其他 Stack Exchange 站点的这些问题：stats.stackexchange.com/q/260505/67965datascience.stackexchange.com/questions/9302/…
标签： python tensorflow neural-network deep-learning tflearn